feat(richie user env): adding app_image_path configuration

Add test_protected_phrases.py covering phrase-matching behavior in the
RAG engine, and update the existing ebook_search tests to use the async
SQLAlchemy engine/session (create_async_engine, AsyncSession) and async
HTTP paths.

.dockerignore

@@ -0,0 +1,15 @@
+.git
+.direnv
+.mypy_cache
+.pytest_cache
+.ruff_cache
+__pycache__
+**/__pycache__
+*.pyc
+*.pyo
+.ebook_search_bm25
+result
+result-*
+*.egg-info
+dist
+build

.github/workflows/build_systems.yml

@@ -17,12 +17,11 @@ jobs:
           - "bob"
           - "brain"
           - "jeeves"
-          - "leviathan"
           - "rhapsody-in-green"
     continue-on-error: true
     steps:
       - uses: actions/checkout@v4
       - name: Build default package
-        run: "nixos-rebuild build --flake ./#${{ matrix.system }}"
+        run: "nixos-rebuild build --accept-flake-config --flake ./#${{ matrix.system }}"
       - name: copy to nix-cache
         run: nix copy --accept-flake-config --to unix:///host-nix/var/nix/daemon-socket/socket .#nixosConfigurations.${{ matrix.system }}.config.system.build.toplevel

.github/workflows/fix_eval_warnings.yml

@@ -1,30 +0,0 @@
-name: fix_eval_warnings
-on:
-  workflow_run:
-    workflows: ["build_systems"]
-    types: [completed]
-
-jobs:
-  check-warnings:
-    if: >-
-      github.event.workflow_run.conclusion != 'cancelled' &&
-      github.event.workflow_run.head_branch == 'main' &&
-      (github.event.workflow_run.event == 'push' || github.event.workflow_run.event == 'schedule')
-    runs-on: self-hosted
-    permissions:
-      contents: write
-      pull-requests: write
-
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Fix eval warnings
-        env:
-          GH_TOKEN: ${{ secrets.GH_TOKEN_FOR_UPDATES }}
-        run: >-
-          nix develop .#devShells.x86_64-linux.default -c
-          python -m python.eval_warnings.main
-          --run-id "${{ github.event.workflow_run.id }}"
-          --repo "${{ github.repository }}"
-          --ollama-url "${{ secrets.OLLAMA_URL }}"
-          --run-url "${{ github.event.workflow_run.html_url }}"

.github/workflows/merge_flake_lock_update.yml

@@ -6,24 +6,18 @@ on:
 
 jobs:
   merge:
-    runs-on: ubuntu-latest
+    runs-on: self-hosted
 
     permissions:
       contents: write
       pull-requests: write
 
     steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
       - name: merge_flake_lock_update
-        run: |
+        run: >-
-          pr_number=$(gh pr list --state open --author RichieCahill --label flake_lock_update --json number --jq '.[0].number')
+          nix develop .#devShells.x86_64-linux.default -c
-          echo "pr_number=$pr_number" >> $GITHUB_ENV
+          python -m python.gitea_flake_lock merge
-          if [ -n "$pr_number" ]; then
+          --repo "${{ github.repository }}"
-            gh pr merge "$pr_number" --rebase
-          else
-            echo "No open PR found with label flake_lock_update"
-          fi
         env:
-          GITHUB_TOKEN: ${{ secrets.GH_TOKEN_FOR_UPDATES }}
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          GITEA_URL: https://gitea.tmmworkshop.com

.github/workflows/pytest.yml

@@ -1,13 +1,13 @@
 name: pytest
 
 on:
+  workflow_dispatch:
   push:
     branches:
       - main
   pull_request:
     branches:
       - main
-  merge_group:
 
 jobs:
   pytest:

.github/workflows/update-flake-lock.yml

@@ -6,18 +6,21 @@ on:
 
 jobs:
   lockfile:
-    runs-on: ubuntu-latest
+    runs-on: self-hosted
+    permissions:
+      actions: write
+      contents: write
+      pull-requests: write
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
-      - name: Install Nix
-        uses: DeterminateSystems/nix-installer-action@main
       - name: Update flake.lock
-        uses: DeterminateSystems/update-flake-lock@main
+        run: nix flake update
-        with:
+      - name: Create or update flake.lock PR
-          token: ${{ secrets.GH_TOKEN_FOR_UPDATES }}
+        env:
-          pr-title: "Update flake.lock"
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
-          pr-labels: |
+          GITEA_URL: https://gitea.tmmworkshop.com
-            dependencies
+        run: >-
-            automated
+          nix develop .#devShells.x86_64-linux.default -c
-            flake_lock_update
+          python -m python.gitea_flake_lock update
+          --repo "${{ github.repository }}"

.gitignore

@@ -172,3 +172,4 @@ frontend/node_modules/
 
 # data from testing llms
 data/*
+.ebook_search_bm25

.sops.yaml

@@ -7,7 +7,6 @@ keys:
   - &system_bob age1q47vup0tjhulkg7d6xwmdsgrw64h4ax3la3evzqpxyy4adsmk9fs56qz3y # cspell:disable-line
   - &system_brain age1jhf7vm0005j60mjq63696frrmjhpy8kpc2d66mw044lqap5mjv4snmwvwm # cspell:disable-line
   - &system_jeeves age13lmqgc3jvkyah5e3vcwmj4s5wsc2akctcga0lpc0x8v8du3fxprqp4ldkv # cspell:disable-line
-  - &system_leviathan age1l272y8udvg60z7edgje42fu49uwt4x2gxn5zvywssnv9h2krms8s094m4k # cspell:disable-line
   - &system_rhapsody age1ufnewppysaq2wwcl4ugngjz8pfzc5a35yg7luq0qmuqvctajcycs5lf6k4 # cspell:disable-line
 
 creation_rules:
@@ -18,5 +17,4 @@ creation_rules:
           - *system_bob
           - *system_brain
           - *system_jeeves
-          - *system_leviathan
           - *system_rhapsody

.vscode/settings.json

@@ -71,6 +71,7 @@
     "ehci",
     "emerg",
     "endlessh",
+    "ents",
     "errorlens",
     "esbenp",
     "esphome",
@@ -172,6 +173,8 @@
     "Networkd",
     "networkmanager",
     "newtabpage",
+    "ngram",
+    "ngrams",
     "nixfmt",
     "nixos",
     "nixpkgs",
@@ -242,6 +245,7 @@
     "referer",
     "REFERERS",
     "relatime",
+    "rerank",
     "Rhosts",
     "ripgrep",
     "roboto",
@@ -297,7 +301,9 @@
     "uiprotect",
     "uitour",
     "unifi",
+    "unjudged",
     "unrar",
+    "unstorable",
     "unsubmitted",
     "uptimekuma",
     "urlbar",
@@ -325,6 +331,7 @@
     "xcursorgen",
     "xdist",
     "xhci",
+    "yake",
     "yazi",
     "yubikey",
     "yubioath",

AGENTS.md

@@ -1,12 +0,0 @@
-## Dev environment tips
-
-- use treefmt to format all files
-- make python code ruff compliant
-- use pytest to test python code
-- always use the minimum amount of complexity
-- if judgment calls are easy to reverse make them. if not ask me first
-- Match existing code style.
-- Use builtin helpers getenv() over os.environ.get.
-- Prefer single-purpose functions over “do everything” helpers.
-- Avoid compatibility branches like PG_USER and POSTGRESQL_URL unless requested.
-- Keep helpers only if reused or they simplify the code otherwise inline.

README.md

@@ -1 +1,51 @@
 # dotfiles
+
+## Installer ISO
+
+Build a bootable NixOS ISO with the installer preinstalled:
+
+```sh
+nix build .#iso
+```
+
+Write `result/iso/nixos-zfs-installer.iso` to a USB stick (for example with `dd`) or boot it in a VM. The image is the minimal NixOS installation CD with ZFS enabled and `nixos-installer` on `PATH`. SSH is enabled and the `nixos` and `root` accounts use the password `nixos`, so you can also run the installer remotely. Once booted:
+
+```sh
+sudo nixos-installer
+```
+
+The ISO bundles the `.#installer-nixos` package, a variant of the binary that keeps its Nix store linkage instead of being patched for foreign distributions.
+
+## Installer binary
+
+Build the self-contained installer executable with:
+
+```sh
+nix build .#installer
+```
+
+The flake package (defined in `python/installer/package.nix`) uses the Python builder in `python/installer/build.py`, which stages only the installer modules before running PyInstaller. You can also call it directly when `pyinstaller` and `patchelf` are on `PATH`:
+
+```sh
+python -m python.installer.build --output ./nixos-installer
+```
+
+Copy `result/bin/nixos-installer` to the installer USB stick and run it as root from the NixOS live environment:
+
+```sh
+sudo ./nixos-installer
+```
+
+Validate the live environment first with:
+
+```sh
+./nixos-installer --check
+```
+
+Paste a value into the TUI encryption password field to enable LUKS during install, or set `ENCRYPT_KEY`:
+
+```sh
+sudo env ENCRYPT_KEY='change-me' ./nixos-installer
+```
+
+The binary bundles the Python runtime and only the installer modules it imports. It still expects the NixOS installer environment to provide system install tools such as `parted`, `zfs`, `zpool`, `cryptsetup`, `nixos-generate-config`, and `nixos-install`.

common/global/default.nix

@@ -23,7 +23,10 @@
   boot = {
     tmp.useTmpfs = true;
     kernelPackages = lib.mkDefault pkgs.linuxPackages_6_12;
-    zfs.package = lib.mkDefault pkgs.zfs_2_4;
+    zfs = {
+      package = lib.mkDefault pkgs.zfs_2_4;
+      forceImportRoot = lib.mkDefault false;
+    };
   };
 
   hardware.enableRedistributableFirmware = true;
@@ -37,10 +40,17 @@
 
   nixpkgs = {
     overlays = builtins.attrValues outputs.overlays;
-    config.allowUnfree = true;
+    config = {
+      allowUnfree = true;
+      permittedInsecurePackages = [
+        "openssl-1.1.1w" # This is for discord-canary
+      ];
+    };
   };
 
   services = {
+    dbus.implementation = "dbus";
+
     # firmware update
     fwupd.enable = true;
 

common/optional/monitoring-agent.nix

@@ -0,0 +1,256 @@
+{
+  config,
+  lib,
+  pkgs,
+  ...
+}:
+let
+  monitoringInterface = "ztwfunumly";
+  nodeTextfileDir = "/var/lib/prometheus-node-exporter-textfile";
+
+  mkProcessNameTemplate =
+    perPid: template: if perPid then "${template}:{{.PID}}:{{.StartTime}}" else template;
+
+  mkProcessMatchers = perPid: [
+    {
+      name = mkProcessNameTemplate perPid "{{.Username}}:{{.Matches.Module}}";
+      cmdline = [ "^/nix/store[^ ]*/bin/python[^ ]* -m (?P<Module>[^ ]+)" ];
+    }
+    {
+      name = mkProcessNameTemplate perPid "{{.Username}}:{{.Matches.Wrapped}}";
+      cmdline = [
+        "^/nix/store[^ ]*/bin/python[^ ]* /nix/store[^ ]*/bin/\\.?(?P<Wrapped>[^ /]+?)(?:-wrapped)?(?:\\s|$)"
+      ];
+    }
+    {
+      name = mkProcessNameTemplate perPid "{{.Username}}:{{.Matches.Wrapped}}";
+      cmdline = [
+        "^/nix/store[^ ]*/bin/node /nix/store[^ ]*-(?P<Wrapped>[A-Za-z0-9._+-]+)-[0-9][^ /]*/"
+      ];
+    }
+    {
+      name = mkProcessNameTemplate perPid "{{.Username}}:{{.Matches.Wrapped}}";
+      cmdline = [ "^/nix/store[^ ]*/(?:bin/|lib/[^ ]*/)?\\.?(?P<Wrapped>[^ /]+?)(?:-wrapped)?(?:\\s|$)" ];
+    }
+    {
+      name = mkProcessNameTemplate perPid "{{.Username}}:{{.ExeBase}}";
+      cmdline = [ ".+" ];
+    }
+  ];
+
+  perPidConfig = pkgs.writeText "process-exporter-per-pid.yaml" (
+    builtins.toJSON {
+      process_names = mkProcessMatchers true;
+    }
+  );
+
+  zpoolLatencyScript = pkgs.writeShellScript "zpool-latency-exporter" ''
+        set -euo pipefail
+
+        out_dir=${lib.escapeShellArg nodeTextfileDir}
+        host=${lib.escapeShellArg config.networking.hostName}
+        tmp_file="$(mktemp "$out_dir/zpool.prom.XXXXXX")"
+        trap 'rm -f "$tmp_file"' EXIT
+
+        pools="$(zpool list -H -o name | paste -sd, -)"
+
+        cat >"$tmp_file" <<'EOF'
+    # HELP zpool_iostat_total_wait_read_ns Average total read wait time reported by zpool iostat.
+    # TYPE zpool_iostat_total_wait_read_ns gauge
+    # HELP zpool_iostat_total_wait_write_ns Average total write wait time reported by zpool iostat.
+    # TYPE zpool_iostat_total_wait_write_ns gauge
+    # HELP zpool_iostat_disk_wait_read_ns Average disk read wait time reported by zpool iostat.
+    # TYPE zpool_iostat_disk_wait_read_ns gauge
+    # HELP zpool_iostat_disk_wait_write_ns Average disk write wait time reported by zpool iostat.
+    # TYPE zpool_iostat_disk_wait_write_ns gauge
+    # HELP zpool_iostat_syncq_wait_read_ns Average synchronous queue read wait time reported by zpool iostat.
+    # TYPE zpool_iostat_syncq_wait_read_ns gauge
+    # HELP zpool_iostat_syncq_wait_write_ns Average synchronous queue write wait time reported by zpool iostat.
+    # TYPE zpool_iostat_syncq_wait_write_ns gauge
+    # HELP zpool_iostat_asyncq_wait_read_ns Average asynchronous queue read wait time reported by zpool iostat.
+    # TYPE zpool_iostat_asyncq_wait_read_ns gauge
+    # HELP zpool_iostat_asyncq_wait_write_ns Average asynchronous queue write wait time reported by zpool iostat.
+    # TYPE zpool_iostat_asyncq_wait_write_ns gauge
+    EOF
+
+        zpool iostat -Hplvy -y 1 1 | awk -F '\t' -v host="$host" -v pools="$pools" '
+          function esc(str, out) {
+            out = str
+            gsub(/\\/, "\\\\", out)
+            gsub(/"/, "\\\"", out)
+            return out
+          }
+
+          function emit(metric, pool, vdev, value) {
+            if (value == "" || value == "-") {
+              return
+            }
+
+            printf "%s{host=\"%s\",pool=\"%s\",vdev=\"%s\"} %s\n",
+              metric,
+              esc(host),
+              esc(pool),
+              esc(vdev),
+              value
+          }
+
+          BEGIN {
+            split(pools, pool_names, ",")
+            for (idx in pool_names) {
+              if (pool_names[idx] != "") {
+                known_pools[pool_names[idx]] = 1
+              }
+            }
+          }
+
+          NF == 0 {
+            next
+          }
+
+          {
+            row_name = $1
+
+            if (row_name in known_pools) {
+              current_pool = row_name
+              current_vdev = "_pool"
+            } else if (current_pool == "") {
+              next
+            } else {
+              current_vdev = row_name
+            }
+
+            emit("zpool_iostat_total_wait_read_ns", current_pool, current_vdev, $8)
+            emit("zpool_iostat_total_wait_write_ns", current_pool, current_vdev, $9)
+            emit("zpool_iostat_disk_wait_read_ns", current_pool, current_vdev, $10)
+            emit("zpool_iostat_disk_wait_write_ns", current_pool, current_vdev, $11)
+            emit("zpool_iostat_syncq_wait_read_ns", current_pool, current_vdev, $12)
+            emit("zpool_iostat_syncq_wait_write_ns", current_pool, current_vdev, $13)
+            emit("zpool_iostat_asyncq_wait_read_ns", current_pool, current_vdev, $14)
+            emit("zpool_iostat_asyncq_wait_write_ns", current_pool, current_vdev, $15)
+          }
+        ' >>"$tmp_file"
+
+        mv "$tmp_file" "$out_dir/zpool.prom"
+        trap - EXIT
+  '';
+in
+{
+  networking.firewall.interfaces.${monitoringInterface}.allowedTCPPorts = [
+    9100
+    9134
+    9256
+    9257
+    9633
+  ];
+
+  services.prometheus.exporters = {
+    node = {
+      enable = true;
+      enabledCollectors = [
+        "pressure"
+        "processes"
+        "systemd"
+      ];
+      extraFlags = [ "--collector.textfile.directory=${nodeTextfileDir}" ];
+    };
+
+    process = {
+      enable = true;
+      user = "root";
+      group = "root";
+      settings.process_names = mkProcessMatchers false;
+      extraFlags = [
+        "-gather-smaps=false"
+        "-remove-empty-groups=true"
+        "-threads=false"
+      ];
+    };
+
+    smartctl.enable = true;
+    zfs.enable = true;
+  };
+
+  programs.atop = {
+    enable = true;
+    atopService.enable = true;
+    atopRotateTimer.enable = true;
+    atopacctService.enable = true;
+    settings.interval = 30;
+  };
+
+  systemd = {
+    services = {
+      prometheus-process-pid-exporter = {
+        description = "Prometheus process exporter with per-PID naming";
+        wantedBy = [ "multi-user.target" ];
+        after = [ "network.target" ];
+        serviceConfig = {
+          ExecStart = ''
+            ${pkgs.prometheus-process-exporter}/bin/process-exporter \
+              --web.listen-address 0.0.0.0:9257 \
+              --config.path ${perPidConfig} \
+              -children=false \
+              -gather-smaps=false \
+              -remove-empty-groups=true \
+              -threads=false
+          '';
+          User = "root";
+          Group = "root";
+          Restart = "always";
+          WorkingDirectory = "/tmp";
+          CapabilityBoundingSet = [ "" ];
+          DeviceAllow = [ "" ];
+          LockPersonality = true;
+          MemoryDenyWriteExecute = true;
+          NoNewPrivileges = true;
+          PrivateDevices = true;
+          PrivateTmp = true;
+          ProtectClock = true;
+          ProtectControlGroups = true;
+          ProtectHome = true;
+          ProtectHostname = true;
+          ProtectKernelLogs = true;
+          ProtectKernelModules = true;
+          ProtectKernelTunables = true;
+          ProtectSystem = "strict";
+          RemoveIPC = true;
+          RestrictAddressFamilies = [
+            "AF_INET"
+            "AF_INET6"
+          ];
+          RestrictNamespaces = true;
+          RestrictRealtime = true;
+          RestrictSUIDSGID = true;
+          SystemCallArchitectures = "native";
+          UMask = "0077";
+        };
+      };
+
+      zpool-latency-exporter = {
+        description = "Exports ZFS latency metrics for node_exporter textfile collection";
+        after = [ "zfs-import.target" ];
+        requires = [ "zfs-import.target" ];
+        path = [
+          config.boot.zfs.package
+          pkgs.coreutils
+          pkgs.gawk
+        ];
+        serviceConfig = {
+          Type = "oneshot";
+          ExecStart = zpoolLatencyScript;
+        };
+      };
+    };
+
+    timers.zpool-latency-exporter = {
+      wantedBy = [ "timers.target" ];
+      timerConfig = {
+        OnBootSec = "2m";
+        OnUnitActiveSec = "60s";
+        Unit = "zpool-latency-exporter.service";
+      };
+    };
+
+    tmpfiles.rules = [ "d ${nodeTextfileDir} 0755 root root - -" ];
+  };
+}

common/optional/update.nix

@@ -4,7 +4,7 @@
     flags = [ "--accept-flake-config" ];
     randomizedDelaySec = "1h";
     persistent = true;
-    flake = "github:RichieCahill/dotfiles";
+    flake = "git+https://gitea.tmmworkshop.com/richie/dotfiles?ref=main";
     allowReboot = true;
     dates = "Sat *-*-* 06:00:00";
   };

docs/zfs_failed_root_import_recovry.md

@@ -0,0 +1,76 @@
+# ZFS failed root import recovery
+
+## Fast path
+
+If the machine fails to boot because ZFS refuses to import `root_pool`:
+
+### GRUB
+
+1. At the bootloader menu, select the normal NixOS entry.
+2. Press `e`.
+3. Find the line that starts with `linux`.
+4. Append this to the end of that line:
+
+```text
+zfs_force=1
+```
+
+5. Boot once with `Ctrl+x` or `F10`.
+
+### systemd-boot
+
+1. At the bootloader menu, highlight the normal NixOS entry.
+2. Press `e`.
+3. Append this to the end of the options line:
+
+```text
+zfs_force=1
+```
+
+4. Press `Enter` to boot once.
+
+## After boot
+
+Run:
+
+```bash
+sudo zpool status
+sudo zpool import
+journalctl -b | rg "ZFS|zfs|import|root_pool"
+```
+
+## Expected result
+
+`sudo zpool status` should show `root_pool` as `ONLINE`.
+
+## Reboot test
+
+Run:
+
+```bash
+sudo reboot
+```
+
+Do not add `zfs_force=1` the second time.
+
+## If it still fails
+
+Boot once more with:
+
+```text
+zfs_force=1
+```
+
+Then run:
+
+```bash
+sudo zpool status -v
+sudo zpool history | tail -n 50
+journalctl -b | rg "ZFS|zfs|import|root_pool"
+```
+
+## Notes
+
+- Root pool name is `root_pool`.
+- This is a one-time recovery path after disk moves, controller changes, dirty exports, or interrupted imports.
+- Some hosts also need the LUKS unlock USB key inserted before boot.

flake.lock

@@ -8,11 +8,11 @@
       },
       "locked": {
         "dir": "pkgs/firefox-addons",
-        "lastModified": 1777435375,
+        "lastModified": 1782964936,
-        "narHash": "sha256-2WRfJbipnTz+EY3rHRnCoG4kWkzPczb/cLcWwhy/0QA=",
+        "narHash": "sha256-wXEBDr7/dFQYhVpDwCKc9fkrYQQE4x0bdirX1bsLBGA=",
         "owner": "rycee",
         "repo": "nur-expressions",
-        "rev": "4d89e8e2c50711ee3fea3a25e662cfa5c6628e07",
+        "rev": "64feee871e0373dd6121e412c3fb12e372d1bfb5",
         "type": "gitlab"
       },
       "original": {
@@ -29,11 +29,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1777434174,
+        "lastModified": 1783005591,
-        "narHash": "sha256-KwTyQ5g2qDhWIs/O6vH8HeF8n4JCzZIT/VYE7nYnukQ=",
+        "narHash": "sha256-NcLHV5uBAeggDUE2wPbKszjfyaSLsoqaYt7izOphkZw=",
         "owner": "nix-community",
         "repo": "home-manager",
-        "rev": "d3b4e4b1bd59aedd3d4eb0a8df7162edb6da4607",
+        "rev": "f469c79b955609d6a8fdd9e689be76a93b1621d7",
         "type": "github"
       },
       "original": {
@@ -43,12 +43,15 @@
       }
     },
     "nixos-hardware": {
+      "inputs": {
+        "nixpkgs": "nixpkgs"
+      },
       "locked": {
-        "lastModified": 1776983936,
+        "lastModified": 1782562157,
-        "narHash": "sha256-ZOQyNqSvJ8UdrrqU1p7vaFcdL53idK+LOM8oRWEWh6o=",
+        "narHash": "sha256-a7+T6QSeowynwZ1ZJJbP8T8ntAytvrui8kFGJmIZt2c=",
         "owner": "nixos",
         "repo": "nixos-hardware",
-        "rev": "2096f3f411ce46e88a79ae4eafcfc9df8ed41c61",
+        "rev": "a9cf7546a938c737b079e738de73934a13de9784",
         "type": "github"
       },
       "original": {
@@ -60,27 +63,24 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1777268161,
+        "lastModified": 1767892417,
-        "narHash": "sha256-bxrdOn8SCOv8tN4JbTF/TXq7kjo9ag4M+C8yzzIRYbE=",
+        "narHash": "sha256-8bW3q88CEg2u4hSP66Vf4lpbLonHz7hqDNBMcCY7E9U=",
-        "owner": "nixos",
+        "rev": "3497aa5c9457a9d88d71fa93a4a8368816fbeeba",
-        "repo": "nixpkgs",
+        "type": "tarball",
-        "rev": "1c3fe55ad329cbcb28471bb30f05c9827f724c76",
+        "url": "https://releases.nixos.org/nixos/unstable/nixos-26.05pre924538.3497aa5c9457/nixexprs.tar.xz"
-        "type": "github"
       },
       "original": {
-        "owner": "nixos",
+        "type": "tarball",
-        "ref": "nixos-unstable",
+        "url": "https://channels.nixos.org/nixos-unstable/nixexprs.tar.xz"
-        "repo": "nixpkgs",
-        "type": "github"
       }
     },
     "nixpkgs-master": {
       "locked": {
-        "lastModified": 1777437048,
+        "lastModified": 1783021952,
-        "narHash": "sha256-Ca4jKXJuYp1D+DqiuQ/vGHRYKPlAZTn1vq7XDU9t18w=",
+        "narHash": "sha256-8PghAtSGGZ0umfVI8Qbd7ZbFrfZPiH1UwtVbgLeikDA=",
         "owner": "nixos",
         "repo": "nixpkgs",
-        "rev": "1e1459dda883651ef85e23c7c6e2224cba195065",
+        "rev": "f136374c679c54171a3ace589d15e9e79a8bd086",
         "type": "github"
       },
       "original": {
@@ -106,12 +106,28 @@
         "type": "github"
       }
     },
+    "nixpkgs_2": {
+      "locked": {
+        "lastModified": 1782723713,
+        "narHash": "sha256-oPXCU/SSUokcGaJREHibG1CBX3+s/W7orDWQOZDsEeQ=",
+        "owner": "nixos",
+        "repo": "nixpkgs",
+        "rev": "b5aa0fbd538984f6e3d201be0005b4463d8b09f8",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nixos",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
     "root": {
       "inputs": {
         "firefox-addons": "firefox-addons",
         "home-manager": "home-manager",
         "nixos-hardware": "nixos-hardware",
-        "nixpkgs": "nixpkgs",
+        "nixpkgs": "nixpkgs_2",
         "nixpkgs-master": "nixpkgs-master",
         "nixpkgs-stable": "nixpkgs-stable",
         "sops-nix": "sops-nix",
@@ -125,11 +141,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1777338324,
+        "lastModified": 1782165805,
-        "narHash": "sha256-bc+ZZCmOTNq86/svGnw0tVpH7vJaLYvGLLKFYP08Q8E=",
+        "narHash": "sha256-478kKQBvK6SYTOdN2h9jhKJv94nbXRbFMfuL1WshErg=",
         "owner": "Mic92",
         "repo": "sops-nix",
-        "rev": "8eaee5c45428b28b8c47a83e4c09dccec5f279b5",
+        "rev": "56b24064fdcaedca53553b1a6d607fd23b613a24",
         "type": "github"
       },
       "original": {

flake.nix

@@ -65,38 +65,48 @@
 
       devShells = forEachSystem (pkgs: import ./shell.nix { inherit pkgs; });
       formatter = forEachSystem (pkgs: pkgs.treefmt);
+      packages = forEachSystem (
+        pkgs:
+        let
+          installer = pkgs.callPackage ./python/installer/package.nix { };
+          installer-nixos = pkgs.callPackage ./python/installer/package.nix { patchElf = false; };
+        in
+        {
+          inherit installer installer-nixos;
+          default = installer;
+        }
+        // lib.optionalAttrs (pkgs.stdenv.hostPlatform.system == "x86_64-linux") {
+          iso = self.nixosConfigurations.iso.config.system.build.isoImage;
+        }
+      );
+      apps = forEachSystem (
+        pkgs:
+        let
+          system = pkgs.stdenv.hostPlatform.system;
+          installer = {
+            type = "app";
+            program = "${self.packages.${system}.installer}/bin/nixos-installer";
+            meta.description = "One-file NixOS ZFS installer.";
+          };
+        in
+        {
+          inherit installer;
+          default = installer;
+        }
+      );
 
-      nixosConfigurations = {
+      nixosConfigurations =
-        bob = lib.nixosSystem {
+        let
-          modules = [
+          hosts = builtins.attrNames (
-            ./systems/bob
+            lib.filterAttrs (_: type: type == "directory") (builtins.readDir ./systems)
-          ];
+          );
+          mkHost =
+            name:
+            lib.nixosSystem {
+              modules = [ ./systems/${name} ];
               specialArgs = { inherit inputs outputs; };
             };
-        brain = lib.nixosSystem {
+        in
-          modules = [
+        lib.genAttrs hosts mkHost;
-            ./systems/brain
-          ];
-          specialArgs = { inherit inputs outputs; };
-        };
-        jeeves = lib.nixosSystem {
-          modules = [
-            ./systems/jeeves
-          ];
-          specialArgs = { inherit inputs outputs; };
-        };
-        rhapsody-in-green = lib.nixosSystem {
-          modules = [
-            ./systems/rhapsody-in-green
-          ];
-          specialArgs = { inherit inputs outputs; };
-        };
-        leviathan = lib.nixosSystem {
-          modules = [
-            ./systems/leviathan
-          ];
-          specialArgs = { inherit inputs outputs; };
-        };
-      };
     };
 }

frontend/.gitignore

@@ -1,24 +0,0 @@
-# Logs
-logs
-*.log
-npm-debug.log*
-yarn-debug.log*
-yarn-error.log*
-pnpm-debug.log*
-lerna-debug.log*
-
-node_modules
-dist
-dist-ssr
-*.local
-
-# Editor directories and files
-.vscode/*
-!.vscode/extensions.json
-.idea
-.DS_Store
-*.suo
-*.ntvs*
-*.njsproj
-*.sln
-*.sw?

overlays/default.nix

@@ -17,17 +17,17 @@
 
   python-env = final: _prev: {
     my_python = final.python314.withPackages (
-      ps: with ps; [
+      ps:
+      with ps;
+      [
         alembic
         apprise
         apscheduler
         fastapi
         fastapi-cli
-        faster-whisper
         httpx
         mypy
-        orjson
+        pgvector
-        polars
         psycopg
         pydantic
         pyfakefs
@@ -37,12 +37,8 @@
         pytest-xdist
         python-multipart
         ruff
-        scalene
-        sqlalchemy
         sqlalchemy
         tenacity
-        textual
-        tiktoken
         tinytuya
         typer
         websockets

pyproject.toml

@@ -3,7 +3,7 @@ name = "system_tools"
 version = "0.1.0"
 description = ""
 authors = [{ name = "Richie Cahill", email = "richie@tmmworkshop.com" }]
-requires-python = "~=3.13.0"
+requires-python = "~=3.14.0"
 readme = "README.md"
 license = "MIT"
 # these dependencies are a best effort and aren't guaranteed to work
@@ -12,26 +12,39 @@ dependencies = [
     "alembic",
     "apprise",
     "apscheduler",
+    "beautifulsoup4",
+    "bm25s",
+    "ebooklib",
+    "fastapi",
+    "fastapi-cli",
     "httpx",
-    "python-multipart",
+    "jinja2",
+    "pgvector",
     "polars",
     "psycopg[binary]",
     "pydantic",
-    "pyyaml",
+    "pydantic-settings",
-    "sqlalchemy",
+    "python-multipart",
+    "sqlalchemy[asyncio]",
+    "tenacity",
+    "tiktoken",
+    "tinytuya",
     "typer",
+    "uvicorn",
     "websockets",
+    "yake",
 ]
 
 [project.scripts]
 database = "python.database_cli:app"
-van-inventory = "python.van_inventory.main:serve"
 whisper-transcribe = "python.tools.whisper.transcribe:main"
 
 [dependency-groups]
 dev = [
+    "aiosqlite",
     "mypy",
     "pyfakefs",
+    "pytest-asyncio",
     "pytest-cov",
     "pytest-mock",
     "pytest-xdist",
@@ -41,7 +54,7 @@ dev = [
 
 [tool.ruff]
 
-target-version = "py313"
+target-version = "py314"
 
 line-length = 120
 
@@ -84,9 +97,6 @@ lint.ignore = [
 "python/alembic/**" = [
     "INP001", # (perm) this creates LSP issues for alembic
 ]
-"python/signal_bot/**" = [
-    "D107", # (perm) class docstrings cover __init__
-]
 
 [tool.ruff.lint.pydocstyle]
 convention = "google"
@@ -110,5 +120,6 @@ exclude_lines = [
 
 [tool.pytest.ini_options]
 addopts = "-n auto -ra"
+asyncio_mode = "auto"
 testpaths = ["tests"]
 # --cov=system_tools --cov-report=term-missing --cov-report=xml --cov-report=html --cov-branch

python/alembic/data_science_dev/versions/2026_03_24-adding_failedingestion_2f43120e3ffc.py

@@ -1,50 +0,0 @@
-"""adding FailedIngestion.
-
-Revision ID: 2f43120e3ffc
-Revises: f99be864fe69
-Create Date: 2026-03-24 23:46:17.277897
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import sqlalchemy as sa
-from alembic import op
-
-from python.orm import DataScienceDevBase
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "2f43120e3ffc"
-down_revision: str | None = "f99be864fe69"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = DataScienceDevBase.schema_name
-
-
-def upgrade() -> None:
-    """Upgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.create_table(
-        "failed_ingestion",
-        sa.Column("raw_line", sa.Text(), nullable=False),
-        sa.Column("error", sa.Text(), nullable=False),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_failed_ingestion")),
-        schema=schema,
-    )
-    # ### end Alembic commands ###
-
-
-def downgrade() -> None:
-    """Downgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_table("failed_ingestion", schema=schema)
-    # ### end Alembic commands ###

python/alembic/data_science_dev/versions/2026_03_25-attach_partitions_to_posts.py

@@ -1,72 +0,0 @@
-"""Attach all partition tables to the posts parent table.
-
-Alembic autogenerate creates partition tables as standalone tables but does not
-emit the ALTER TABLE ... ATTACH PARTITION statements needed for PostgreSQL to
-route inserts to the correct partition.
-
-Revision ID: a1b2c3d4e5f6
-Revises: 605b1794838f
-Create Date: 2026-03-25 10:00:00.000000
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-from alembic import op
-from sqlalchemy import text
-
-from python.orm import DataScienceDevBase
-from python.orm.data_science_dev.posts.partitions import (
-    PARTITION_END_YEAR,
-    PARTITION_START_YEAR,
-    iso_weeks_in_year,
-    week_bounds,
-)
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "a1b2c3d4e5f6"
-down_revision: str | None = "605b1794838f"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = DataScienceDevBase.schema_name
-
-ALREADY_ATTACHED_QUERY = text("""
-    SELECT inhrelid::regclass::text
-    FROM pg_inherits
-    WHERE inhparent = :parent::regclass
-""")
-
-
-def upgrade() -> None:
-    """Attach all weekly partition tables to the posts parent table."""
-    connection = op.get_bind()
-    already_attached = {row[0] for row in connection.execute(ALREADY_ATTACHED_QUERY, {"parent": f"{schema}.posts"})}
-
-    for year in range(PARTITION_START_YEAR, PARTITION_END_YEAR + 1):
-        for week in range(1, iso_weeks_in_year(year) + 1):
-            table_name = f"posts_{year}_{week:02d}"
-            qualified_name = f"{schema}.{table_name}"
-            if qualified_name in already_attached:
-                continue
-            start, end = week_bounds(year, week)
-            start_str = start.strftime("%Y-%m-%d %H:%M:%S")
-            end_str = end.strftime("%Y-%m-%d %H:%M:%S")
-            op.execute(
-                f"ALTER TABLE {schema}.posts "
-                f"ATTACH PARTITION {qualified_name} "
-                f"FOR VALUES FROM ('{start_str}') TO ('{end_str}')"
-            )
-
-
-def downgrade() -> None:
-    """Detach all weekly partition tables from the posts parent table."""
-    for year in range(PARTITION_START_YEAR, PARTITION_END_YEAR + 1):
-        for week in range(1, iso_weeks_in_year(year) + 1):
-            table_name = f"posts_{year}_{week:02d}"
-            op.execute(f"ALTER TABLE {schema}.posts DETACH PARTITION {schema}.{table_name}")

python/alembic/data_science_dev/versions/2026_03_27-adding_congress_data_83bfc8af92d8.py

@@ -1,153 +0,0 @@
-"""adding congress data.
-
-Revision ID: 83bfc8af92d8
-Revises: a1b2c3d4e5f6
-Create Date: 2026-03-27 10:43:02.324510
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import sqlalchemy as sa
-from alembic import op
-
-from python.orm import DataScienceDevBase
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "83bfc8af92d8"
-down_revision: str | None = "a1b2c3d4e5f6"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = DataScienceDevBase.schema_name
-
-
-def upgrade() -> None:
-    """Upgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.create_table(
-        "bill",
-        sa.Column("congress", sa.Integer(), nullable=False),
-        sa.Column("bill_type", sa.String(), nullable=False),
-        sa.Column("number", sa.Integer(), nullable=False),
-        sa.Column("title", sa.String(), nullable=True),
-        sa.Column("title_short", sa.String(), nullable=True),
-        sa.Column("official_title", sa.String(), nullable=True),
-        sa.Column("status", sa.String(), nullable=True),
-        sa.Column("status_at", sa.Date(), nullable=True),
-        sa.Column("sponsor_bioguide_id", sa.String(), nullable=True),
-        sa.Column("subjects_top_term", sa.String(), nullable=True),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_bill")),
-        sa.UniqueConstraint("congress", "bill_type", "number", name="uq_bill_congress_type_number"),
-        schema=schema,
-    )
-    op.create_index("ix_bill_congress", "bill", ["congress"], unique=False, schema=schema)
-    op.create_table(
-        "legislator",
-        sa.Column("bioguide_id", sa.Text(), nullable=False),
-        sa.Column("thomas_id", sa.String(), nullable=True),
-        sa.Column("lis_id", sa.String(), nullable=True),
-        sa.Column("govtrack_id", sa.Integer(), nullable=True),
-        sa.Column("opensecrets_id", sa.String(), nullable=True),
-        sa.Column("fec_ids", sa.String(), nullable=True),
-        sa.Column("first_name", sa.String(), nullable=False),
-        sa.Column("last_name", sa.String(), nullable=False),
-        sa.Column("official_full_name", sa.String(), nullable=True),
-        sa.Column("nickname", sa.String(), nullable=True),
-        sa.Column("birthday", sa.Date(), nullable=True),
-        sa.Column("gender", sa.String(), nullable=True),
-        sa.Column("current_party", sa.String(), nullable=True),
-        sa.Column("current_state", sa.String(), nullable=True),
-        sa.Column("current_district", sa.Integer(), nullable=True),
-        sa.Column("current_chamber", sa.String(), nullable=True),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_legislator")),
-        schema=schema,
-    )
-    op.create_index(op.f("ix_legislator_bioguide_id"), "legislator", ["bioguide_id"], unique=True, schema=schema)
-    op.create_table(
-        "bill_text",
-        sa.Column("bill_id", sa.Integer(), nullable=False),
-        sa.Column("version_code", sa.String(), nullable=False),
-        sa.Column("version_name", sa.String(), nullable=True),
-        sa.Column("text_content", sa.String(), nullable=True),
-        sa.Column("date", sa.Date(), nullable=True),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.ForeignKeyConstraint(
-            ["bill_id"], [f"{schema}.bill.id"], name=op.f("fk_bill_text_bill_id_bill"), ondelete="CASCADE"
-        ),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_bill_text")),
-        sa.UniqueConstraint("bill_id", "version_code", name="uq_bill_text_bill_id_version_code"),
-        schema=schema,
-    )
-    op.create_table(
-        "vote",
-        sa.Column("congress", sa.Integer(), nullable=False),
-        sa.Column("chamber", sa.String(), nullable=False),
-        sa.Column("session", sa.Integer(), nullable=False),
-        sa.Column("number", sa.Integer(), nullable=False),
-        sa.Column("vote_type", sa.String(), nullable=True),
-        sa.Column("question", sa.String(), nullable=True),
-        sa.Column("result", sa.String(), nullable=True),
-        sa.Column("result_text", sa.String(), nullable=True),
-        sa.Column("vote_date", sa.Date(), nullable=False),
-        sa.Column("yea_count", sa.Integer(), nullable=True),
-        sa.Column("nay_count", sa.Integer(), nullable=True),
-        sa.Column("not_voting_count", sa.Integer(), nullable=True),
-        sa.Column("present_count", sa.Integer(), nullable=True),
-        sa.Column("bill_id", sa.Integer(), nullable=True),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.ForeignKeyConstraint(["bill_id"], [f"{schema}.bill.id"], name=op.f("fk_vote_bill_id_bill")),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_vote")),
-        sa.UniqueConstraint("congress", "chamber", "session", "number", name="uq_vote_congress_chamber_session_number"),
-        schema=schema,
-    )
-    op.create_index("ix_vote_congress_chamber", "vote", ["congress", "chamber"], unique=False, schema=schema)
-    op.create_index("ix_vote_date", "vote", ["vote_date"], unique=False, schema=schema)
-    op.create_table(
-        "vote_record",
-        sa.Column("vote_id", sa.Integer(), nullable=False),
-        sa.Column("legislator_id", sa.Integer(), nullable=False),
-        sa.Column("position", sa.String(), nullable=False),
-        sa.ForeignKeyConstraint(
-            ["legislator_id"],
-            [f"{schema}.legislator.id"],
-            name=op.f("fk_vote_record_legislator_id_legislator"),
-            ondelete="CASCADE",
-        ),
-        sa.ForeignKeyConstraint(
-            ["vote_id"], [f"{schema}.vote.id"], name=op.f("fk_vote_record_vote_id_vote"), ondelete="CASCADE"
-        ),
-        sa.PrimaryKeyConstraint("vote_id", "legislator_id", name=op.f("pk_vote_record")),
-        schema=schema,
-    )
-    # ### end Alembic commands ###
-
-
-def downgrade() -> None:
-    """Downgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_table("vote_record", schema=schema)
-    op.drop_index("ix_vote_date", table_name="vote", schema=schema)
-    op.drop_index("ix_vote_congress_chamber", table_name="vote", schema=schema)
-    op.drop_table("vote", schema=schema)
-    op.drop_table("bill_text", schema=schema)
-    op.drop_index(op.f("ix_legislator_bioguide_id"), table_name="legislator", schema=schema)
-    op.drop_table("legislator", schema=schema)
-    op.drop_index("ix_bill_congress", table_name="bill", schema=schema)
-    op.drop_table("bill", schema=schema)
-    # ### end Alembic commands ###

python/alembic/data_science_dev/versions/2026_03_29-adding_legislatorsocialmedia_5cd7eee3549d.py

@@ -1,58 +0,0 @@
-"""adding LegislatorSocialMedia.
-
-Revision ID: 5cd7eee3549d
-Revises: 83bfc8af92d8
-Create Date: 2026-03-29 11:53:44.224799
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import sqlalchemy as sa
-from alembic import op
-
-from python.orm import DataScienceDevBase
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "5cd7eee3549d"
-down_revision: str | None = "83bfc8af92d8"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = DataScienceDevBase.schema_name
-
-
-def upgrade() -> None:
-    """Upgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.create_table(
-        "legislator_social_media",
-        sa.Column("legislator_id", sa.Integer(), nullable=False),
-        sa.Column("platform", sa.String(), nullable=False),
-        sa.Column("account_name", sa.String(), nullable=False),
-        sa.Column("url", sa.String(), nullable=True),
-        sa.Column("source", sa.String(), nullable=False),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.ForeignKeyConstraint(
-            ["legislator_id"],
-            [f"{schema}.legislator.id"],
-            name=op.f("fk_legislator_social_media_legislator_id_legislator"),
-        ),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_legislator_social_media")),
-        schema=schema,
-    )
-    # ### end Alembic commands ###
-
-
-def downgrade() -> None:
-    """Downgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_table("legislator_social_media", schema=schema)
-    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_03-adding_audiobook_libreary_metadata_d7864d1ffc17.py

@@ -0,0 +1,93 @@
+"""adding audiobook libreary metadata.
+
+Revision ID: d7864d1ffc17
+Revises: c8a794340928
+Create Date: 2026-06-03 20:24:09.200837
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import sqlalchemy as sa
+from alembic import op
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "d7864d1ffc17"
+down_revision: str | None = "c8a794340928"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "audiobook_author",
+        sa.Column("name", sa.String(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_audiobook_author")),
+        sa.UniqueConstraint("name", name=op.f("uq_audiobook_author_name")),
+        schema=schema,
+    )
+    op.create_table(
+        "audiobook_series",
+        sa.Column("name", sa.String(), nullable=False),
+        sa.Column("author_id", sa.Integer(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["author_id"],
+            [f"{schema}.audiobook_author.id"],
+            name=op.f("fk_audiobook_series_author_id_audiobook_author"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_audiobook_series")),
+        sa.UniqueConstraint("author_id", "name", name=op.f("uq_audiobook_series_author_id")),
+        schema=schema,
+    )
+    op.create_table(
+        "audiobook",
+        sa.Column("title", sa.String(), nullable=False),
+        sa.Column("author_id", sa.Integer(), nullable=False),
+        sa.Column("series_id", sa.Integer(), nullable=True),
+        sa.Column("series_index", sa.Integer(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["author_id"],
+            [f"{schema}.audiobook_author.id"],
+            name=op.f("fk_audiobook_author_id_audiobook_author"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["series_id"],
+            [f"{schema}.audiobook_series.id"],
+            name=op.f("fk_audiobook_series_id_audiobook_series"),
+            ondelete="SET NULL",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_audiobook")),
+        schema=schema,
+    )
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_table("audiobook", schema=schema)
+    op.drop_table("audiobook_series", schema=schema)
+    op.drop_table("audiobook_author", schema=schema)
+    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_10-add_ebook_search_tables_2db132cace1a.py

@@ -0,0 +1,200 @@
+"""add ebook search tables.
+
+Revision ID: 2db132cace1a
+Revises: b3c60cc5beb5
+Create Date: 2026-06-10 22:10:54.379159
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import pgvector
+import sqlalchemy as sa
+from alembic import op
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "2db132cace1a"
+down_revision: str | None = "b3c60cc5beb5"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "ebook_embedding_model",
+        sa.Column("name", sa.String(), nullable=False),
+        sa.Column("dimension", sa.Integer(), nullable=False),
+        sa.Column("is_default", sa.Boolean(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_embedding_model")),
+        sa.UniqueConstraint("name", name=op.f("uq_ebook_embedding_model_name")),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_source",
+        sa.Column("title", sa.String(), nullable=False),
+        sa.Column("author", sa.String(), nullable=True),
+        sa.Column("language", sa.String(), nullable=True),
+        sa.Column("publisher", sa.String(), nullable=True),
+        sa.Column("identifier", sa.String(), nullable=True),
+        sa.Column("file_path", sa.String(), nullable=False),
+        sa.Column("file_sha256", sa.String(length=64), nullable=False),
+        sa.Column("file_mtime", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("file_size", sa.BigInteger(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_source")),
+        sa.UniqueConstraint("file_path", name=op.f("uq_ebook_source_file_path")),
+        sa.UniqueConstraint("file_sha256", name=op.f("uq_ebook_source_file_sha256")),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_chapter",
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("spine_index", sa.Integer(), nullable=False),
+        sa.Column("title", sa.String(), nullable=True),
+        sa.Column("href", sa.String(), nullable=True),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["source_id"],
+            [f"{schema}.ebook_source.id"],
+            name=op.f("fk_ebook_chapter_source_id_ebook_source"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_chapter")),
+        sa.UniqueConstraint("source_id", "spine_index", name=op.f("uq_ebook_chapter_source_id")),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_chunk",
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("chapter_id", sa.Integer(), nullable=True),
+        sa.Column("chunk_index", sa.Integer(), nullable=False),
+        sa.Column("text", sa.String(), nullable=False),
+        sa.Column("token_start", sa.Integer(), nullable=False),
+        sa.Column("token_count", sa.Integer(), nullable=False),
+        sa.Column("page_label", sa.String(), nullable=True),
+        sa.Column("content_sha256", sa.String(length=64), nullable=False),
+        sa.Column("search_text", sa.String(), nullable=False),
+        sa.Column("id", sa.BigInteger(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["chapter_id"],
+            [f"{schema}.ebook_chapter.id"],
+            name=op.f("fk_ebook_chunk_chapter_id_ebook_chapter"),
+            ondelete="SET NULL",
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_id"],
+            [f"{schema}.ebook_source.id"],
+            name=op.f("fk_ebook_chunk_source_id_ebook_source"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_chunk")),
+        sa.UniqueConstraint("source_id", "chunk_index", name="uq_ebook_chunk_source_id_chunk_index"),
+        sa.UniqueConstraint("source_id", "content_sha256", name="uq_ebook_chunk_source_id_content_sha256"),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_chunk_embedding_1024",
+        sa.Column("chunk_id", sa.BigInteger(), nullable=False),
+        sa.Column("model_id", sa.Integer(), nullable=False),
+        sa.Column("embedding", pgvector.sqlalchemy.vector.VECTOR(dim=1024), nullable=False),
+        sa.Column("id", sa.BigInteger(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["chunk_id"],
+            [f"{schema}.ebook_chunk.id"],
+            name=op.f("fk_ebook_chunk_embedding_1024_chunk_id_ebook_chunk"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["model_id"],
+            [f"{schema}.ebook_embedding_model.id"],
+            name=op.f("fk_ebook_chunk_embedding_1024_model_id_ebook_embedding_model"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_chunk_embedding_1024")),
+        sa.UniqueConstraint("chunk_id", "model_id", name=op.f("uq_ebook_chunk_embedding_1024_chunk_id")),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_chunk_embedding_2560",
+        sa.Column("chunk_id", sa.BigInteger(), nullable=False),
+        sa.Column("model_id", sa.Integer(), nullable=False),
+        sa.Column("embedding", pgvector.sqlalchemy.vector.VECTOR(dim=2560), nullable=False),
+        sa.Column("id", sa.BigInteger(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["chunk_id"],
+            [f"{schema}.ebook_chunk.id"],
+            name=op.f("fk_ebook_chunk_embedding_2560_chunk_id_ebook_chunk"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["model_id"],
+            [f"{schema}.ebook_embedding_model.id"],
+            name=op.f("fk_ebook_chunk_embedding_2560_model_id_ebook_embedding_model"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_chunk_embedding_2560")),
+        sa.UniqueConstraint("chunk_id", "model_id", name=op.f("uq_ebook_chunk_embedding_2560_chunk_id")),
+        schema=schema,
+    )
+    op.create_table(
+        "ebook_chunk_embedding_4096",
+        sa.Column("chunk_id", sa.BigInteger(), nullable=False),
+        sa.Column("model_id", sa.Integer(), nullable=False),
+        sa.Column("embedding", pgvector.sqlalchemy.vector.VECTOR(dim=4096), nullable=False),
+        sa.Column("id", sa.BigInteger(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["chunk_id"],
+            [f"{schema}.ebook_chunk.id"],
+            name=op.f("fk_ebook_chunk_embedding_4096_chunk_id_ebook_chunk"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["model_id"],
+            [f"{schema}.ebook_embedding_model.id"],
+            name=op.f("fk_ebook_chunk_embedding_4096_model_id_ebook_embedding_model"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_ebook_chunk_embedding_4096")),
+        sa.UniqueConstraint("chunk_id", "model_id", name=op.f("uq_ebook_chunk_embedding_4096_chunk_id")),
+        schema=schema,
+    )
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_table("ebook_chunk_embedding_4096", schema=schema)
+    op.drop_table("ebook_chunk_embedding_2560", schema=schema)
+    op.drop_table("ebook_chunk_embedding_1024", schema=schema)
+    op.drop_table("ebook_chunk", schema=schema)
+    op.drop_table("ebook_chapter", schema=schema)
+    op.drop_table("ebook_source", schema=schema)
+    op.drop_table("ebook_embedding_model", schema=schema)
+    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_10-updated_series_index_to_float_and_added__b3c60cc5beb5.py

@@ -0,0 +1,63 @@
+"""updated series_index to float and added UniqueConstraint to audiobook and audiobook_author.
+
+Revision ID: b3c60cc5beb5
+Revises: d7864d1ffc17
+Create Date: 2026-06-10 20:02:43.073725
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import sqlalchemy as sa
+from alembic import op
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "b3c60cc5beb5"
+down_revision: str | None = "d7864d1ffc17"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.alter_column(
+        "audiobook",
+        "series_index",
+        existing_type=sa.INTEGER(),
+        type_=sa.Float(),
+        existing_nullable=False,
+        schema=schema,
+    )
+    op.create_unique_constraint(
+        op.f("uq_audiobook_author_id"),
+        "audiobook",
+        ["author_id", "series_id", "title"],
+        schema=schema,
+        postgresql_nulls_not_distinct=True,
+    )
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_constraint(op.f("uq_audiobook_author_id"), "audiobook", schema=schema, type_="unique")
+    op.alter_column(
+        "audiobook",
+        "series_index",
+        existing_type=sa.Float(),
+        type_=sa.INTEGER(),
+        existing_nullable=False,
+        schema=schema,
+    )
+    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_13-add_1024_ebook_embedding_cosine_index_c460105682d2.py

@@ -0,0 +1,54 @@
+"""add 1024 ebook embedding cosine index.
+
+Revision ID: c460105682d2
+Revises: 2db132cace1a
+Create Date: 2026-06-13 19:53:45.680289
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from alembic import op
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "c460105682d2"
+down_revision: str | None = "2db132cace1a"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_index(
+        "ix_ebook_chunk_embedding_1024_embedding_cosine",
+        "ebook_chunk_embedding_1024",
+        ["embedding"],
+        unique=False,
+        schema=schema,
+        postgresql_using="hnsw",
+        postgresql_ops={"embedding": "vector_cosine_ops"},
+    )
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_index(
+        "ix_ebook_chunk_embedding_1024_embedding_cosine",
+        table_name="ebook_chunk_embedding_1024",
+        schema=schema,
+        postgresql_using="hnsw",
+        postgresql_ops={"embedding": "vector_cosine_ops"},
+    )
+    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_23-adding_haproxy_data_96d72c748c24.py

@@ -0,0 +1,103 @@
+"""adding haproxy data.
+
+Revision ID: 96d72c748c24
+Revises: c460105682d2
+Create Date: 2026-06-23 16:37:17.768851
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import sqlalchemy as sa
+from alembic import op
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "96d72c748c24"
+down_revision: str | None = "c460105682d2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "haproxy_request",
+        sa.Column("line_hash", sa.String(), nullable=False),
+        sa.Column("requested_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("client_ip", sa.String(), nullable=False),
+        sa.Column("client_port", sa.Integer(), nullable=False),
+        sa.Column("frontend", sa.String(), nullable=False),
+        sa.Column("ssl", sa.Boolean(), nullable=False),
+        sa.Column("backend", sa.String(), nullable=False),
+        sa.Column("server", sa.String(), nullable=False),
+        sa.Column("time_request", sa.Integer(), nullable=False),
+        sa.Column("time_queue", sa.Integer(), nullable=False),
+        sa.Column("time_connect", sa.Integer(), nullable=False),
+        sa.Column("time_response", sa.Integer(), nullable=False),
+        sa.Column("time_total", sa.Integer(), nullable=False),
+        sa.Column("status_code", sa.Integer(), nullable=False),
+        sa.Column("bytes_read", sa.BigInteger(), nullable=False),
+        sa.Column("termination_state", sa.String(), nullable=False),
+        sa.Column("active_connections", sa.Integer(), nullable=False),
+        sa.Column("frontend_connections", sa.Integer(), nullable=False),
+        sa.Column("backend_connections", sa.Integer(), nullable=False),
+        sa.Column("server_connections", sa.Integer(), nullable=False),
+        sa.Column("retries", sa.Integer(), nullable=False),
+        sa.Column("server_queue", sa.Integer(), nullable=False),
+        sa.Column("backend_queue", sa.Integer(), nullable=False),
+        sa.Column("host", sa.String(), nullable=True),
+        sa.Column("user_agent", sa.String(), nullable=True),
+        sa.Column("method", sa.String(), nullable=False),
+        sa.Column("target", sa.String(), nullable=False),
+        sa.Column("path", sa.String(), nullable=False),
+        sa.Column("query", sa.String(), nullable=True),
+        sa.Column("http_version", sa.String(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_haproxy_request")),
+        sa.UniqueConstraint("line_hash", name=op.f("uq_haproxy_request_line_hash")),
+        schema=schema,
+    )
+    op.create_index(op.f("ix_haproxy_request_backend"), "haproxy_request", ["backend"], unique=False, schema=schema)
+    op.create_index(op.f("ix_haproxy_request_client_ip"), "haproxy_request", ["client_ip"], unique=False, schema=schema)
+    op.create_index(op.f("ix_haproxy_request_host"), "haproxy_request", ["host"], unique=False, schema=schema)
+    op.create_index(op.f("ix_haproxy_request_path"), "haproxy_request", ["path"], unique=False, schema=schema)
+    op.create_index(
+        op.f("ix_haproxy_request_requested_at"), "haproxy_request", ["requested_at"], unique=False, schema=schema
+    )
+    op.create_index(
+        op.f("ix_haproxy_request_status_code"), "haproxy_request", ["status_code"], unique=False, schema=schema
+    )
+    op.create_index(
+        op.f("ix_haproxy_request_time_response"), "haproxy_request", ["time_response"], unique=False, schema=schema
+    )
+    op.create_index(
+        op.f("ix_haproxy_request_user_agent"), "haproxy_request", ["user_agent"], unique=False, schema=schema
+    )
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_index(op.f("ix_haproxy_request_user_agent"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_time_response"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_status_code"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_requested_at"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_path"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_host"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_client_ip"), table_name="haproxy_request", schema=schema)
+    op.drop_index(op.f("ix_haproxy_request_backend"), table_name="haproxy_request", schema=schema)
+    op.drop_table("haproxy_request", schema=schema)
+    # ### end Alembic commands ###

python/alembic/richie/versions/2026_06_29-adding_phrase_metadata_tables_dddee09eddcc.py

@@ -0,0 +1,206 @@
+"""adding Phrase metadata tables.
+
+Revision ID: dddee09eddcc
+Revises: 96d72c748c24
+Create Date: 2026-06-29 00:49:07.344159
+
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+from python.orm import RichieBase
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+# revision identifiers, used by Alembic.
+revision: str = "dddee09eddcc"
+down_revision: str | None = "96d72c748c24"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+schema = RichieBase.schema_name
+
+
+def upgrade() -> None:
+    """Upgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "candidate_phrases",
+        sa.Column("book_id", sa.Integer(), nullable=False),
+        sa.Column("series_id", sa.Integer(), nullable=True),
+        sa.Column("phrase_text", sa.Text(), nullable=False),
+        sa.Column("phrase_norm", sa.Text(), nullable=False),
+        sa.Column("token_count", sa.Integer(), nullable=False),
+        sa.Column("source_raw_ngram", sa.Boolean(), nullable=False),
+        sa.Column("source_yake", sa.Boolean(), nullable=False),
+        sa.Column("source_spacy_ner", sa.Boolean(), nullable=False),
+        sa.Column("source_spacy_noun_chunk", sa.Boolean(), nullable=False),
+        sa.Column("source_capitalized", sa.Boolean(), nullable=False),
+        sa.Column("source_metadata", sa.Boolean(), nullable=False),
+        sa.Column("spacy_label", sa.String(), nullable=True),
+        sa.Column("raw_count", sa.Integer(), nullable=False),
+        sa.Column("chapter_count", sa.Integer(), nullable=False),
+        sa.Column("yake_score", sa.Float(), nullable=True),
+        sa.Column("candidate_score", sa.Float(), nullable=False),
+        sa.Column(
+            "sample_contexts",
+            sa.JSON().with_variant(postgresql.JSONB(astext_type=sa.Text()), "postgresql"),
+            nullable=True,
+        ),
+        sa.Column("llm_judged", sa.Boolean(), nullable=False),
+        sa.Column("llm_keep", sa.Boolean(), nullable=True),
+        sa.Column("llm_confidence", sa.Float(), nullable=True),
+        sa.Column("llm_category", sa.String(), nullable=True),
+        sa.Column("llm_reason", sa.Text(), nullable=True),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["book_id"],
+            [f"{schema}.ebook_source.id"],
+            name=op.f("fk_candidate_phrases_book_id_ebook_source"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_candidate_phrases")),
+        sa.UniqueConstraint("book_id", "phrase_norm", name="uq_candidate_phrases_book_id_phrase_norm"),
+        schema=schema,
+    )
+    op.create_index(
+        "candidate_phrases_book_norm_idx", "candidate_phrases", ["book_id", "phrase_norm"], unique=False, schema=schema
+    )
+    op.create_index(
+        "candidate_phrases_book_score_idx",
+        "candidate_phrases",
+        ["book_id", "candidate_score"],
+        unique=False,
+        schema=schema,
+    )
+    op.create_table(
+        "protected_phrases",
+        sa.Column("book_id", sa.Integer(), nullable=True),
+        sa.Column("series_id", sa.Integer(), nullable=True),
+        sa.Column("phrase_text", sa.Text(), nullable=False),
+        sa.Column("phrase_norm", sa.Text(), nullable=False),
+        sa.Column("canonical_id", sa.String(), nullable=False),
+        sa.Column("phrase_type", sa.String(), nullable=True),
+        sa.Column("token_count", sa.Integer(), nullable=False),
+        sa.Column("confidence", sa.Float(), nullable=False),
+        sa.Column("importance", sa.Float(), nullable=False),
+        sa.Column("allow_nested", sa.Boolean(), nullable=False),
+        sa.Column("suppress_children", sa.Boolean(), nullable=False),
+        sa.Column("source_candidate_id", sa.Integer(), nullable=True),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["book_id"],
+            [f"{schema}.ebook_source.id"],
+            name=op.f("fk_protected_phrases_book_id_ebook_source"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_candidate_id"],
+            [f"{schema}.candidate_phrases.id"],
+            name=op.f("fk_protected_phrases_source_candidate_id_candidate_phrases"),
+            ondelete="SET NULL",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_protected_phrases")),
+        sa.UniqueConstraint("book_id", "phrase_norm", name="uq_protected_phrases_book_id_phrase_norm"),
+        schema=schema,
+    )
+    op.create_index(
+        "protected_phrases_book_norm_idx", "protected_phrases", ["book_id", "phrase_norm"], unique=False, schema=schema
+    )
+    op.create_index("protected_phrases_norm_idx", "protected_phrases", ["phrase_norm"], unique=False, schema=schema)
+    op.create_index(
+        "protected_phrases_series_norm_idx",
+        "protected_phrases",
+        ["series_id", "phrase_norm"],
+        unique=False,
+        schema=schema,
+    )
+    op.create_table(
+        "chunk_phrase_mentions",
+        sa.Column("chunk_id", sa.BigInteger(), nullable=False),
+        sa.Column("phrase_id", sa.Integer(), nullable=False),
+        sa.Column("book_id", sa.Integer(), nullable=True),
+        sa.Column("series_id", sa.Integer(), nullable=True),
+        sa.Column("start_char", sa.Integer(), nullable=False),
+        sa.Column("end_char", sa.Integer(), nullable=True),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["book_id"],
+            [f"{schema}.ebook_source.id"],
+            name=op.f("fk_chunk_phrase_mentions_book_id_ebook_source"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["chunk_id"],
+            [f"{schema}.ebook_chunk.id"],
+            name=op.f("fk_chunk_phrase_mentions_chunk_id_ebook_chunk"),
+            ondelete="CASCADE",
+        ),
+        sa.ForeignKeyConstraint(
+            ["phrase_id"],
+            [f"{schema}.protected_phrases.id"],
+            name=op.f("fk_chunk_phrase_mentions_phrase_id_protected_phrases"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_chunk_phrase_mentions")),
+        sa.UniqueConstraint("chunk_id", "phrase_id", "start_char", name="uq_chunk_phrase_mentions_chunk_phrase_start"),
+        schema=schema,
+    )
+    op.create_index(
+        "chunk_phrase_mentions_chunk_idx", "chunk_phrase_mentions", ["chunk_id"], unique=False, schema=schema
+    )
+    op.create_index(
+        "chunk_phrase_mentions_phrase_idx", "chunk_phrase_mentions", ["phrase_id"], unique=False, schema=schema
+    )
+    op.create_table(
+        "phrase_aliases",
+        sa.Column("phrase_id", sa.Integer(), nullable=False),
+        sa.Column("alias_text", sa.Text(), nullable=False),
+        sa.Column("alias_norm", sa.Text(), nullable=False),
+        sa.Column("confidence", sa.Float(), nullable=False),
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["phrase_id"],
+            [f"{schema}.protected_phrases.id"],
+            name=op.f("fk_phrase_aliases_phrase_id_protected_phrases"),
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_phrase_aliases")),
+        sa.UniqueConstraint("phrase_id", "alias_norm", name="uq_phrase_aliases_phrase_id_alias_norm"),
+        schema=schema,
+    )
+    op.create_index("phrase_aliases_norm_idx", "phrase_aliases", ["alias_norm"], unique=False, schema=schema)
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    """Downgrade."""
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.drop_index("phrase_aliases_norm_idx", table_name="phrase_aliases", schema=schema)
+    op.drop_table("phrase_aliases", schema=schema)
+    op.drop_index("chunk_phrase_mentions_phrase_idx", table_name="chunk_phrase_mentions", schema=schema)
+    op.drop_index("chunk_phrase_mentions_chunk_idx", table_name="chunk_phrase_mentions", schema=schema)
+    op.drop_table("chunk_phrase_mentions", schema=schema)
+    op.drop_index("protected_phrases_series_norm_idx", table_name="protected_phrases", schema=schema)
+    op.drop_index("protected_phrases_norm_idx", table_name="protected_phrases", schema=schema)
+    op.drop_index("protected_phrases_book_norm_idx", table_name="protected_phrases", schema=schema)
+    op.drop_table("protected_phrases", schema=schema)
+    op.drop_index("candidate_phrases_book_score_idx", table_name="candidate_phrases", schema=schema)
+    op.drop_index("candidate_phrases_book_norm_idx", table_name="candidate_phrases", schema=schema)
+    op.drop_table("candidate_phrases", schema=schema)
+    # ### end Alembic commands ###

python/alembic/signal_bot/versions/2026_03_17-seprating_signal_bot_database_6eaf696e07a5.py

@@ -1,100 +0,0 @@
-"""seprating signal_bot database.
-
-Revision ID: 6eaf696e07a5
-Revises:
-Create Date: 2026-03-17 21:35:37.612672
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import sqlalchemy as sa
-from alembic import op
-from sqlalchemy.dialects import postgresql
-
-from python.orm import SignalBotBase
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "6eaf696e07a5"
-down_revision: str | None = None
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = SignalBotBase.schema_name
-
-
-def upgrade() -> None:
-    """Upgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.create_table(
-        "dead_letter_message",
-        sa.Column("source", sa.String(), nullable=False),
-        sa.Column("message", sa.Text(), nullable=False),
-        sa.Column("received_at", sa.DateTime(timezone=True), nullable=False),
-        sa.Column(
-            "status", postgresql.ENUM("UNPROCESSED", "PROCESSED", name="message_status", schema=schema), nullable=False
-        ),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_dead_letter_message")),
-        schema=schema,
-    )
-    op.create_table(
-        "role",
-        sa.Column("name", sa.String(length=50), nullable=False),
-        sa.Column("id", sa.SmallInteger(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_role")),
-        sa.UniqueConstraint("name", name=op.f("uq_role_name")),
-        schema=schema,
-    )
-    op.create_table(
-        "signal_device",
-        sa.Column("phone_number", sa.String(length=50), nullable=False),
-        sa.Column("safety_number", sa.String(), nullable=True),
-        sa.Column(
-            "trust_level",
-            postgresql.ENUM("VERIFIED", "UNVERIFIED", "BLOCKED", name="trust_level", schema=schema),
-            nullable=False,
-        ),
-        sa.Column("last_seen", sa.DateTime(timezone=True), nullable=False),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_signal_device")),
-        sa.UniqueConstraint("phone_number", name=op.f("uq_signal_device_phone_number")),
-        schema=schema,
-    )
-    op.create_table(
-        "device_role",
-        sa.Column("device_id", sa.Integer(), nullable=False),
-        sa.Column("role_id", sa.SmallInteger(), nullable=False),
-        sa.Column("id", sa.Integer(), nullable=False),
-        sa.Column("created", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.Column("updated", sa.DateTime(timezone=True), server_default=sa.text("now()"), nullable=False),
-        sa.ForeignKeyConstraint(
-            ["device_id"], [f"{schema}.signal_device.id"], name=op.f("fk_device_role_device_id_signal_device")
-        ),
-        sa.ForeignKeyConstraint(["role_id"], [f"{schema}.role.id"], name=op.f("fk_device_role_role_id_role")),
-        sa.PrimaryKeyConstraint("id", name=op.f("pk_device_role")),
-        sa.UniqueConstraint("device_id", "role_id", name="uq_device_role_device_role"),
-        schema=schema,
-    )
-    # ### end Alembic commands ###
-
-
-def downgrade() -> None:
-    """Downgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_table("device_role", schema=schema)
-    op.drop_table("signal_device", schema=schema)
-    op.drop_table("role", schema=schema)
-    op.drop_table("dead_letter_message", schema=schema)
-    # ### end Alembic commands ###

python/alembic/signal_bot/versions/2026_03_18-test_66bdd532bcab.py

@@ -1,72 +0,0 @@
-"""test.
-
-Revision ID: 66bdd532bcab
-Revises: 6eaf696e07a5
-Create Date: 2026-03-18 19:21:14.561568
-
-"""
-
-from __future__ import annotations
-
-from typing import TYPE_CHECKING
-
-import sqlalchemy as sa
-from alembic import op
-from sqlalchemy.dialects import postgresql
-
-from python.orm import SignalBotBase
-
-if TYPE_CHECKING:
-    from collections.abc import Sequence
-
-# revision identifiers, used by Alembic.
-revision: str = "66bdd532bcab"
-down_revision: str | None = "6eaf696e07a5"
-branch_labels: str | Sequence[str] | None = None
-depends_on: str | Sequence[str] | None = None
-
-schema = SignalBotBase.schema_name
-
-
-def upgrade() -> None:
-    """Upgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.alter_column(
-        "dead_letter_message",
-        "status",
-        existing_type=postgresql.ENUM("UNPROCESSED", "PROCESSED", name="message_status", schema=schema),
-        type_=sa.Enum("UNPROCESSED", "PROCESSED", name="message_status", native_enum=False),
-        existing_nullable=False,
-        schema=schema,
-    )
-    op.alter_column(
-        "signal_device",
-        "trust_level",
-        existing_type=postgresql.ENUM("VERIFIED", "UNVERIFIED", "BLOCKED", name="trust_level", schema=schema),
-        type_=sa.Enum("VERIFIED", "UNVERIFIED", "BLOCKED", name="trust_level", native_enum=False),
-        existing_nullable=False,
-        schema=schema,
-    )
-    # ### end Alembic commands ###
-
-
-def downgrade() -> None:
-    """Downgrade."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.alter_column(
-        "signal_device",
-        "trust_level",
-        existing_type=sa.Enum("VERIFIED", "UNVERIFIED", "BLOCKED", name="trust_level", native_enum=False),
-        type_=postgresql.ENUM("VERIFIED", "UNVERIFIED", "BLOCKED", name="trust_level", schema=schema),
-        existing_nullable=False,
-        schema=schema,
-    )
-    op.alter_column(
-        "dead_letter_message",
-        "status",
-        existing_type=sa.Enum("UNPROCESSED", "PROCESSED", name="message_status", native_enum=False),
-        type_=postgresql.ENUM("UNPROCESSED", "PROCESSED", name="message_status", schema=schema),
-        existing_nullable=False,
-        schema=schema,
-    )
-    # ### end Alembic commands ###

python/api/dependencies.py

@@ -1,16 +0,0 @@
-"""FastAPI dependencies."""
-
-from collections.abc import Iterator
-from typing import Annotated
-
-from fastapi import Depends, Request
-from sqlalchemy.orm import Session
-
-
-def get_db(request: Request) -> Iterator[Session]:
-    """Get database session from app state."""
-    with Session(request.app.state.engine) as session:
-        yield session
-
-
-DbSession = Annotated[Session, Depends(get_db)]

python/api/main.py

@@ -1,19 +1,23 @@
 """FastAPI interface for Contact database."""
 
+from __future__ import annotations
+
 import logging
-from collections.abc import AsyncIterator
 from contextlib import asynccontextmanager
-from typing import Annotated
+from typing import TYPE_CHECKING, Annotated
 
 import typer
 import uvicorn
 from fastapi import FastAPI
 
-from python.api.middleware import ZstdMiddleware
 from python.api.routers import contact_router, views_router
 from python.common import configure_logger
+from python.fastapi_tools import ZstdMiddleware
 from python.orm.common import get_postgres_engine
 
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator
+
 logger = logging.getLogger(__name__)
 
 

python/api/routers/contact.py

@@ -9,7 +9,7 @@ from pydantic import BaseModel
 from sqlalchemy import select
 from sqlalchemy.orm import selectinload
 
-from python.api.dependencies import DbSession
+from python.fastapi_tools.db import DbSession  # noqa: TC001 this is a FastAPI needed at runtime
 from python.orm.richie.contact import Contact, ContactRelationship, Need, RelationshipType
 
 TEMPLATES_DIR = Path(__file__).parent.parent / "templates"

python/api/routers/views.py

@@ -9,7 +9,7 @@ from fastapi.templating import Jinja2Templates
 from sqlalchemy import select
 from sqlalchemy.orm import Session, selectinload
 
-from python.api.dependencies import DbSession
+from python.fastapi_tools.db import DbSession  # noqa: TC001 this is a FastAPI needed at runtime
 from python.orm.richie.contact import Contact, ContactRelationship, Need, RelationshipType
 
 TEMPLATES_DIR = Path(__file__).parent.parent / "templates"

python/common.py

@@ -3,28 +3,23 @@
 from __future__ import annotations
 
 import logging
-import sys
 from datetime import UTC, datetime
-from os import getenv
+from pathlib import Path
 from subprocess import PIPE, Popen
 
-from apprise import Apprise
+from python.logging_config import configure_logger as _configure_logger
 
 logger = logging.getLogger(__name__)
 
 
-def configure_logger(level: str = "INFO") -> None:
+def get_repo_dir() -> Path:
-    """Configure the logger.
+    """Return the repository root directory."""
+    return Path(__file__).resolve().parents[1]
 
-    Args:
+
-        level (str, optional): The logging level. Defaults to "INFO".
+def configure_logger(level: str = "INFO") -> None:
-    """
+    """Configure the logger."""
-    logging.basicConfig(
+    _configure_logger(level)
-        level=level,
-        datefmt="%Y-%m-%dT%H:%M:%S%z",
-        format="%(asctime)s %(levelname)s %(filename)s:%(lineno)d - %(message)s",
-        handlers=[logging.StreamHandler(sys.stdout)],
-    )
 
 
 def bash_wrapper(command: str) -> tuple[str, int]:
@@ -47,26 +42,6 @@ def bash_wrapper(command: str) -> tuple[str, int]:
     return output.decode(), process.returncode
 
 
-def signal_alert(body: str, title: str = "") -> None:
-    """Send a signal alert.
-
-    Args:
-        body (str): The body of the alert.
-        title (str, optional): The title of the alert. Defaults to "".
-    """
-    apprise_client = Apprise()
-
-    from_phone = getenv("SIGNAL_ALERT_FROM_PHONE")
-    to_phone = getenv("SIGNAL_ALERT_TO_PHONE")
-    if not from_phone or not to_phone:
-        logger.info("SIGNAL_ALERT_FROM_PHONE or SIGNAL_ALERT_TO_PHONE not set")
-        return
-
-    apprise_client.add(f"signal://localhost:8989/{from_phone}/{to_phone}")
-
-    apprise_client.notify(title=title, body=body)
-
-
 def utcnow() -> datetime:
     """Get the current UTC time."""
     return datetime.now(tz=UTC)

python/data_science/__init__.py

@@ -1,3 +0,0 @@
-"""Data science CLI tools."""
-
-from __future__ import annotations

python/data_science/ingest_congress.py

@@ -1,613 +0,0 @@
-"""Ingestion pipeline for loading congress data from unitedstates/congress JSON files.
-
-Loads legislators, bills, votes, vote records, and bill text into the data_science_dev database.
-Expects the parent directory to contain congress-tracker/ and congress-legislators/ as siblings.
-
-Usage:
-    ingest-congress /path/to/parent/
-    ingest-congress /path/to/parent/ --congress 118
-    ingest-congress /path/to/parent/ --congress 118 --only bills
-"""
-
-from __future__ import annotations
-
-import logging
-from pathlib import Path  # noqa: TC003 needed at runtime for typer CLI argument
-from typing import TYPE_CHECKING, Annotated
-
-import orjson
-import typer
-import yaml
-from sqlalchemy import select
-from sqlalchemy.orm import Session
-
-from python.common import configure_logger
-from python.orm.common import get_postgres_engine
-from python.orm.data_science_dev.congress import Bill, BillText, Legislator, LegislatorSocialMedia, Vote, VoteRecord
-
-if TYPE_CHECKING:
-    from collections.abc import Iterator
-
-    from sqlalchemy.engine import Engine
-
-logger = logging.getLogger(__name__)
-
-BATCH_SIZE = 10_000
-
-app = typer.Typer(help="Ingest unitedstates/congress data into data_science_dev.")
-
-
-@app.command()
-def main(
-    parent_dir: Annotated[
-        Path,
-        typer.Argument(help="Parent directory containing congress-tracker/ and congress-legislators/"),
-    ],
-    congress: Annotated[int | None, typer.Option(help="Only ingest a specific congress number")] = None,
-    only: Annotated[
-        str | None,
-        typer.Option(help="Only run a specific step: legislators, social-media, bills, votes, bill-text"),
-    ] = None,
-) -> None:
-    """Ingest congress data from unitedstates/congress JSON files."""
-    configure_logger(level="INFO")
-
-    data_dir = parent_dir / "congress-tracker/congress/data/"
-    legislators_dir = parent_dir / "congress-legislators"
-
-    if not data_dir.is_dir():
-        typer.echo(f"Expected congress-tracker/ directory: {data_dir}", err=True)
-        raise typer.Exit(code=1)
-
-    if not legislators_dir.is_dir():
-        typer.echo(f"Expected congress-legislators/ directory: {legislators_dir}", err=True)
-        raise typer.Exit(code=1)
-
-    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
-
-    congress_dirs = _resolve_congress_dirs(data_dir, congress)
-    if not congress_dirs:
-        typer.echo("No congress directories found.", err=True)
-        raise typer.Exit(code=1)
-
-    logger.info("Found %d congress directories to process", len(congress_dirs))
-
-    steps: dict[str, tuple] = {
-        "legislators": (ingest_legislators, (engine, legislators_dir)),
-        "legislators-social-media": (ingest_social_media, (engine, legislators_dir)),
-        "bills": (ingest_bills, (engine, congress_dirs)),
-        "votes": (ingest_votes, (engine, congress_dirs)),
-        "bill-text": (ingest_bill_text, (engine, congress_dirs)),
-    }
-
-    if only:
-        if only not in steps:
-            typer.echo(f"Unknown step: {only}. Choose from: {', '.join(steps)}", err=True)
-            raise typer.Exit(code=1)
-        steps = {only: steps[only]}
-
-    for step_name, (step_func, step_args) in steps.items():
-        logger.info("=== Starting step: %s ===", step_name)
-        step_func(*step_args)
-        logger.info("=== Finished step: %s ===", step_name)
-
-    logger.info("ingest-congress done")
-
-
-def _resolve_congress_dirs(data_dir: Path, congress: int | None) -> list[Path]:
-    """Find congress number directories under data_dir."""
-    if congress is not None:
-        target = data_dir / str(congress)
-        return [target] if target.is_dir() else []
-    return sorted(path for path in data_dir.iterdir() if path.is_dir() and path.name.isdigit())
-
-
-def _flush_batch(session: Session, batch: list[object], label: str) -> int:
-    """Add a batch of ORM objects to the session and commit. Returns count added."""
-    if not batch:
-        return 0
-    session.add_all(batch)
-    session.commit()
-    count = len(batch)
-    logger.info("Committed %d %s", count, label)
-    batch.clear()
-    return count
-
-
-# ---------------------------------------------------------------------------
-# Legislators — loaded from congress-legislators YAML files
-# ---------------------------------------------------------------------------
-
-
-def ingest_legislators(engine: Engine, legislators_dir: Path) -> None:
-    """Load legislators from congress-legislators YAML files."""
-    legislators_data = _load_legislators_yaml(legislators_dir)
-    logger.info("Loaded %d legislators from YAML files", len(legislators_data))
-
-    with Session(engine) as session:
-        existing_legislators = {
-            legislator.bioguide_id: legislator for legislator in session.scalars(select(Legislator)).all()
-        }
-        logger.info("Found %d existing legislators in DB", len(existing_legislators))
-
-        total_inserted = 0
-        total_updated = 0
-        for entry in legislators_data:
-            bioguide_id = entry.get("id", {}).get("bioguide")
-            if not bioguide_id:
-                continue
-
-            fields = _parse_legislator(entry)
-            if existing := existing_legislators.get(bioguide_id):
-                changed = False
-                for field, value in fields.items():
-                    if value is not None and getattr(existing, field) != value:
-                        setattr(existing, field, value)
-                        changed = True
-                if changed:
-                    total_updated += 1
-            else:
-                session.add(Legislator(bioguide_id=bioguide_id, **fields))
-                total_inserted += 1
-
-        session.commit()
-    logger.info("Inserted %d new legislators, updated %d existing", total_inserted, total_updated)
-
-
-def _load_legislators_yaml(legislators_dir: Path) -> list[dict]:
-    """Load and combine legislators-current.yaml and legislators-historical.yaml."""
-    legislators: list[dict] = []
-    for filename in ("legislators-current.yaml", "legislators-historical.yaml"):
-        path = legislators_dir / filename
-        if not path.exists():
-            logger.warning("Legislators file not found: %s", path)
-            continue
-        with path.open() as file:
-            data = yaml.safe_load(file)
-            if isinstance(data, list):
-                legislators.extend(data)
-    return legislators
-
-
-def _parse_legislator(entry: dict) -> dict:
-    """Extract Legislator fields from a congress-legislators YAML entry."""
-    ids = entry.get("id", {})
-    name = entry.get("name", {})
-    bio = entry.get("bio", {})
-    terms = entry.get("terms", [])
-    latest_term = terms[-1] if terms else {}
-
-    fec_ids = ids.get("fec")
-    fec_ids_joined = ",".join(fec_ids) if isinstance(fec_ids, list) else fec_ids
-
-    chamber = latest_term.get("type")
-    chamber_normalized = {"rep": "House", "sen": "Senate"}.get(chamber, chamber)
-
-    return {
-        "thomas_id": ids.get("thomas"),
-        "lis_id": ids.get("lis"),
-        "govtrack_id": ids.get("govtrack"),
-        "opensecrets_id": ids.get("opensecrets"),
-        "fec_ids": fec_ids_joined,
-        "first_name": name.get("first"),
-        "last_name": name.get("last"),
-        "official_full_name": name.get("official_full"),
-        "nickname": name.get("nickname"),
-        "birthday": bio.get("birthday"),
-        "gender": bio.get("gender"),
-        "current_party": latest_term.get("party"),
-        "current_state": latest_term.get("state"),
-        "current_district": latest_term.get("district"),
-        "current_chamber": chamber_normalized,
-    }
-
-
-# ---------------------------------------------------------------------------
-# Social Media — loaded from legislators-social-media.yaml
-# ---------------------------------------------------------------------------
-
-SOCIAL_MEDIA_PLATFORMS = {
-    "twitter": "https://twitter.com/{account}",
-    "facebook": "https://facebook.com/{account}",
-    "youtube": "https://youtube.com/{account}",
-    "instagram": "https://instagram.com/{account}",
-    "mastodon": None,
-}
-
-
-def ingest_social_media(engine: Engine, legislators_dir: Path) -> None:
-    """Load social media accounts from legislators-social-media.yaml."""
-    social_media_path = legislators_dir / "legislators-social-media.yaml"
-    if not social_media_path.exists():
-        logger.warning("Social media file not found: %s", social_media_path)
-        return
-
-    with social_media_path.open() as file:
-        social_media_data = yaml.safe_load(file)
-
-    if not isinstance(social_media_data, list):
-        logger.warning("Unexpected format in %s", social_media_path)
-        return
-
-    logger.info("Loaded %d entries from legislators-social-media.yaml", len(social_media_data))
-
-    with Session(engine) as session:
-        legislator_map = _build_legislator_map(session)
-        existing_accounts = {
-            (account.legislator_id, account.platform)
-            for account in session.scalars(select(LegislatorSocialMedia)).all()
-        }
-        logger.info("Found %d existing social media accounts in DB", len(existing_accounts))
-
-        total_inserted = 0
-        total_updated = 0
-        for entry in social_media_data:
-            bioguide_id = entry.get("id", {}).get("bioguide")
-            if not bioguide_id:
-                continue
-
-            legislator_id = legislator_map.get(bioguide_id)
-            if legislator_id is None:
-                continue
-
-            social = entry.get("social", {})
-            for platform, url_template in SOCIAL_MEDIA_PLATFORMS.items():
-                account_name = social.get(platform)
-                if not account_name:
-                    continue
-
-                url = url_template.format(account=account_name) if url_template else None
-
-                if (legislator_id, platform) in existing_accounts:
-                    total_updated += 1
-                else:
-                    session.add(
-                        LegislatorSocialMedia(
-                            legislator_id=legislator_id,
-                            platform=platform,
-                            account_name=str(account_name),
-                            url=url,
-                            source="https://github.com/unitedstates/congress-legislators",
-                        )
-                    )
-                    existing_accounts.add((legislator_id, platform))
-                    total_inserted += 1
-
-        session.commit()
-    logger.info("Inserted %d new social media accounts, updated %d existing", total_inserted, total_updated)
-
-
-def _iter_voters(position_group: object) -> Iterator[dict]:
-    """Yield voter dicts from a vote position group (handles list, single dict, or string)."""
-    if isinstance(position_group, dict):
-        yield position_group
-    elif isinstance(position_group, list):
-        for voter in position_group:
-            if isinstance(voter, dict):
-                yield voter
-
-
-# ---------------------------------------------------------------------------
-# Bills
-# ---------------------------------------------------------------------------
-
-
-def ingest_bills(engine: Engine, congress_dirs: list[Path]) -> None:
-    """Load bill data.json files."""
-    with Session(engine) as session:
-        existing_bills = {(bill.congress, bill.bill_type, bill.number) for bill in session.scalars(select(Bill)).all()}
-        logger.info("Found %d existing bills in DB", len(existing_bills))
-
-        total_inserted = 0
-        batch: list[Bill] = []
-        for congress_dir in congress_dirs:
-            bills_dir = congress_dir / "bills"
-            if not bills_dir.is_dir():
-                continue
-            logger.info("Scanning bills from %s", congress_dir.name)
-            for bill_file in bills_dir.rglob("data.json"):
-                data = _read_json(bill_file)
-                if data is None:
-                    continue
-                bill = _parse_bill(data, existing_bills)
-                if bill is not None:
-                    batch.append(bill)
-                    if len(batch) >= BATCH_SIZE:
-                        total_inserted += _flush_batch(session, batch, "bills")
-
-        total_inserted += _flush_batch(session, batch, "bills")
-    logger.info("Inserted %d new bills total", total_inserted)
-
-
-def _parse_bill(data: dict, existing_bills: set[tuple[int, str, int]]) -> Bill | None:
-    """Parse a bill data.json dict into a Bill ORM object, skipping existing."""
-    raw_congress = data.get("congress")
-    bill_type = data.get("bill_type")
-    raw_number = data.get("number")
-    if raw_congress is None or bill_type is None or raw_number is None:
-        return None
-    congress = int(raw_congress)
-    number = int(raw_number)
-    if (congress, bill_type, number) in existing_bills:
-        return None
-
-    sponsor_bioguide = None
-    sponsor = data.get("sponsor")
-    if sponsor:
-        sponsor_bioguide = sponsor.get("bioguide_id")
-
-    return Bill(
-        congress=congress,
-        bill_type=bill_type,
-        number=number,
-        title=data.get("short_title") or data.get("official_title"),
-        title_short=data.get("short_title"),
-        official_title=data.get("official_title"),
-        status=data.get("status"),
-        status_at=data.get("status_at"),
-        sponsor_bioguide_id=sponsor_bioguide,
-        subjects_top_term=data.get("subjects_top_term"),
-    )
-
-
-# ---------------------------------------------------------------------------
-# Votes (and vote records)
-# ---------------------------------------------------------------------------
-
-
-def ingest_votes(engine: Engine, congress_dirs: list[Path]) -> None:
-    """Load vote data.json files with their vote records."""
-    with Session(engine) as session:
-        legislator_map = _build_legislator_map(session)
-        logger.info("Loaded %d legislators into lookup map", len(legislator_map))
-        bill_map = _build_bill_map(session)
-        logger.info("Loaded %d bills into lookup map", len(bill_map))
-        existing_votes = {
-            (vote.congress, vote.chamber, vote.session, vote.number) for vote in session.scalars(select(Vote)).all()
-        }
-        logger.info("Found %d existing votes in DB", len(existing_votes))
-
-        total_inserted = 0
-        batch: list[Vote] = []
-        for congress_dir in congress_dirs:
-            votes_dir = congress_dir / "votes"
-            if not votes_dir.is_dir():
-                continue
-            logger.info("Scanning votes from %s", congress_dir.name)
-            for vote_file in votes_dir.rglob("data.json"):
-                data = _read_json(vote_file)
-                if data is None:
-                    continue
-                vote = _parse_vote(data, legislator_map, bill_map, existing_votes)
-                if vote is not None:
-                    batch.append(vote)
-                    if len(batch) >= BATCH_SIZE:
-                        total_inserted += _flush_batch(session, batch, "votes")
-
-        total_inserted += _flush_batch(session, batch, "votes")
-    logger.info("Inserted %d new votes total", total_inserted)
-
-
-def _build_legislator_map(session: Session) -> dict[str, int]:
-    """Build a mapping of bioguide_id -> legislator.id."""
-    return {legislator.bioguide_id: legislator.id for legislator in session.scalars(select(Legislator)).all()}
-
-
-def _build_bill_map(session: Session) -> dict[tuple[int, str, int], int]:
-    """Build a mapping of (congress, bill_type, number) -> bill.id."""
-    return {(bill.congress, bill.bill_type, bill.number): bill.id for bill in session.scalars(select(Bill)).all()}
-
-
-def _parse_vote(
-    data: dict,
-    legislator_map: dict[str, int],
-    bill_map: dict[tuple[int, str, int], int],
-    existing_votes: set[tuple[int, str, int, int]],
-) -> Vote | None:
-    """Parse a vote data.json dict into a Vote ORM object with records."""
-    raw_congress = data.get("congress")
-    chamber = data.get("chamber")
-    raw_number = data.get("number")
-    vote_date = data.get("date")
-    if raw_congress is None or chamber is None or raw_number is None or vote_date is None:
-        return None
-
-    raw_session = data.get("session")
-    if raw_session is None:
-        return None
-
-    congress = int(raw_congress)
-    number = int(raw_number)
-    session_number = int(raw_session)
-
-    # Normalize chamber from "h"/"s" to "House"/"Senate"
-    chamber_normalized = {"h": "House", "s": "Senate"}.get(chamber, chamber)
-
-    if (congress, chamber_normalized, session_number, number) in existing_votes:
-        return None
-
-    # Resolve linked bill
-    bill_id = None
-    bill_ref = data.get("bill")
-    if bill_ref:
-        bill_key = (
-            int(bill_ref.get("congress", congress)),
-            bill_ref.get("type"),
-            int(bill_ref.get("number", 0)),
-        )
-        bill_id = bill_map.get(bill_key)
-
-    raw_votes = data.get("votes", {})
-    vote_counts = _count_votes(raw_votes)
-    vote_records = _build_vote_records(raw_votes, legislator_map)
-
-    return Vote(
-        congress=congress,
-        chamber=chamber_normalized,
-        session=session_number,
-        number=number,
-        vote_type=data.get("type"),
-        question=data.get("question"),
-        result=data.get("result"),
-        result_text=data.get("result_text"),
-        vote_date=vote_date[:10] if isinstance(vote_date, str) else vote_date,
-        bill_id=bill_id,
-        vote_records=vote_records,
-        **vote_counts,
-    )
-
-
-def _count_votes(raw_votes: dict) -> dict[str, int]:
-    """Count voters per position category, correctly handling dict and list formats."""
-    yea_count = 0
-    nay_count = 0
-    not_voting_count = 0
-    present_count = 0
-
-    for position, position_group in raw_votes.items():
-        voter_count = sum(1 for _ in _iter_voters(position_group))
-        if position in ("Yea", "Aye"):
-            yea_count += voter_count
-        elif position in ("Nay", "No"):
-            nay_count += voter_count
-        elif position == "Not Voting":
-            not_voting_count += voter_count
-        elif position == "Present":
-            present_count += voter_count
-
-    return {
-        "yea_count": yea_count,
-        "nay_count": nay_count,
-        "not_voting_count": not_voting_count,
-        "present_count": present_count,
-    }
-
-
-def _build_vote_records(raw_votes: dict, legislator_map: dict[str, int]) -> list[VoteRecord]:
-    """Build VoteRecord objects from raw vote data."""
-    records: list[VoteRecord] = []
-    for position, position_group in raw_votes.items():
-        for voter in _iter_voters(position_group):
-            bioguide_id = voter.get("id")
-            if not bioguide_id:
-                continue
-            legislator_id = legislator_map.get(bioguide_id)
-            if legislator_id is None:
-                continue
-            records.append(
-                VoteRecord(
-                    legislator_id=legislator_id,
-                    position=position,
-                )
-            )
-    return records
-
-
-# ---------------------------------------------------------------------------
-# Bill Text
-# ---------------------------------------------------------------------------
-
-
-def ingest_bill_text(engine: Engine, congress_dirs: list[Path]) -> None:
-    """Load bill text from text-versions directories."""
-    with Session(engine) as session:
-        bill_map = _build_bill_map(session)
-        logger.info("Loaded %d bills into lookup map", len(bill_map))
-        existing_bill_texts = {
-            (bill_text.bill_id, bill_text.version_code) for bill_text in session.scalars(select(BillText)).all()
-        }
-        logger.info("Found %d existing bill text versions in DB", len(existing_bill_texts))
-
-        total_inserted = 0
-        batch: list[BillText] = []
-        for congress_dir in congress_dirs:
-            logger.info("Scanning bill texts from %s", congress_dir.name)
-            for bill_text in _iter_bill_texts(congress_dir, bill_map, existing_bill_texts):
-                batch.append(bill_text)
-                if len(batch) >= BATCH_SIZE:
-                    total_inserted += _flush_batch(session, batch, "bill texts")
-
-        total_inserted += _flush_batch(session, batch, "bill texts")
-    logger.info("Inserted %d new bill text versions total", total_inserted)
-
-
-def _iter_bill_texts(
-    congress_dir: Path,
-    bill_map: dict[tuple[int, str, int], int],
-    existing_bill_texts: set[tuple[int, str]],
-) -> Iterator[BillText]:
-    """Yield BillText objects for a single congress directory, skipping existing."""
-    bills_dir = congress_dir / "bills"
-    if not bills_dir.is_dir():
-        return
-
-    for bill_dir in bills_dir.rglob("text-versions"):
-        if not bill_dir.is_dir():
-            continue
-        bill_key = _bill_key_from_dir(bill_dir.parent, congress_dir)
-        if bill_key is None:
-            continue
-        bill_id = bill_map.get(bill_key)
-        if bill_id is None:
-            continue
-
-        for version_dir in sorted(bill_dir.iterdir()):
-            if not version_dir.is_dir():
-                continue
-            if (bill_id, version_dir.name) in existing_bill_texts:
-                continue
-            text_content = _read_bill_text(version_dir)
-            version_data = _read_json(version_dir / "data.json")
-            yield BillText(
-                bill_id=bill_id,
-                version_code=version_dir.name,
-                version_name=version_data.get("version_name") if version_data else None,
-                date=version_data.get("issued_on") if version_data else None,
-                text_content=text_content,
-            )
-
-
-def _bill_key_from_dir(bill_dir: Path, congress_dir: Path) -> tuple[int, str, int] | None:
-    """Extract (congress, bill_type, number) from directory structure."""
-    congress = int(congress_dir.name)
-    bill_type = bill_dir.parent.name
-    name = bill_dir.name
-    # Directory name is like "hr3590" — strip the type prefix to get the number
-    number_str = name[len(bill_type) :]
-    if not number_str.isdigit():
-        return None
-    return (congress, bill_type, int(number_str))
-
-
-def _read_bill_text(version_dir: Path) -> str | None:
-    """Read bill text from a version directory, preferring .txt over .xml."""
-    for extension in ("txt", "htm", "html", "xml"):
-        candidates = list(version_dir.glob(f"document.{extension}"))
-        if not candidates:
-            candidates = list(version_dir.glob(f"*.{extension}"))
-        if candidates:
-            try:
-                return candidates[0].read_text(encoding="utf-8")
-            except Exception:
-                logger.exception("Failed to read %s", candidates[0])
-    return None
-
-
-# ---------------------------------------------------------------------------
-# Helpers
-# ---------------------------------------------------------------------------
-
-
-def _read_json(path: Path) -> dict | None:
-    """Read and parse a JSON file, returning None on failure."""
-    try:
-        return orjson.loads(path.read_bytes())
-    except FileNotFoundError:
-        return None
-    except Exception:
-        logger.exception("Failed to parse %s", path)
-        return None
-
-
-if __name__ == "__main__":
-    app()

python/data_science/ingest_posts.py

@@ -1,247 +0,0 @@
-"""Ingestion pipeline for loading JSONL post files into the weekly-partitioned posts table.
-
-Usage:
-    ingest-posts /path/to/files/
-    ingest-posts /path/to/single_file.jsonl
-    ingest-posts /data/dir/ --workers 4 --batch-size 5000
-"""
-
-from __future__ import annotations
-
-import logging
-from datetime import UTC, datetime
-from pathlib import Path  # noqa: TC003 this is needed for typer
-from typing import TYPE_CHECKING, Annotated
-
-import orjson
-import psycopg
-import typer
-
-from python.common import configure_logger
-from python.orm.common import get_connection_info
-from python.parallelize import parallelize_process
-
-if TYPE_CHECKING:
-    from collections.abc import Iterator
-
-logger = logging.getLogger(__name__)
-
-
-app = typer.Typer(help="Ingest JSONL post files into the partitioned posts table.")
-
-
-@app.command()
-def main(
-    path: Annotated[Path, typer.Argument(help="Directory containing JSONL files, or a single JSONL file")],
-    batch_size: Annotated[int, typer.Option(help="Rows per INSERT batch")] = 10000,
-    workers: Annotated[int, typer.Option(help="Parallel workers for multi-file ingestion")] = 4,
-    pattern: Annotated[str, typer.Option(help="Glob pattern for JSONL files")] = "*.jsonl",
-) -> None:
-    """Ingest JSONL post files into the weekly-partitioned posts table."""
-    configure_logger(level="INFO")
-
-    logger.info("starting ingest-posts")
-    logger.info("path=%s batch_size=%d workers=%d pattern=%s", path, batch_size, workers, pattern)
-    if path.is_file():
-        ingest_file(path, batch_size=batch_size)
-    elif path.is_dir():
-        ingest_directory(path, batch_size=batch_size, max_workers=workers, pattern=pattern)
-    else:
-        typer.echo(f"Path does not exist: {path}", err=True)
-        raise typer.Exit(code=1)
-
-    logger.info("ingest-posts done")
-
-
-def ingest_directory(
-    directory: Path,
-    *,
-    batch_size: int,
-    max_workers: int,
-    pattern: str = "*.jsonl",
-) -> None:
-    """Ingest all JSONL files in a directory using parallel workers."""
-    files = sorted(directory.glob(pattern))
-    if not files:
-        logger.warning("No JSONL files found in %s", directory)
-        return
-
-    logger.info("Found %d JSONL files to ingest", len(files))
-
-    kwargs_list = [{"path": fp, "batch_size": batch_size} for fp in files]
-    parallelize_process(ingest_file, kwargs_list, max_workers=max_workers)
-
-
-SCHEMA = "main"
-
-COLUMNS = (
-    "post_id",
-    "user_id",
-    "instance",
-    "date",
-    "text",
-    "langs",
-    "like_count",
-    "reply_count",
-    "repost_count",
-    "reply_to",
-    "replied_author",
-    "thread_root",
-    "thread_root_author",
-    "repost_from",
-    "reposted_author",
-    "quotes",
-    "quoted_author",
-    "labels",
-    "sent_label",
-    "sent_score",
-)
-
-INSERT_FROM_STAGING = f"""
-    INSERT INTO {SCHEMA}.posts ({", ".join(COLUMNS)})
-    SELECT {", ".join(COLUMNS)} FROM pg_temp.staging
-    ON CONFLICT (post_id, date) DO NOTHING
-"""  # noqa: S608
-
-FAILED_INSERT = f"""
-    INSERT INTO {SCHEMA}.failed_ingestion (raw_line, error)
-    VALUES (%(raw_line)s, %(error)s)
-"""  # noqa: S608
-
-
-def get_psycopg_connection() -> psycopg.Connection:
-    """Create a raw psycopg3 connection from environment variables."""
-    database, host, port, username, password = get_connection_info("DATA_SCIENCE_DEV")
-    return psycopg.connect(
-        dbname=database,
-        host=host,
-        port=int(port),
-        user=username,
-        password=password,
-        autocommit=False,
-    )
-
-
-def ingest_file(path: Path, *, batch_size: int) -> None:
-    """Ingest a single JSONL file into the posts table."""
-    log_trigger = max(100_000 // batch_size, 1)
-    failed_lines: list[dict] = []
-    try:
-        with get_psycopg_connection() as connection:
-            for index, batch in enumerate(read_jsonl_batches(path, batch_size, failed_lines), 1):
-                ingest_batch(connection, batch)
-                if index % log_trigger == 0:
-                    logger.info("Ingested %d batches (%d rows) from %s", index, index * batch_size, path)
-
-            if failed_lines:
-                logger.warning("Recording %d malformed lines from %s", len(failed_lines), path.name)
-                with connection.cursor() as cursor:
-                    cursor.executemany(FAILED_INSERT, failed_lines)
-                connection.commit()
-    except Exception:
-        logger.exception("Failed to ingest file: %s", path)
-        raise
-
-
-def ingest_batch(connection: psycopg.Connection, batch: list[dict]) -> None:
-    """COPY batch into a temp staging table, then INSERT ... ON CONFLICT into posts."""
-    if not batch:
-        return
-
-    try:
-        with connection.cursor() as cursor:
-            cursor.execute(f"""
-                CREATE TEMP TABLE IF NOT EXISTS staging
-                (LIKE {SCHEMA}.posts INCLUDING DEFAULTS)
-                ON COMMIT DELETE ROWS
-            """)
-            cursor.execute("TRUNCATE pg_temp.staging")
-
-            with cursor.copy(f"COPY pg_temp.staging ({', '.join(COLUMNS)}) FROM STDIN") as copy:
-                for row in batch:
-                    copy.write_row(tuple(row.get(column) for column in COLUMNS))
-
-            cursor.execute(INSERT_FROM_STAGING)
-        connection.commit()
-    except Exception as error:
-        connection.rollback()
-
-        if len(batch) == 1:
-            logger.exception("Skipping bad row post_id=%s", batch[0].get("post_id"))
-            with connection.cursor() as cursor:
-                cursor.execute(
-                    FAILED_INSERT,
-                    {
-                        "raw_line": orjson.dumps(batch[0], default=str).decode(),
-                        "error": str(error),
-                    },
-                )
-            connection.commit()
-            return
-
-        midpoint = len(batch) // 2
-        ingest_batch(connection, batch[:midpoint])
-        ingest_batch(connection, batch[midpoint:])
-
-
-def read_jsonl_batches(file_path: Path, batch_size: int, failed_lines: list[dict]) -> Iterator[list[dict]]:
-    """Stream a JSONL file and yield batches of transformed rows."""
-    batch: list[dict] = []
-    with file_path.open("r", encoding="utf-8") as handle:
-        for raw_line in handle:
-            line = raw_line.strip()
-            if not line:
-                continue
-            batch.extend(parse_line(line, file_path, failed_lines))
-            if len(batch) >= batch_size:
-                yield batch
-                batch = []
-    if batch:
-        yield batch
-
-
-def parse_line(line: str, file_path: Path, failed_lines: list[dict]) -> Iterator[dict]:
-    """Parse a JSONL line, handling concatenated JSON objects."""
-    try:
-        yield transform_row(orjson.loads(line))
-    except orjson.JSONDecodeError:
-        if "}{" not in line:
-            logger.warning("Skipping malformed line in %s: %s", file_path.name, line[:120])
-            failed_lines.append({"raw_line": line, "error": "malformed JSON"})
-            return
-        fragments = line.replace("}{", "}\n{").split("\n")
-        for fragment in fragments:
-            try:
-                yield transform_row(orjson.loads(fragment))
-            except (orjson.JSONDecodeError, KeyError, ValueError) as error:
-                logger.warning("Skipping malformed fragment in %s: %s", file_path.name, fragment[:120])
-                failed_lines.append({"raw_line": fragment, "error": str(error)})
-    except Exception as error:
-        logger.exception("Skipping bad row in %s: %s", file_path.name, line[:120])
-        failed_lines.append({"raw_line": line, "error": str(error)})
-
-
-def transform_row(raw: dict) -> dict:
-    """Transform a raw JSONL row into a dict matching the Posts table columns."""
-    raw["date"] = parse_date(raw["date"])
-    if raw.get("langs") is not None:
-        raw["langs"] = orjson.dumps(raw["langs"])
-    if raw.get("text") is not None:
-        raw["text"] = raw["text"].replace("\x00", "")
-    return raw
-
-
-def parse_date(raw_date: int) -> datetime:
-    """Parse compact YYYYMMDDHHmm integer into a naive datetime (input is UTC by spec)."""
-    return datetime(
-        raw_date // 100000000,
-        (raw_date // 1000000) % 100,
-        (raw_date // 10000) % 100,
-        (raw_date // 100) % 100,
-        raw_date % 100,
-        tzinfo=UTC,
-    )
-
-
-if __name__ == "__main__":
-    app()

python/database_cli.py

@@ -4,12 +4,10 @@ Usage:
     database <db_name> <command> [args...]
 
 Examples:
-    database van_inventory upgrade head
-    database van_inventory downgrade head-1
-    database van_inventory revision --autogenerate -m "add meals table"
-    database van_inventory check
     database richie check
     database richie upgrade head
+    database richie downgrade head-1
+    database richie revision --autogenerate -m "add meals table"
 """
 
 from __future__ import annotations
@@ -48,10 +46,7 @@ class DatabaseConfig:
 
     def alembic_config(self) -> Config:
         """Build an alembic Config for this database."""
-        # Runtime import needed — Config is in TYPE_CHECKING for the return type annotation
+        cfg = Config()
-        from alembic.config import Config as AlembicConfig  # noqa: PLC0415
-
-        cfg = AlembicConfig()
         cfg.set_main_option("script_location", self.script_location)
         cfg.set_main_option("file_template", self.file_template)
         cfg.set_main_option("prepend_sys_path", ".")
@@ -76,27 +71,6 @@ DATABASES: dict[str, DatabaseConfig] = {
         base_class_name="RichieBase",
         models_module="python.orm.richie",
     ),
-    "van_inventory": DatabaseConfig(
-        env_prefix="VAN_INVENTORY",
-        version_location="python/alembic/van_inventory/versions",
-        base_module="python.orm.van_inventory.base",
-        base_class_name="VanInventoryBase",
-        models_module="python.orm.van_inventory.models",
-    ),
-    "signal_bot": DatabaseConfig(
-        env_prefix="SIGNALBOT",
-        version_location="python/alembic/signal_bot/versions",
-        base_module="python.orm.signal_bot.base",
-        base_class_name="SignalBotBase",
-        models_module="python.orm.signal_bot.models",
-    ),
-    "data_science_dev": DatabaseConfig(
-        env_prefix="DATA_SCIENCE_DEV",
-        version_location="python/alembic/data_science_dev/versions",
-        base_module="python.orm.data_science_dev.base",
-        base_class_name="DataScienceDevBase",
-        models_module="python.orm.data_science_dev.models",
-    ),
 }
 
 

python/ebook_search/__init__.py

@@ -0,0 +1 @@
+"""EPUB search package."""

python/ebook_search/answer.py

@@ -0,0 +1,65 @@
+"""Grounded answer generation."""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+from python.ebook_search.llm_interface import request_chat_completion
+
+if TYPE_CHECKING:
+    import httpx
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.search import SearchResult
+
+logger = logging.getLogger(__name__)
+
+
+async def answer_query(
+    client: httpx.AsyncClient,
+    query: str,
+    results: list[SearchResult],
+    config: EbookSearchConfig,
+) -> str:
+    """Answer a question using only retrieved chunks."""
+    if not config.answer_enabled:
+        logger.info("ebook_answer_skipped_disabled")
+        return "Answer generation is disabled. Source chunks are shown below."
+
+    if not results:
+        logger.info("ebook_answer_skipped_no_results")
+        return "No relevant sources were found."
+
+    logger.info(
+        "ebook_answer_request_start base_url=%s model=%s sources=%s query_length=%s",
+        config.vllm_base_url,
+        config.chat_model,
+        len(results),
+        len(query),
+    )
+    context = "\n\n".join(
+        f"[{index}] {result.source_title}{' - ' + result.chapter_title if result.chapter_title else ''}\n{result.text}"
+        for index, result in enumerate(results, start=1)
+    )
+    content = await request_chat_completion(
+        client,
+        config,
+        [
+            {
+                "role": "system",
+                "content": (
+                    "Answer only from the provided context. Cite sources with bracketed numbers like [1]. "
+                    "If the context is insufficient, say so."
+                ),
+            },
+            {"role": "user", "content": f"Question:\n{query}\n\nContext:\n{context}"},
+        ],
+    )
+
+    logger.info(
+        "ebook_answer_request_complete model=%s answer_length=%s",
+        config.chat_model,
+        len(content),
+    )
+    return content or "The model returned an empty answer."

python/ebook_search/api/__init__.py

@@ -0,0 +1 @@
+"""Web and external API adapters for EPUB search."""

python/ebook_search/api/bm25_tasks.py

@@ -0,0 +1,73 @@
+"""Background BM25 refresh tasks for the web app.
+
+The refresh is scheduled on the event loop instead of a thread because the async psycopg
+driver only works from the loop; a bare thread cannot open a session on the async engine.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from typing import TYPE_CHECKING
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from python.ebook_search.bm25_corpus import load_bm25_corpus, refresh_bm25_corpus
+
+if TYPE_CHECKING:
+    from fastapi import FastAPI
+    from sqlalchemy.ext.asyncio import AsyncEngine
+
+    from python.ebook_search.config import EbookSearchConfig
+
+logger = logging.getLogger(__name__)
+
+
+def schedule_bm25_refresh(app: FastAPI) -> None:
+    """Schedule a delayed BM25 corpus refresh, replacing any pending refresh.
+
+    Only called from route handlers, so a running event loop is guaranteed.
+    """
+    cancel_bm25_refresh(app)
+
+    loop = asyncio.get_running_loop()
+
+    def start_refresh() -> None:
+        app.state.bm25_refresh_task = loop.create_task(refresh_bm25_for_app(app))
+
+    app.state.bm25_refresh_timer = loop.call_later(app.state.config.bm25_refresh_delay_seconds, start_refresh)
+    logger.info(
+        "ebook_bm25_refresh_scheduled delay_seconds=%s",
+        app.state.config.bm25_refresh_delay_seconds,
+    )
+
+
+def cancel_bm25_refresh(app: FastAPI) -> None:
+    """Cancel any pending BM25 corpus refresh timer and in-flight refresh task."""
+    existing_timer = getattr(app.state, "bm25_refresh_timer", None)
+    if existing_timer is not None:
+        existing_timer.cancel()
+        app.state.bm25_refresh_timer = None
+        logger.info("ebook_bm25_refresh_cancelled")
+
+    existing_task = getattr(app.state, "bm25_refresh_task", None)
+    if existing_task is not None:
+        if not existing_task.done():
+            existing_task.cancel()
+        app.state.bm25_refresh_task = None
+
+
+async def refresh_bm25_for_app(app: FastAPI) -> None:
+    """Refresh the BM25 corpus using the app engine and config."""
+    try:
+        await refresh_bm25_for_engine(app.state.engine, app.state.config)
+    except Exception:
+        logger.exception("ebook_bm25_refresh_failed")
+
+
+async def refresh_bm25_for_engine(engine: AsyncEngine, config: EbookSearchConfig) -> None:
+    """Refresh the BM25 corpus using an async SQLAlchemy engine."""
+    async with AsyncSession(engine) as session:
+        await refresh_bm25_corpus(session, config)
+    load_bm25_corpus.cache_clear()
+    logger.info("ebook_bm25_corpus_cache_cleared_after_refresh")

python/ebook_search/api/dependencies.py

@@ -0,0 +1,31 @@
+"""FastAPI dependencies for the EPUB search app."""
+
+from __future__ import annotations
+
+from typing import Annotated
+
+import httpx
+from fastapi import Depends, Request
+from sqlalchemy.ext.asyncio import AsyncEngine
+
+from python.ebook_search.config import EbookSearchConfig
+
+
+def get_config(request: Request) -> EbookSearchConfig:
+    """Get the loaded search config from app state."""
+    return request.app.state.config
+
+
+def get_engine(request: Request) -> AsyncEngine:
+    """Get the database engine from app state."""
+    return request.app.state.engine
+
+
+def get_http_client(request: Request) -> httpx.AsyncClient:
+    """Get the shared LLM HTTP client from app state."""
+    return request.app.state.http_client
+
+
+AppConfig = Annotated[EbookSearchConfig, Depends(get_config)]
+AppEngine = Annotated[AsyncEngine, Depends(get_engine)]
+AppHttpClient = Annotated[httpx.AsyncClient, Depends(get_http_client)]

python/ebook_search/api/judge_tasks.py

@@ -0,0 +1,131 @@
+"""Background phrase-judging tasks for the web app.
+
+Judging a book sends one LLM request per candidate phrase, which can take minutes, so it must
+not run inside the request where it would block the UI. Judgments run as async FastAPI
+background tasks, awaited on the event loop after the response is sent, and are tracked per
+book in app state so a second judge request for a book that is already being judged is
+rejected instead of doubling the work.
+
+State is loop-confined: every read and mutation happens on the event loop (async route
+handlers and async background tasks) and no critical section contains an ``await``, so each
+mutation is atomic per loop iteration and no locking is needed.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+from python.ebook_search.protected_phrases.judge_ngrams import judge_candidate_phrases_for_books
+
+if TYPE_CHECKING:
+    from fastapi import BackgroundTasks, FastAPI
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class JudgeTaskState:
+    """Running book judgments and last outcome messages, keyed by book id."""
+
+    running_book_ids: set[int] = field(default_factory=set)
+    outcome_messages: dict[int, str] = field(default_factory=dict)
+
+
+def get_judge_task_state(app: FastAPI) -> JudgeTaskState:
+    """Return the app's judge task state, creating it on first use.
+
+    Args:
+        app (FastAPI): App whose state holds the judge task registry.
+
+    Returns:
+        JudgeTaskState: The shared judge task state for this app.
+    """
+    state = getattr(app.state, "judge_tasks", None)
+    if state is None:
+        state = JudgeTaskState()
+        app.state.judge_tasks = state
+    return state
+
+
+def start_book_phrase_judgment(app: FastAPI, background_tasks: BackgroundTasks, source_id: int) -> bool:
+    """Queue judging of one book's candidate phrases as a FastAPI background task.
+
+    The book is claimed before the response returns, so a repeated judge request cannot queue
+    a second run while one is pending or running.
+
+    Args:
+        app (FastAPI): App supplying the engine, config, and judge task state.
+        background_tasks (BackgroundTasks): Request's background tasks to queue the judgment on.
+        source_id (int): Book to judge candidates for.
+
+    Returns:
+        bool: True when a judgment was queued, False when one is already running for this book.
+    """
+    state = get_judge_task_state(app)
+    if source_id in state.running_book_ids:
+        logger.info("ebook_book_phrase_judgment_already_running source_id=%s", source_id)
+        return False
+    state.running_book_ids.add(source_id)
+    state.outcome_messages.pop(source_id, None)
+    background_tasks.add_task(judge_book_phrases_for_app, app, source_id)
+    logger.info("ebook_book_phrase_judgment_queued source_id=%s", source_id)
+    return True
+
+
+async def judge_book_phrases_for_app(app: FastAPI, source_id: int) -> None:
+    """Judge one book using the app engine and config, recording the outcome message.
+
+    Args:
+        app (FastAPI): App supplying the engine, config, and judge task state.
+        source_id (int): Book to judge candidates for.
+    """
+    state = get_judge_task_state(app)
+    try:
+        result = await judge_candidate_phrases_for_books(app.state.engine, app.state.config, source_ids=[source_id])
+        logger.info(
+            "ebook_book_phrase_judgment_complete source_id=%s judged=%s protected=%s mentions=%s failed=%s",
+            source_id,
+            result.candidates_judged,
+            result.protected_phrases,
+            result.phrase_mentions,
+            result.books_failed,
+        )
+        if result.books_failed:
+            message = "Judging failed; see server logs for details"
+        else:
+            message = (
+                f"Judged {result.candidates_judged} candidates; {result.protected_phrases} protected phrases promoted"
+            )
+    except Exception:
+        logger.exception("ebook_book_phrase_judgment_task_failed source_id=%s", source_id)
+        message = "Judging failed; see server logs for details"
+    state.running_book_ids.discard(source_id)
+    state.outcome_messages[source_id] = message
+
+
+def is_judging_book(app: FastAPI, source_id: int) -> bool:
+    """Report whether a judgment is currently queued or running for one book.
+
+    Args:
+        app (FastAPI): App supplying the judge task state.
+        source_id (int): Book to check.
+
+    Returns:
+        bool: True while the book's judgment is pending or running.
+    """
+    return source_id in get_judge_task_state(app).running_book_ids
+
+
+def pop_book_judgment_outcome(app: FastAPI, source_id: int) -> str | None:
+    """Return and clear the outcome message from one book's last finished judgment.
+
+    Args:
+        app (FastAPI): App supplying the judge task state.
+        source_id (int): Book to fetch the outcome for.
+
+    Returns:
+        str | None: The outcome message, or None when there is nothing new to report.
+    """
+    return get_judge_task_state(app).outcome_messages.pop(source_id, None)

python/ebook_search/api/main.py

@@ -0,0 +1,98 @@
+"""FastAPI HTMX app for EPUB search."""
+
+from __future__ import annotations
+
+import logging
+from contextlib import asynccontextmanager
+from typing import TYPE_CHECKING, Annotated
+
+import httpx
+import typer
+import uvicorn
+from fastapi import FastAPI
+from fastapi.staticfiles import StaticFiles
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from python.common import configure_logger
+from python.ebook_search.api.bm25_tasks import cancel_bm25_refresh
+from python.ebook_search.api.routes import admin_router, health_router, page_router, search_router
+from python.ebook_search.api.web import STATIC_DIR
+from python.ebook_search.bm25_corpus import ensure_bm25_corpus
+from python.ebook_search.config import load_config
+from python.ebook_search.protected_phrases.pool import shutdown_extraction_pool
+from python.fastapi_tools import ZstdMiddleware
+from python.orm.common import get_async_postgres_engine
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator
+
+
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI) -> AsyncIterator[None]:
+    """Manage application startup and shutdown resources."""
+    logger.info("ebook_search_startup")
+    config = load_config()
+    app.state.config = config
+    logger.info(
+        "ebook_search_config_loaded top_k=%s embedding_model=%s embedding_base_url=%s vllm_base_url=%s "
+        "rerank_enabled=%s phrase_matching_enabled=%s answer_enabled=%s library_paths=%s",
+        config.top_k,
+        config.embedding_model,
+        config.embedding_base_url,
+        config.vllm_base_url,
+        config.rerank.enabled,
+        config.phrase_matching_enabled,
+        config.answer_enabled,
+        len(config.library_paths),
+    )
+    if not config.library_paths:
+        logger.warning("ebook_search_no_library_paths_configured")
+    # Concurrent phrase judging opens one session per book worker on this engine, so size the pool
+    # to cover those plus headroom for ordinary web requests.
+    app.state.engine = get_async_postgres_engine(
+        name="RICHIE",
+        vector_engine=True,
+        pool_size=config.phrase_judge_book_workers + 10,
+    )
+    app.state.http_client = httpx.AsyncClient()
+    async with AsyncSession(app.state.engine, expire_on_commit=False) as session:
+        await ensure_bm25_corpus(session, config)
+    try:
+        yield
+    finally:
+        logger.info("ebook_search_shutdown")
+        cancel_bm25_refresh(app)
+        shutdown_extraction_pool()
+        await app.state.http_client.aclose()
+        await app.state.engine.dispose()
+
+
+def create_app() -> FastAPI:
+    """Create the EPUB search web app."""
+    app = FastAPI(title="EPUB Search", lifespan=lifespan)
+    app.add_middleware(ZstdMiddleware)
+    app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")
+
+    app.include_router(admin_router)
+    app.include_router(health_router)
+    app.include_router(page_router)
+    app.include_router(search_router)
+
+    return app
+
+
+def serve(
+    host: Annotated[str, typer.Option("--host", "-h", help="Host to bind to")] = "127.0.0.1",
+    port: Annotated[int, typer.Option("--port", "-p", help="Port to bind to")] = 8070,
+    log_level: Annotated[str, typer.Option("--log-level", "-l", help="Log level")] = "INFO",
+) -> None:
+    """Start the EPUB search server."""
+    configure_logger(log_level)
+    uvicorn.run(create_app(), host=host, port=port)
+
+
+if __name__ == "__main__":
+    typer.run(serve)

python/ebook_search/api/routes/__init__.py

@@ -0,0 +1,13 @@
+"""EPUB search web route modules."""
+
+from python.ebook_search.api.routes.admin import router as admin_router
+from python.ebook_search.api.routes.health import router as health_router
+from python.ebook_search.api.routes.page import router as page_router
+from python.ebook_search.api.routes.search import router as search_router
+
+__all__ = [
+    "admin_router",
+    "health_router",
+    "page_router",
+    "search_router",
+]

python/ebook_search/api/routes/admin.py

@@ -0,0 +1,263 @@
+"""Admin routes for the EPUB search web UI."""
+
+from __future__ import annotations
+
+import logging
+
+from fastapi import APIRouter, Request
+from fastapi.responses import HTMLResponse
+
+from python.ebook_search.api.bm25_tasks import schedule_bm25_refresh
+from python.ebook_search.api.dependencies import (  # noqa: TC001 FastAPI resolves these annotated dependencies at runtime
+    AppConfig,
+    AppEngine,
+    AppHttpClient,
+)
+from python.ebook_search.api.web import templates
+from python.ebook_search.embeddings import embed_missing_chunks, embedding_model_stats
+from python.ebook_search.ingest import ingest_configured_paths
+from python.ebook_search.protected_phrases.generate_ngrams import generate_candidate_phrases_for_books
+from python.ebook_search.protected_phrases.judge_ngrams import judge_candidate_phrases_for_books
+from python.ebook_search.protected_phrases.store import book_ids_pending_first_judgment, corpus_phrase_stats
+from python.fastapi_tools import AsyncDbSession  # noqa: TC001 FastAPI resolves this annotated dependency at runtime
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/admin")
+
+
+@router.get("", response_class=HTMLResponse)
+async def admin(request: Request, config: AppConfig, session: AsyncDbSession) -> HTMLResponse:
+    """Render the admin page."""
+    stats = await embedding_model_stats(session)
+    phrase_stats = await corpus_phrase_stats(session)
+    logger.info(
+        "ebook_admin_page_loaded models=%s candidate_phrases=%s protected_phrases=%s",
+        len(stats),
+        phrase_stats.candidate_phrases,
+        phrase_stats.protected_phrases,
+    )
+    return templates.TemplateResponse(
+        request,
+        "admin.html",
+        {"config": config, "stats": stats, "phrase_stats": phrase_stats},
+    )
+
+
+@router.post("/scan", response_class=HTMLResponse)
+async def scan_library(request: Request, config: AppConfig, session: AsyncDbSession) -> HTMLResponse:
+    """Scan configured library paths for EPUB changes."""
+    try:
+        count = await ingest_configured_paths(session, config)
+        await session.commit()
+    except Exception as error:
+        logger.exception("ebook_admin_scan_failed")
+        return templates.TemplateResponse(request, "partials/error.html", {"message": str(error)}, status_code=500)
+
+    logger.info("ebook_admin_scan_complete changed_files=%s", count)
+    if count > 0:
+        schedule_bm25_refresh(request.app)
+    return templates.TemplateResponse(request, "partials/admin_status.html", {"message": f"Indexed {count} EPUBs"})
+
+
+@router.post("/phrases/generate-all", response_class=HTMLResponse)
+async def generate_all_phrases(request: Request, config: AppConfig, session: AsyncDbSession) -> HTMLResponse:
+    """Regenerate candidate phrases for every indexed book without LLM judging."""
+    return await run_phrase_generation(request, config, session, only_missing=False)
+
+
+@router.post("/phrases/generate-missing", response_class=HTMLResponse)
+async def generate_missing_phrases(request: Request, config: AppConfig, session: AsyncDbSession) -> HTMLResponse:
+    """Generate candidate phrases only for books that have none yet."""
+    return await run_phrase_generation(request, config, session, only_missing=True)
+
+
+async def run_phrase_generation(
+    request: Request,
+    config: AppConfig,
+    session: AsyncDbSession,
+    *,
+    only_missing: bool,
+) -> HTMLResponse:
+    """Run candidate phrase generation and render the outcome as an admin status partial.
+
+    Args:
+        request (Request): Current request, for template rendering.
+        config (AppConfig): Runtime phrase-tuning settings.
+        session (AsyncDbSession): Active database session.
+        only_missing (bool): Only generate for books without candidates instead of every book.
+
+    Returns:
+        HTMLResponse: Status partial describing the generation outcome.
+    """
+    try:
+        result = await generate_candidate_phrases_for_books(session, config, only_missing=only_missing)
+        await session.commit()
+    except Exception as error:
+        await session.rollback()
+        logger.exception("ebook_admin_generate_phrases_failed only_missing=%s", only_missing)
+        return templates.TemplateResponse(request, "partials/error.html", {"message": str(error)}, status_code=500)
+
+    logger.info(
+        "ebook_admin_generate_phrases_complete only_missing=%s books_seen=%s books_built=%s candidates=%s",
+        only_missing,
+        result.books_seen,
+        result.books_built,
+        result.candidate_phrases,
+    )
+    if only_missing and result.books_seen == 0:
+        return templates.TemplateResponse(
+            request,
+            "partials/admin_status.html",
+            {"message": "All books already have candidate phrases"},
+        )
+    return templates.TemplateResponse(
+        request,
+        "partials/admin_status.html",
+        {
+            "message": (
+                f"Generated phrases for {result.books_built} of {result.books_seen} books; "
+                f"{result.candidate_phrases} candidates stored"
+            )
+        },
+    )
+
+
+@router.post("/phrases/judge-all", response_class=HTMLResponse)
+async def judge_all_phrases(request: Request, engine: AppEngine, config: AppConfig) -> HTMLResponse:
+    """Judge unjudged candidate phrases across every indexed book."""
+    return await run_phrase_judgment(request, engine, config, source_ids=None)
+
+
+@router.post("/phrases/judge-missing", response_class=HTMLResponse)
+async def judge_missing_phrases(
+    request: Request,
+    engine: AppEngine,
+    config: AppConfig,
+    session: AsyncDbSession,
+) -> HTMLResponse:
+    """Judge candidate phrases only for books where judging has never run."""
+    source_ids = await book_ids_pending_first_judgment(session)
+    if not source_ids:
+        return templates.TemplateResponse(
+            request,
+            "partials/admin_status.html",
+            {"message": "All books with candidate phrases have been judged"},
+        )
+    return await run_phrase_judgment(request, engine, config, source_ids=source_ids)
+
+
+async def run_phrase_judgment(
+    request: Request,
+    engine: AppEngine,
+    config: AppConfig,
+    *,
+    source_ids: list[int] | None,
+) -> HTMLResponse:
+    """Run LLM judging for candidate phrases and render the outcome as an admin status partial.
+
+    Args:
+        request (Request): Current request, for template rendering.
+        engine (AppEngine): Engine used to open per-book judging sessions.
+        config (AppConfig): Runtime phrase-tuning settings.
+        source_ids (list[int] | None): Books to judge; ``None`` judges every indexed book.
+
+    Returns:
+        HTMLResponse: Status partial describing the judging outcome.
+    """
+    try:
+        result = await judge_candidate_phrases_for_books(engine, config, source_ids=source_ids)
+    except Exception as error:
+        logger.exception("ebook_admin_judge_phrases_failed")
+        return templates.TemplateResponse(request, "partials/error.html", {"message": str(error)}, status_code=500)
+
+    logger.info(
+        "ebook_admin_judge_phrases_complete books_seen=%s books_judged=%s books_failed=%s candidates_judged=%s "
+        "protected=%s mentions=%s",
+        result.books_seen,
+        result.books_judged,
+        result.books_failed,
+        result.candidates_judged,
+        result.protected_phrases,
+        result.phrase_mentions,
+    )
+    return templates.TemplateResponse(
+        request,
+        "partials/admin_status.html",
+        {
+            "message": (
+                f"Judged {result.candidates_judged} candidates across {result.books_judged} of "
+                f"{result.books_seen} books; {result.protected_phrases} protected phrases, "
+                f"{result.phrase_mentions} mentions"
+                + (f"; {result.books_failed} books failed" if result.books_failed else "")
+            )
+        },
+    )
+
+
+@router.post("/embed-missing", response_class=HTMLResponse)
+async def embed_missing(
+    request: Request,
+    config: AppConfig,
+    session: AsyncDbSession,
+    client: AppHttpClient,
+) -> HTMLResponse:
+    """Embed chunks missing vectors for the configured model."""
+    try:
+        count = await embed_missing_chunks(session, client, config)
+        await session.commit()
+    except Exception as error:
+        logger.exception("ebook_admin_embed_missing_failed")
+        return templates.TemplateResponse(request, "partials/error.html", {"message": str(error)}, status_code=500)
+
+    logger.info("ebook_admin_embed_missing_complete chunks=%s", count)
+    return templates.TemplateResponse(
+        request,
+        "partials/admin_status.html",
+        {"message": f"Embedded {count} chunks"},
+    )
+
+
+@router.post("/embed-all", response_class=HTMLResponse)
+async def embed_all(
+    request: Request,
+    config: AppConfig,
+    session: AsyncDbSession,
+    client: AppHttpClient,
+) -> HTMLResponse:
+    """Embed all chunks missing vectors in fixed-size batches."""
+    total = 0
+    batches = 0
+    try:
+        while True:
+            count = await embed_missing_chunks(session, client, config)
+            if count == 0:
+                break
+            await session.commit()
+            total += count
+            batches += 1
+            logger.info(
+                "ebook_admin_embed_all_batch_complete batch=%s chunks=%s total_chunks=%s",
+                batches,
+                count,
+                total,
+            )
+    except Exception as error:
+        logger.exception(
+            "ebook_admin_embed_all_failed batches=%s chunks=%s",
+            batches,
+            total,
+        )
+        return templates.TemplateResponse(
+            request,
+            "partials/error.html",
+            {"message": f"Embed all failed after {total} chunks in {batches} batches: {error}"},
+            status_code=500,
+        )
+
+    logger.info("ebook_admin_embed_all_complete batches=%s chunks=%s", batches, total)
+    return templates.TemplateResponse(
+        request,
+        "partials/admin_status.html",
+        {"message": f"Embedded {total} chunks in {batches} batches of {config.embedding_batch_size}"},
+    )

python/ebook_search/api/routes/health.py

@@ -0,0 +1,99 @@
+"""Liveness and readiness routes for the EPUB search service."""
+
+from __future__ import annotations
+
+import logging
+from http import HTTPStatus
+from typing import TYPE_CHECKING
+
+from fastapi import APIRouter
+from fastapi.responses import JSONResponse
+from sqlalchemy import literal, select
+from sqlalchemy.exc import SQLAlchemyError
+
+from python.ebook_search.api.dependencies import (  # noqa: TC001 FastAPI resolves these annotated dependencies at runtime
+    AppConfig,
+    AppHttpClient,
+)
+from python.ebook_search.bm25_corpus import bm25_index_exists, bm25_index_path, read_bm25_manifest
+from python.ebook_search.llm_interface import check_chat_endpoint, check_embedding_endpoint
+from python.fastapi_tools import AsyncDbSession  # noqa: TC001 FastAPI resolves this annotated dependency at runtime
+
+if TYPE_CHECKING:
+    import httpx
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter()
+
+
+@router.get("/health")
+async def health() -> dict[str, str]:
+    """Liveness probe that returns ok without touching dependencies."""
+    return {"status": "ok"}
+
+
+@router.get("/ready")
+async def ready(config: AppConfig, session: AsyncDbSession, client: AppHttpClient) -> JSONResponse:
+    """Readiness probe reporting database, embedding endpoint, and BM25 index status."""
+    database_ok = await check_database(session)
+    embedding_ok = await check_embedding_endpoint(client, config)
+    chat_status = await chat_endpoint_status(client, config)
+    bm25_status = check_bm25_status(config)
+
+    checks = {
+        "database": "ok" if database_ok else "fail",
+        "embedding": "ok" if embedding_ok else "fail",
+        "chat": chat_status,
+        "bm25": bm25_status,
+    }
+    if not database_ok:
+        status = "unavailable"
+        status_code = HTTPStatus.SERVICE_UNAVAILABLE
+    elif not embedding_ok or chat_status == "fail" or bm25_status == "missing":
+        status = "degraded"
+        status_code = HTTPStatus.OK
+    else:
+        status = "ready"
+        status_code = HTTPStatus.OK
+
+    logger.info(
+        "ebook_ready_check status=%s database=%s embedding=%s chat=%s bm25=%s",
+        status,
+        database_ok,
+        embedding_ok,
+        chat_status,
+        bm25_status,
+    )
+    return JSONResponse(content={"status": status, "checks": checks}, status_code=status_code)
+
+
+async def chat_endpoint_status(client: httpx.AsyncClient, config: EbookSearchConfig) -> str:
+    """Return the answering chat endpoint status, or disabled when answers are off."""
+    if not config.answer_enabled:
+        return "disabled"
+    return "ok" if await check_chat_endpoint(client, config) else "fail"
+
+
+async def check_database(session: AsyncSession) -> bool:
+    """Return whether the database answers a trivial query."""
+    try:
+        await session.execute(select(literal(1)))
+    except SQLAlchemyError as error:
+        logger.warning("ebook_ready_database_unavailable error=%s", error)
+        return False
+    return True
+
+
+def check_bm25_status(config: EbookSearchConfig) -> str:
+    """Return the persisted BM25 index status without loading it into memory."""
+    index_path = bm25_index_path(config)
+    manifest = read_bm25_manifest(index_path)
+    if manifest is None or not bm25_index_exists(index_path, manifest):
+        return "missing"
+    if manifest.chunk_count == 0:
+        return "empty"
+    return "ok"

python/ebook_search/api/routes/page.py

@@ -0,0 +1,202 @@
+"""Page routes for the EPUB search web UI."""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+from fastapi import APIRouter, BackgroundTasks, HTTPException, Request
+from fastapi.responses import HTMLResponse, RedirectResponse
+from sqlalchemy import func, select
+
+from python.ebook_search.api.dependencies import (
+    AppConfig,  # noqa: TC001 FastAPI resolves this annotated dependency at runtime
+)
+from python.ebook_search.api.judge_tasks import is_judging_book, pop_book_judgment_outcome, start_book_phrase_judgment
+from python.ebook_search.api.web import templates
+from python.ebook_search.protected_phrases.generate_ngrams import recalculate_candidate_phrases_for_book
+from python.fastapi_tools import AsyncDbSession  # noqa: TC001 FastAPI resolves this annotated dependency at runtime
+from python.orm.richie import EbookCandidatePhrase, EbookChapter, EbookChunk, EbookProtectedPhrase, EbookSource
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter()
+
+
+@router.get("/", response_class=HTMLResponse)
+async def index(request: Request, config: AppConfig) -> HTMLResponse:
+    """Render the search page."""
+    return templates.TemplateResponse(request, "search.html", {"config": config})
+
+
+@router.get("/books", response_class=HTMLResponse)
+async def books(request: Request, session: AsyncDbSession) -> HTMLResponse:
+    """Render the indexed books page."""
+    sources = list((await session.scalars(select(EbookSource).order_by(EbookSource.title))).all())
+    logger.info("ebook_books_page_loaded count=%s", len(sources))
+    return templates.TemplateResponse(request, "books.html", {"sources": sources})
+
+
+async def get_chapter_count(session: AsyncSession, book_id: int) -> int:
+    """Return the number of indexed chapters for one book."""
+    return await session.scalar(select(func.count(EbookChapter.id)).where(EbookChapter.source_id == book_id)) or 0
+
+
+async def get_chunk_count(session: AsyncSession, book_id: int) -> int:
+    """Return the number of indexed chunks for one book."""
+    return await session.scalar(select(func.count(EbookChunk.id)).where(EbookChunk.source_id == book_id)) or 0
+
+
+async def get_candidate_count(session: AsyncSession, book_id: int) -> int:
+    """Return the number of indexed candidates for one book."""
+    return (
+        await session.scalar(select(func.count(EbookCandidatePhrase.id)).where(EbookCandidatePhrase.book_id == book_id))
+        or 0
+    )
+
+
+async def get_judged_candidate_count(session: AsyncSession, book_id: int) -> int:
+    """Return the number of judged candidates for one book."""
+    return (
+        await session.scalar(
+            select(func.count(EbookCandidatePhrase.id)).where(
+                EbookCandidatePhrase.book_id == book_id,
+                EbookCandidatePhrase.llm_judged.is_(True),
+            )
+        )
+        or 0
+    )
+
+
+async def get_protected_count(session: AsyncSession, book_id: int) -> int:
+    """Return the number of protected phrases for one book."""
+    return (
+        await session.scalar(select(func.count(EbookProtectedPhrase.id)).where(EbookProtectedPhrase.book_id == book_id))
+        or 0
+    )
+
+
+async def get_candidates(session: AsyncSession, book_id: int) -> list[EbookCandidatePhrase]:
+    """Return the indexed candidates for one book."""
+    return list(
+        await session.scalars(
+            select(EbookCandidatePhrase)
+            .where(EbookCandidatePhrase.book_id == book_id)
+            .order_by(EbookCandidatePhrase.candidate_score.desc())
+            .limit(100)
+        )
+    )
+
+
+async def get_protected_phrases(session: AsyncSession, book_id: int) -> list[EbookProtectedPhrase]:
+    """Return the protected phrases for one book."""
+    return list(
+        await session.scalars(
+            select(EbookProtectedPhrase)
+            .where(EbookProtectedPhrase.book_id == book_id)
+            .order_by(EbookProtectedPhrase.importance.desc())
+            .limit(100)
+        )
+    )
+
+
+@router.get("/books/{source_id}", response_class=HTMLResponse)
+async def book_detail(source_id: int, request: Request, session: AsyncDbSession) -> HTMLResponse:
+    """Render details for one indexed book."""
+    source = await session.get(EbookSource, source_id)
+    phrase_status_message = None
+    recalculated = request.query_params.get("phrases_recalculated")
+    if recalculated is not None:
+        phrase_status_message = f"Recalculated phrases; {recalculated} candidates generated"
+    judgment_outcome = pop_book_judgment_outcome(request.app, source_id)
+    if judgment_outcome is not None:
+        phrase_status_message = judgment_outcome
+    judging_in_progress = is_judging_book(request.app, source_id)
+    if judging_in_progress:
+        phrase_status_message = "Judging candidate phrases in the background; refresh to see progress"
+    if source is not None:
+        chapter_count = await get_chapter_count(session, source.id)
+        chunk_count = await get_chunk_count(session, source.id)
+        candidate_count = await get_candidate_count(session, source.id)
+        judged_candidate_count = await get_judged_candidate_count(session, source.id)
+        protected_count = await get_protected_count(session, source.id)
+        candidates = await get_candidates(session, source.id)
+        protected_phrases = await get_protected_phrases(session, source.id)
+    else:
+        chapter_count = 0
+        chunk_count = 0
+        candidate_count = 0
+        judged_candidate_count = 0
+        protected_count = 0
+        candidates = []
+        protected_phrases = []
+    logger.info(
+        "ebook_book_detail_loaded source_id=%s found=%s chapters=%s chunks=%s candidates=%s judged=%s protected=%s",
+        source_id,
+        source is not None,
+        chapter_count,
+        chunk_count,
+        candidate_count,
+        judged_candidate_count,
+        protected_count,
+    )
+    return templates.TemplateResponse(
+        request,
+        "book_detail.html",
+        {
+            "candidate_count": candidate_count,
+            "candidates": candidates,
+            "chapter_count": chapter_count,
+            "chunk_count": chunk_count,
+            "judged_candidate_count": judged_candidate_count,
+            "judging_in_progress": judging_in_progress,
+            "protected_count": protected_count,
+            "protected_phrases": protected_phrases,
+            "phrase_status_message": phrase_status_message,
+            "source": source,
+        },
+    )
+
+
+@router.post("/books/{source_id}/recalculate-phrases")
+async def recalculate_book_phrases(source_id: int, config: AppConfig, session: AsyncDbSession) -> RedirectResponse:
+    """Clear and regenerate candidate phrases for one indexed book."""
+    source = await session.get(EbookSource, source_id)
+    if source is None:
+        raise HTTPException(status_code=404, detail="Book not found")
+
+    result = await recalculate_candidate_phrases_for_book(session, source, config, use_process_pool=True)
+    logger.info(
+        "ebook_book_phrase_recalculation_complete source_id=%s candidates=%s deleted_candidates=%s "
+        "deleted_protected=%s deleted_aliases=%s deleted_mentions=%s",
+        source_id,
+        result.candidate_phrases,
+        result.deleted_candidates,
+        result.deleted_protected_phrases,
+        result.deleted_aliases,
+        result.deleted_mentions,
+    )
+    return RedirectResponse(
+        url=f"/books/{source_id}?phrases_recalculated={result.candidate_phrases}",
+        status_code=303,
+    )
+
+
+@router.post("/books/{source_id}/judge-phrases")
+async def judge_book_phrases(
+    source_id: int,
+    request: Request,
+    background_tasks: BackgroundTasks,
+    session: AsyncDbSession,
+) -> RedirectResponse:
+    """Queue background judging of one book's candidate phrases and return immediately."""
+    source = await session.get(EbookSource, source_id)
+    if source is None:
+        raise HTTPException(status_code=404, detail="Book not found")
+
+    started = start_book_phrase_judgment(request.app, background_tasks, source.id)
+    logger.info("ebook_book_phrase_judgment_requested source_id=%s started=%s", source_id, started)
+    return RedirectResponse(url=f"/books/{source_id}", status_code=303)

python/ebook_search/api/routes/search.py

@@ -0,0 +1,129 @@
+"""Search routes for the EPUB search web UI."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import replace
+from time import perf_counter
+from typing import TYPE_CHECKING, Annotated
+
+from fastapi import APIRouter, Form, Request
+from fastapi.responses import HTMLResponse
+
+from python.ebook_search.answer import answer_query
+from python.ebook_search.api.dependencies import (  # noqa: TC001 FastAPI resolves these annotated dependencies at runtime
+    AppConfig,
+    AppEngine,
+    AppHttpClient,
+)
+from python.ebook_search.api.web import templates
+from python.ebook_search.guardrails import (
+    CitationReport,
+    is_confident,
+    retrieval_confidence,
+    validate_citations,
+)
+from python.ebook_search.search import SearchResponse, search_ebooks
+from python.ebook_search.timing import runtime_step_from_start
+
+if TYPE_CHECKING:
+    import httpx
+
+    from python.ebook_search.config import EbookSearchConfig
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter()
+
+
+async def build_answer(
+    client: httpx.AsyncClient,
+    query: str,
+    response: SearchResponse,
+    config: EbookSearchConfig,
+) -> tuple[str, bool, CitationReport | None]:
+    """Generate the answer for a search, returning ``(answer, low_confidence, citation_report)``."""
+    if not config.answer_enabled:
+        logger.info("ebook_answer_skipped_disabled")
+        return "Answer generation is disabled. Source chunks are shown below.", False, None
+
+    if not is_confident(response.results, config):
+        logger.info(
+            "ebook_answer_low_confidence confidence=%.4f threshold=%.4f",
+            retrieval_confidence(response.results),
+            config.min_retrieval_confidence,
+        )
+        answer = (
+            "Retrieval confidence is low for this query, so answer generation was skipped. "
+            "Source chunks are shown below."
+        )
+        return answer, True, None
+
+    try:
+        answer = await answer_query(client, query, response.results, config)
+    except RuntimeError as error:
+        logger.warning("ebook_answer_request_failed_falling_back error=%s", error)
+        return "Answer generation failed. Source chunks are still shown below.", False, None
+
+    citation_report = None
+    if config.validate_citations_enabled and response.results:
+        citation_report = validate_citations(answer, len(response.results))
+        if citation_report.invalid or not citation_report.grounded:
+            logger.warning(
+                "ebook_answer_citation_issue invalid=%s grounded=%s",
+                citation_report.invalid,
+                citation_report.grounded,
+            )
+    return answer, False, citation_report
+
+
+@router.post("/search", response_class=HTMLResponse)
+async def search(
+    request: Request,
+    config: AppConfig,
+    engine: AppEngine,
+    client: AppHttpClient,
+    query: Annotated[str, Form()],
+    rerank: Annotated[str | None, Form()] = None,
+    phrase_matching: Annotated[str | None, Form()] = None,
+) -> HTMLResponse:
+    """Run a search and render HTMX results."""
+    try:
+        response = await search_ebooks(
+            engine,
+            client,
+            query,
+            config,
+            rerank=rerank == "true",
+            phrase_matching=phrase_matching == "true",
+        )
+    except Exception as error:
+        logger.exception("ebook_search_request_failed")
+        return templates.TemplateResponse(request, "partials/error.html", {"message": str(error)}, status_code=500)
+
+    answer_start = perf_counter()
+    answer, low_confidence, citation_report = await build_answer(client, query, response, config)
+    answer_step_name = "Answer generation" if config.answer_enabled else "Answer skipped"
+    response = replace(
+        response,
+        timings=(*response.timings, runtime_step_from_start(answer_step_name, answer_start)),
+    )
+
+    for step in response.timings:
+        logger.info("ebook_search_timing step=%r runtime_ms=%.1f", step.name, step.duration_ms)
+    logger.info(
+        "ebook_search_request_complete results=%s rank_label=%s runtime_ms=%.1f",
+        len(response.results),
+        response.rank_label,
+        response.total_runtime_ms,
+    )
+    return templates.TemplateResponse(
+        request,
+        "partials/results.html",
+        {
+            "answer": answer,
+            "response": response,
+            "low_confidence": low_confidence,
+            "citation_report": citation_report,
+        },
+    )

python/ebook_search/api/static/style.css

@@ -0,0 +1,447 @@
+:root {
+    --bg: #f4f5f7;
+    --surface: #ffffff;
+    --border: #e3e5ea;
+    --text: #1c1f24;
+    --muted: #6b7280;
+    --accent: #4f46e5;
+    --accent-soft: #eef0fe;
+    --danger: #b42318;
+    --warn-bg: #fff8eb;
+    --warn-border: #e0a92e;
+    --warn-text: #7a5008;
+    --radius: 12px;
+    --shadow: 0 1px 2px rgba(16, 24, 40, 0.04), 0 1px 3px rgba(16, 24, 40, 0.08);
+}
+
+html.theme-dark {
+    --bg: #0f1117;
+    --surface: #1a1d25;
+    --border: #2b303b;
+    --text: #e6e8ec;
+    --muted: #9aa1ad;
+    --accent: #818cf8;
+    --accent-soft: #262b45;
+    --danger: #f97066;
+    --warn-bg: #2a2410;
+    --warn-border: #b9881f;
+    --warn-text: #e8c97a;
+    --shadow: 0 1px 2px rgba(0, 0, 0, 0.3), 0 1px 3px rgba(0, 0, 0, 0.4);
+    color-scheme: dark;
+}
+
+* {
+    box-sizing: border-box;
+}
+
+body {
+    margin: 0;
+    background: var(--bg);
+    color: var(--text);
+    font-family: system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
+    line-height: 1.55;
+}
+
+main {
+    max-width: 820px;
+    margin: 0 auto;
+    padding: 32px 20px 64px;
+}
+
+/* Header / nav */
+.site-header {
+    background: var(--surface);
+    border-bottom: 1px solid var(--border);
+    position: sticky;
+    top: 0;
+    z-index: 10;
+}
+
+.site-nav {
+    max-width: 820px;
+    margin: 0 auto;
+    padding: 12px 20px;
+    display: flex;
+    align-items: center;
+    gap: 20px;
+}
+
+.brand {
+    font-weight: 700;
+    font-size: 1.05rem;
+    color: var(--text);
+    text-decoration: none;
+}
+
+.nav-links {
+    display: flex;
+    gap: 6px;
+    margin-right: auto;
+}
+
+.nav-links a {
+    padding: 6px 12px;
+    border-radius: 8px;
+    color: var(--muted);
+    text-decoration: none;
+    font-size: 0.94rem;
+    transition: background 0.15s, color 0.15s;
+}
+
+.nav-links a:hover {
+    background: var(--accent-soft);
+    color: var(--accent);
+}
+
+.dev-toggle {
+    display: inline-flex;
+    align-items: center;
+    gap: 6px;
+    font-size: 0.85rem;
+    color: var(--muted);
+    cursor: pointer;
+    user-select: none;
+}
+
+.theme-toggle {
+    display: inline-flex;
+    align-items: center;
+    justify-content: center;
+    width: 34px;
+    height: 34px;
+    padding: 0;
+    font-size: 1rem;
+    line-height: 1;
+    color: var(--text);
+    background: var(--bg);
+    border: 1px solid var(--border);
+    border-radius: 8px;
+    cursor: pointer;
+}
+
+.theme-toggle:hover {
+    border-color: var(--accent);
+    filter: none;
+}
+
+h1 {
+    font-size: 1.6rem;
+    margin: 0 0 20px;
+}
+
+h2 {
+    font-size: 1.15rem;
+    margin: 0 0 8px;
+}
+
+/* Cards */
+.card {
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    box-shadow: var(--shadow);
+    padding: 20px;
+}
+
+/* Search form */
+form {
+    margin: 0;
+}
+
+label {
+    font-weight: 600;
+    font-size: 0.92rem;
+}
+
+textarea {
+    display: block;
+    width: 100%;
+    margin: 8px 0 16px;
+    padding: 12px 14px;
+    font: inherit;
+    color: var(--text);
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: 10px;
+    resize: vertical;
+    transition: border-color 0.15s, box-shadow 0.15s;
+}
+
+textarea:focus {
+    outline: none;
+    border-color: var(--accent);
+    box-shadow: 0 0 0 3px var(--accent-soft);
+}
+
+.form-row {
+    display: flex;
+    align-items: center;
+    justify-content: space-between;
+    gap: 12px;
+    flex-wrap: wrap;
+}
+
+.search-toggles {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 14px;
+}
+
+button {
+    padding: 10px 20px;
+    font: inherit;
+    font-weight: 600;
+    color: #fff;
+    background: var(--accent);
+    border: none;
+    border-radius: 10px;
+    cursor: pointer;
+    transition: filter 0.15s;
+}
+
+button:hover {
+    filter: brightness(1.08);
+}
+
+.check {
+    display: inline-flex;
+    gap: 8px;
+    align-items: center;
+    font-weight: 500;
+    color: var(--muted);
+}
+
+.actions {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 12px;
+    margin-bottom: 24px;
+}
+
+.actions-grid {
+    display: grid;
+    grid-template-columns: repeat(2, max-content);
+}
+
+/* Answer + results */
+#results {
+    display: block;
+    margin-top: 28px;
+}
+
+.rank-label {
+    font-size: 0.82rem;
+    font-weight: 600;
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+    color: var(--muted);
+    margin-bottom: 16px;
+}
+
+.answer {
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    box-shadow: var(--shadow);
+    padding: 20px;
+    margin-bottom: 24px;
+}
+
+.answer p:last-child {
+    margin-bottom: 0;
+}
+
+.results {
+    list-style: none;
+    padding: 0;
+    margin: 0;
+    display: grid;
+    gap: 16px;
+}
+
+.results > li {
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    box-shadow: var(--shadow);
+    padding: 18px 20px;
+}
+
+.results h2 {
+    font-size: 1.05rem;
+}
+
+.results h2 a {
+    color: var(--text);
+    text-decoration: none;
+}
+
+.results h2 a:hover {
+    color: var(--accent);
+}
+
+.meta {
+    color: var(--muted);
+    font-size: 0.88rem;
+    margin: 0 0 10px;
+}
+
+.scores {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 8px;
+    margin: 14px 0 0;
+}
+
+.scores div {
+    display: inline-flex;
+    gap: 6px;
+    align-items: baseline;
+    padding: 3px 10px;
+    background: var(--bg);
+    border: 1px solid var(--border);
+    border-radius: 999px;
+    font-size: 0.78rem;
+}
+
+.scores dt {
+    font-weight: 600;
+    color: var(--muted);
+}
+
+.scores dd {
+    margin: 0;
+    font-variant-numeric: tabular-nums;
+}
+
+.phrase-matches {
+    display: flex;
+    flex-wrap: wrap;
+    gap: 8px;
+    align-items: baseline;
+    margin: 10px 0 0;
+    font-size: 0.78rem;
+}
+
+.phrase-matches-label {
+    color: var(--muted);
+    font-weight: 600;
+}
+
+.phrase-match {
+    padding: 3px 10px;
+    background: var(--bg);
+    border: 1px solid var(--border);
+    border-radius: 999px;
+    color: var(--accent);
+}
+
+/* Runtime — developer diagnostics, hidden unless dev mode is on */
+.runtime {
+    display: none;
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    box-shadow: var(--shadow);
+    padding: 18px 20px;
+    margin-bottom: 24px;
+}
+
+html.dev .runtime {
+    display: block;
+}
+
+.timing-chart {
+    display: grid;
+    gap: 8px;
+    padding: 0;
+    margin: 12px 0 0;
+    list-style: none;
+}
+
+.timing-chart li {
+    display: grid;
+    grid-template-columns: minmax(150px, 1fr) minmax(160px, 2fr) auto auto;
+    gap: 10px;
+    align-items: center;
+    font-size: 0.85rem;
+}
+
+.timing-bar {
+    height: 8px;
+    overflow: hidden;
+    background: var(--bg);
+    border-radius: 999px;
+}
+
+.timing-bar span {
+    display: block;
+    height: 100%;
+    background: var(--accent);
+    border-radius: 999px;
+}
+
+.timing-value,
+.timing-remaining {
+    color: var(--muted);
+    font-variant-numeric: tabular-nums;
+    text-align: right;
+}
+
+/* Tables */
+table {
+    width: 100%;
+    border-collapse: collapse;
+    background: var(--surface);
+    border: 1px solid var(--border);
+    border-radius: var(--radius);
+    overflow: hidden;
+}
+
+th,
+td {
+    padding: 10px 14px;
+    border-bottom: 1px solid var(--border);
+    text-align: left;
+    font-size: 0.9rem;
+}
+
+th {
+    font-weight: 600;
+    color: var(--muted);
+    background: var(--bg);
+}
+
+tbody tr:last-child td {
+    border-bottom: none;
+}
+
+dl dt {
+    font-weight: 600;
+    color: var(--muted);
+    font-size: 0.85rem;
+}
+
+dl dd {
+    margin: 0 0 12px;
+}
+
+/* States */
+.error {
+    color: var(--danger);
+    font-weight: 600;
+}
+
+.notice {
+    margin: 12px 0;
+    padding: 10px 14px;
+    border-left: 3px solid var(--warn-border);
+    border-radius: 6px;
+    background: var(--warn-bg);
+    color: var(--warn-text);
+    font-weight: 500;
+}
+
+.status {
+    color: var(--muted);
+}

python/ebook_search/api/templates/admin.html

@@ -0,0 +1,110 @@
+{% extends "base.html" %} {% block title %}EPUB Admin{% endblock %} {% block
+head %}
+<script src="https://unpkg.com/htmx.org@2.0.4"></script>
+{% endblock %} {% block content %}
+<h1>Admin</h1>
+<section id="admin-status"></section>
+<section class="actions">
+  <form hx-post="/admin/scan" hx-target="#admin-status" hx-swap="innerHTML">
+    <button type="submit">Scan</button>
+  </form>
+</section>
+<section>
+  <h2>Embeddings</h2>
+  <section class="actions">
+    <form
+      hx-post="/admin/embed-missing"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Embed</button>
+    </form>
+    <form
+      hx-post="/admin/embed-all"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Embed all</button>
+    </form>
+  </section>
+  <table>
+    <thead>
+      <tr>
+        <th>Model</th>
+        <th>Dimensions</th>
+        <th>Embedded</th>
+        <th>Missing</th>
+        <th>Total chunks</th>
+      </tr>
+    </thead>
+    <tbody>
+      {% for item in stats %}
+      <tr>
+        <td>{{ item.model_name }}</td>
+        <td>{{ item.dimension }}</td>
+        <td>{{ item.embedded_chunks }}</td>
+        <td>{{ item.missing_chunks }}</td>
+        <td>{{ item.total_chunks }}</td>
+      </tr>
+      {% endfor %}
+    </tbody>
+  </table>
+</section>
+<section>
+  <h2>Protected phrases</h2>
+  <section class="actions actions-grid">
+    <form
+      hx-post="/admin/phrases/generate-all"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Regenerate all phrases</button>
+    </form>
+    <form
+      hx-post="/admin/phrases/generate-missing"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Add missing phrases</button>
+    </form>
+    <form
+      hx-post="/admin/phrases/judge-all"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Judge all phrases</button>
+    </form>
+    <form
+      hx-post="/admin/phrases/judge-missing"
+      hx-target="#admin-status"
+      hx-swap="innerHTML"
+    >
+      <button type="submit">Judge missing phrases</button>
+    </form>
+  </section>
+  <table>
+    <thead>
+      <tr>
+        <th>Candidates</th>
+        <th>Judged</th>
+        <th>Unjudged</th>
+        <th>Protected</th>
+        <th>Books indexed</th>
+        <th>Books generated</th>
+        <th>Books fully judged</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr>
+        <td>{{ phrase_stats.candidate_phrases }}</td>
+        <td>{{ phrase_stats.judged_candidates }}</td>
+        <td>{{ phrase_stats.unjudged_candidates }}</td>
+        <td>{{ phrase_stats.protected_phrases }}</td>
+        <td>{{ phrase_stats.total_books }}</td>
+        <td>{{ phrase_stats.books_with_candidates }}</td>
+        <td>{{ phrase_stats.books_fully_judged }}</td>
+      </tr>
+    </tbody>
+  </table>
+</section>
+{% endblock %}

python/ebook_search/api/templates/base.html

@@ -0,0 +1,71 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>{% block title %}EPUB Search{% endblock %}</title>
+    {% block head %}{% endblock %}
+    <link rel="stylesheet" href="/static/style.css?v={{ static_version('style.css') }}">
+    <script>
+        // Apply theme and dev mode before paint to avoid a flash of unstyled/wrong content.
+        (function () {
+            var stored = localStorage.getItem("ebook-theme");
+            var prefersDark = window.matchMedia("(prefers-color-scheme: dark)").matches;
+            var theme = stored || (prefersDark ? "dark" : "light");
+            document.documentElement.classList.add("theme-" + theme);
+            if (localStorage.getItem("ebook-dev-mode") === "on") {
+                document.documentElement.classList.add("dev");
+            }
+        })();
+    </script>
+</head>
+<body>
+    <header class="site-header">
+        <nav class="site-nav">
+            <a class="brand" href="/">EPUB Search</a>
+            <div class="nav-links">
+                <a href="/">Search</a>
+                <a href="/books">Books</a>
+                <a href="/admin">Admin</a>
+            </div>
+            <button type="button" id="theme-toggle" class="theme-toggle" title="Toggle light / dark theme" aria-label="Toggle theme"></button>
+            <label class="dev-toggle" title="Show developer diagnostics">
+                <input type="checkbox" id="dev-mode-toggle">
+                <span>Dev</span>
+            </label>
+        </nav>
+    </header>
+    <main>
+        {% block content %}{% endblock %}
+    </main>
+    <script>
+        (function () {
+            var toggle = document.getElementById("dev-mode-toggle");
+            if (toggle) {
+                toggle.checked = document.documentElement.classList.contains("dev");
+                toggle.addEventListener("change", function () {
+                    document.documentElement.classList.toggle("dev", toggle.checked);
+                    localStorage.setItem("ebook-dev-mode", toggle.checked ? "on" : "off");
+                });
+            }
+
+            var themeButton = document.getElementById("theme-toggle");
+            if (themeButton) {
+                var root = document.documentElement;
+                var sync = function () {
+                    var isDark = root.classList.contains("theme-dark");
+                    themeButton.textContent = isDark ? "☀️" : "🌙";
+                };
+                sync();
+                themeButton.addEventListener("click", function () {
+                    var next = root.classList.contains("theme-dark") ? "light" : "dark";
+                    root.classList.remove("theme-dark", "theme-light");
+                    root.classList.add("theme-" + next);
+                    localStorage.setItem("ebook-theme", next);
+                    sync();
+                });
+            }
+        })();
+    </script>
+</body>
+</html>

python/ebook_search/api/templates/book_detail.html

@@ -0,0 +1,109 @@
+{% extends "base.html" %}
+
+{% block title %}{% if source %}{{ source.title }}{% else %}Book not found{% endif %}{% endblock %}
+
+{% block content %}
+    {% if source %}
+        <h1>{{ source.title }}</h1>
+        <p class="meta">{{ source.author or "Unknown author" }}</p>
+        {% if phrase_status_message %}
+            <p class="status">{{ phrase_status_message }}</p>
+        {% endif %}
+        <dl class="card">
+            <dt>File</dt>
+            <dd>{{ source.file_path }}</dd>
+            <dt>Chapters</dt>
+            <dd>{{ chapter_count }}</dd>
+            <dt>Chunks</dt>
+            <dd>{{ chunk_count }}</dd>
+            <dt>Candidates</dt>
+            <dd>{{ candidate_count }}</dd>
+            <dt>Judged</dt>
+            <dd>{{ judged_candidate_count }}</dd>
+            <dt>Protected</dt>
+            <dd>{{ protected_count }}</dd>
+        </dl>
+        <form
+            method="post"
+            action="/books/{{ source.id }}/recalculate-phrases"
+            onsubmit="return confirm('Remove old phrases for this book and generate new candidates?');"
+        >
+            <button type="submit">Recalculate phrases</button>
+        </form>
+        <form
+            method="post"
+            action="/books/{{ source.id }}/judge-phrases"
+            onsubmit="return confirm('Judge candidate phrases for this book with the LLM?');"
+        >
+            <button type="submit"{% if judging_in_progress %} disabled{% endif %}>
+                {% if judging_in_progress %}Judging&hellip;{% else %}Judge phrases{% endif %}
+            </button>
+        </form>
+
+        <section>
+            <h2>Candidate n-grams</h2>
+            {% if candidates %}
+                <table>
+                    <thead>
+                        <tr>
+                            <th>Phrase</th>
+                            <th>Status</th>
+                            <th>Score</th>
+                            <th>Count</th>
+                            <th>Chapters</th>
+                        </tr>
+                    </thead>
+                    <tbody>
+                        {% for candidate in candidates %}
+                            <tr>
+                                <td>{{ candidate.phrase_text }}</td>
+                                <td>
+                                    {% if candidate.llm_judged %}
+                                        {% if candidate.llm_keep %}Kept{% else %}Rejected{% endif %}
+                                    {% else %}
+                                        Candidate
+                                    {% endif %}
+                                </td>
+                                <td>{{ "%.2f"|format(candidate.candidate_score) }}</td>
+                                <td>{{ candidate.raw_count }}</td>
+                                <td>{{ candidate.chapter_count }}</td>
+                            </tr>
+                        {% endfor %}
+                    </tbody>
+                </table>
+            {% else %}
+                <p>No candidate n-grams.</p>
+            {% endif %}
+        </section>
+
+        <section>
+            <h2>Protected phrases</h2>
+            {% if protected_phrases %}
+                <table>
+                    <thead>
+                        <tr>
+                            <th>Phrase</th>
+                            <th>Type</th>
+                            <th>Confidence</th>
+                            <th>Importance</th>
+                        </tr>
+                    </thead>
+                    <tbody>
+                        {% for phrase in protected_phrases %}
+                            <tr>
+                                <td>{{ phrase.phrase_text }}</td>
+                                <td>{{ phrase.phrase_type or "phrase" }}</td>
+                                <td>{{ "%.2f"|format(phrase.confidence) }}</td>
+                                <td>{{ "%.2f"|format(phrase.importance) }}</td>
+                            </tr>
+                        {% endfor %}
+                    </tbody>
+                </table>
+            {% else %}
+                <p>No protected phrases.</p>
+            {% endif %}
+        </section>
+    {% else %}
+        <h1>Book not found</h1>
+    {% endif %}
+{% endblock %}

python/ebook_search/api/templates/books.html

@@ -0,0 +1,19 @@
+{% extends "base.html" %}
+
+{% block title %}EPUB Books{% endblock %}
+
+{% block content %}
+    <h1>Books</h1>
+    {% if sources %}
+        <ol class="results">
+            {% for source in sources %}
+                <li>
+                    <h2><a href="/books/{{ source.id }}">{{ source.title }}</a></h2>
+                    <p class="meta">{{ source.author or "Unknown author" }}</p>
+                </li>
+            {% endfor %}
+        </ol>
+    {% else %}
+        <p>No EPUBs indexed.</p>
+    {% endif %}
+{% endblock %}

python/ebook_search/api/templates/partials/admin_status.html

@@ -0,0 +1 @@
+<p class="status">{{ message }}</p>

python/ebook_search/api/templates/partials/error.html

@@ -0,0 +1 @@
+<p class="error">{{ message }}</p>

python/ebook_search/api/templates/partials/results.html

@@ -0,0 +1,98 @@
+<div class="rank-label">{{ response.rank_label }}</div>
+{% if response.timings %}
+    <section class="runtime">
+        <h2>Runtime</h2>
+        <p class="meta">Total {{ "%.1f"|format(response.total_runtime_ms) }} ms</p>
+        <ol class="timing-chart">
+            {% set total = response.total_runtime_ms %}
+            {% set ns = namespace(remaining=total) %}
+            {% for step in response.timings %}
+                {% set width = (step.duration_ms / total * 100) if total else 0 %}
+                {% if step.counts_toward_total %}
+                    {% set ns.remaining = ns.remaining - step.duration_ms %}
+                {% endif %}
+                <li>
+                    <span class="timing-label">{{ step.name }}</span>
+                    <span class="timing-bar"><span style="width: {{ "%.2f"|format(width) }}%"></span></span>
+                    <span class="timing-value">{{ "%.1f"|format(step.duration_ms) }} ms</span>
+                    <span class="timing-remaining">{{ "%.1f"|format([ns.remaining, 0]|max) }} ms left</span>
+                </li>
+            {% endfor %}
+        </ol>
+    </section>
+{% endif %}
+<section class="answer">
+    <h2>Answer</h2>
+    {% if low_confidence|default(false) %}
+        <p class="notice">Low retrieval confidence — answer generation was skipped.</p>
+    {% endif %}
+    {% set report = citation_report|default(none) %}
+    {% if report is not none and not report.grounded %}
+        <p class="notice">Unverified — no source citations were found in this answer.</p>
+    {% endif %}
+    {% if report is not none and report.invalid %}
+        <p class="notice">Invalid citations: {{ report.invalid|join(", ") }} (no matching source).</p>
+    {% endif %}
+    <p>{{ answer }}</p>
+</section>
+{% if response.results %}
+    <ol class="results">
+        {% for result in response.results %}
+            <li>
+                <h2>
+                    {% if result.source_id %}
+                        <a href="/books/{{ result.source_id }}">{{ result.source_title }}</a>
+                    {% else %}
+                        {{ result.source_title }}
+                    {% endif %}
+                </h2>
+                <p class="meta">
+                    {% if result.source_author %}{{ result.source_author }}{% endif %}
+                    {% if result.chapter_title %} · {{ result.chapter_title }}{% endif %}
+                    {% if result.page_label %} · page {{ result.page_label }}{% endif %}
+                </p>
+                <p>{{ result.text }}</p>
+                <dl class="scores">
+                    <div>
+                        <dt>final</dt>
+                        <dd>{{ "%.3f"|format(result.score) }}</dd>
+                    </div>
+                    {% if result.rerank_score is not none %}
+                        <div>
+                            <dt>rerank</dt>
+                            <dd>{{ "%.3f"|format(result.rerank_score) }}</dd>
+                        </div>
+                    {% endif %}
+                    {% if result.vector_score is not none %}
+                        <div>
+                            <dt>vector cosine</dt>
+                            <dd>{{ "%.3f"|format(result.vector_score) }}</dd>
+                        </div>
+                    {% endif %}
+                    {% if result.bm25_score is not none %}
+                        <div>
+                            <dt>BM25</dt>
+                            <dd>{{ "%.6f"|format(result.bm25_score) }}</dd>
+                        </div>
+                    {% endif %}
+                    {% if result.fused_score is not none %}
+                        <div>
+                            <dt>RRF</dt>
+                            <dd>{{ "%.3f"|format(result.fused_score) }}</dd>
+                        </div>
+                    {% endif %}
+                </dl>
+                {% if result.matched_phrases %}
+                    <p class="phrase-matches">
+                        <span class="phrase-matches-label">boosted by</span>
+                        {% for phrase in result.matched_phrases %}
+                            <span class="phrase-match">{{ phrase }}</span>
+                        {% endfor %}
+                    </p>
+                {% endif %}
+            </li>
+        {% endfor %}
+    </ol>
+{% else %}
+    <p>No results.</p>
+{% endif %}

python/ebook_search/api/templates/search.html

@@ -0,0 +1,32 @@
+{% extends "base.html" %}
+
+{% block title %}EPUB Search{% endblock %}
+{% block head %}<script src="https://unpkg.com/htmx.org@2.0.4"></script>{% endblock %}
+
+{% block content %}
+    <h1>Search</h1>
+    <form class="card" hx-post="/search" hx-target="#results" hx-swap="innerHTML">
+        <label for="query">What are you looking for?</label>
+        <textarea id="query" name="query" rows="4" placeholder="Ask a question or paste a passage…" required
+            onkeydown="if (event.key === 'Enter' && !event.shiftKey) { event.preventDefault(); this.form.requestSubmit(); }"></textarea>
+        <div class="form-row">
+            <div class="search-toggles">
+                <label class="check">
+                    <input type="checkbox" name="rerank" value="true" {% if config.rerank.enabled %}checked{% endif %}>
+                    Rerank
+                </label>
+                <label class="check">
+                    <input
+                        type="checkbox"
+                        name="phrase_matching"
+                        value="true"
+                        {% if config.phrase_matching_enabled %}checked{% endif %}
+                    >
+                    Phrase matching
+                </label>
+            </div>
+            <button type="submit">Search</button>
+        </div>
+    </form>
+    <section id="results"></section>
+{% endblock %}

python/ebook_search/api/web.py

@@ -0,0 +1,23 @@
+"""Shared web UI resources for EPUB search."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from fastapi.templating import Jinja2Templates
+
+PACKAGE_DIR = Path(__file__).resolve().parent
+TEMPLATE_DIR = PACKAGE_DIR / "templates"
+STATIC_DIR = PACKAGE_DIR / "static"
+
+
+def static_version(filename: str) -> int:
+    """Return a cache-busting token for a static file based on its modification time."""
+    try:
+        return int((STATIC_DIR / filename).stat().st_mtime)
+    except OSError:
+        return 0
+
+
+templates = Jinja2Templates(directory=TEMPLATE_DIR)
+templates.env.globals["static_version"] = static_version

python/ebook_search/bm25_corpus.py

@@ -0,0 +1,286 @@
+"""Persisted BM25 corpus management."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import shutil
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from functools import cache
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import bm25s
+from sqlalchemy import func, select, union_all
+
+from python.orm.richie import EbookChapter, EbookChunk, EbookSource
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+
+logger = logging.getLogger(__name__)
+MANIFEST_NAME = "manifest.json"
+REQUIRED_INDEX_FILES = frozenset(
+    {
+        "data.csc.index.npy",
+        "indices.csc.index.npy",
+        "indptr.csc.index.npy",
+        "params.index.json",
+        "vocab.index.json",
+        "corpus.jsonl",
+    }
+)
+
+
+@dataclass(frozen=True)
+class BM25Manifest:
+    """Metadata describing a persisted BM25 corpus."""
+
+    created_at: datetime
+    db_updated_at: datetime | None
+    chunk_count: int
+
+
+@dataclass(frozen=True)
+class BM25Corpus:
+    """Loaded persisted BM25 corpus and retriever."""
+
+    retriever: object | None
+    records: tuple[dict[str, object], ...]
+    manifest: BM25Manifest
+
+
+class BM25CorpusUnavailableError(RuntimeError):
+    """Raised when the persisted BM25 corpus cannot be loaded."""
+
+
+def bm25_index_path(config: EbookSearchConfig) -> Path:
+    """Return the configured BM25 index root path relative to the current working directory."""
+    path = Path(config.bm25_index_dir).expanduser()
+    if path.is_absolute():
+        return path
+    return Path.cwd() / path
+
+
+def get_current_bm25_index(index_path: Path) -> Path:
+    """Return the live BM25 index directory."""
+    current_path = index_path / "current"
+    if current_path.exists() or current_path.is_symlink():
+        return current_path
+    return index_path
+
+
+async def ensure_bm25_corpus(session: AsyncSession, config: EbookSearchConfig) -> None:
+    """Create or refresh the persisted BM25 corpus when it is missing or stale."""
+    index_path = bm25_index_path(config)
+    manifest = read_bm25_manifest(index_path)
+    db_updated_at = await corpus_last_updated_at(session)
+    if not bm25_index_exists(index_path, manifest):
+        logger.info("ebook_bm25_index_missing path=%s", index_path)
+        await refresh_bm25_corpus(session, config, db_updated_at=db_updated_at)
+        return
+    if db_updated_at is not None and manifest is not None and manifest.created_at < db_updated_at:
+        logger.info(
+            "ebook_bm25_index_stale path=%s created_at=%s db_updated_at=%s",
+            index_path,
+            manifest.created_at.isoformat(),
+            db_updated_at.isoformat(),
+        )
+        await refresh_bm25_corpus(session, config, db_updated_at=db_updated_at)
+        return
+    logger.info(
+        "ebook_bm25_index_current path=%s chunks=%s created_at=%s",
+        index_path,
+        manifest.chunk_count if manifest else 0,
+        manifest.created_at.isoformat() if manifest else None,
+    )
+
+
+async def refresh_bm25_corpus(
+    session: AsyncSession,
+    config: EbookSearchConfig,
+    *,
+    db_updated_at: datetime | None = None,
+) -> BM25Manifest:
+    """Rebuild and persist the BM25 corpus from the current database chunks.
+
+    The index build is CPU and disk work, so it runs in a worker thread.
+    """
+    index_path = bm25_index_path(config)
+    records, texts = await fetch_bm25_corpus_records(session)
+    manifest = BM25Manifest(
+        created_at=datetime.now(tz=UTC),
+        db_updated_at=db_updated_at if db_updated_at is not None else await corpus_last_updated_at(session),
+        chunk_count=len(records),
+    )
+    await asyncio.to_thread(write_bm25_corpus, index_path, records, texts, manifest)
+    logger.info(
+        "ebook_bm25_index_refreshed path=%s chunks=%s created_at=%s",
+        index_path,
+        manifest.chunk_count,
+        manifest.created_at.isoformat(),
+    )
+    return manifest
+
+
+@cache
+def load_bm25_corpus(config: EbookSearchConfig) -> BM25Corpus:
+    """Load the BM25 corpus into memory once per process.
+
+    Background refresh tasks clear this cache after rebuilding the on-disk corpus.
+    """
+    index_path = bm25_index_path(config)
+    active_index_path = get_current_bm25_index(index_path)
+    logger.info("ebook_bm25_corpus_cache_load path=%s active_path=%s", index_path, active_index_path)
+    manifest = read_bm25_manifest(index_path)
+    if manifest is None or not bm25_index_exists(index_path, manifest):
+        msg = f"BM25 corpus is not available: {index_path}"
+        raise BM25CorpusUnavailableError(msg)
+    if manifest.chunk_count == 0:
+        return BM25Corpus(retriever=None, records=(), manifest=manifest)
+
+    retriever = bm25s.BM25.load(active_index_path, load_corpus=True, mmap=True)
+    records = tuple(dict(record) for record in retriever.corpus)
+    return BM25Corpus(retriever=retriever, records=records, manifest=manifest)
+
+
+def score_bm25_corpus(query: str, corpus: BM25Corpus, *, limit: int) -> list[tuple[dict[str, object], float]]:
+    """Score a query against a loaded BM25 corpus."""
+    if corpus.retriever is None or not corpus.records:
+        return []
+    k = min(limit, len(corpus.records))
+    documents, scores = corpus.retriever.retrieve(
+        bm25s.tokenize(query, show_progress=False),
+        corpus=list(corpus.records),
+        k=k,
+        show_progress=False,
+    )
+    results: list[tuple[dict[str, object], float]] = []
+    for document, score in zip(documents[0], scores[0], strict=True):
+        score_value = float(score)
+        if score_value <= 0:
+            continue
+        results.append((dict(document), score_value))
+    return results
+
+
+async def fetch_bm25_corpus_records(session: AsyncSession) -> tuple[list[dict[str, object]], list[str]]:
+    """Fetch persistable BM25 corpus records and their matching index texts from the database.
+
+    search_text is only needed to build the index, so it is returned separately instead of
+    being persisted into the corpus records, which would double the corpus size.
+    """
+    statement = (
+        select(
+            EbookChunk.id.label("chunk_id"),
+            EbookChunk.text.label("text"),
+            EbookSource.id.label("source_id"),
+            EbookSource.title.label("source_title"),
+            EbookSource.author.label("source_author"),
+            EbookChapter.title.label("chapter_title"),
+            EbookChunk.page_label.label("page_label"),
+            EbookChunk.search_text.label("bm25_text"),
+        )
+        .select_from(EbookChunk)
+        .join(EbookSource, EbookSource.id == EbookChunk.source_id)
+        .outerjoin(EbookChapter, EbookChapter.id == EbookChunk.chapter_id)
+        .order_by(EbookChunk.id)
+    )
+    records: list[dict[str, object]] = []
+    texts: list[str] = []
+    for row in (await session.execute(statement)).mappings():
+        record = dict(row)
+        texts.append(str(record.pop("bm25_text")))
+        records.append(record)
+    return records, texts
+
+
+async def corpus_last_updated_at(session: AsyncSession) -> datetime | None:
+    """Return the latest source/chapter/chunk update timestamp relevant to BM25 text."""
+    update_times = union_all(
+        select(func.max(EbookSource.updated).label("updated")),
+        select(func.max(EbookChapter.updated).label("updated")),
+        select(func.max(EbookChunk.updated).label("updated")),
+    ).subquery()
+    return await session.scalar(select(func.max(update_times.c.updated)))
+
+
+def write_bm25_corpus(
+    index_path: Path,
+    records: list[dict[str, object]],
+    texts: list[str],
+    manifest: BM25Manifest,
+) -> None:
+    """Write a BM25 corpus generation and publish it through the current symlink."""
+    index_path.mkdir(parents=True, exist_ok=True)
+
+    generations_path = index_path / "generations"
+    generations_path.mkdir(exist_ok=True)
+
+    generation_path = next_bm25_generation_path(generations_path, manifest.created_at)
+    current_path = index_path / "current"
+    next_current_path = index_path / f".current.{generation_path.name}.tmp"
+    try:
+        generation_path.mkdir()
+
+        # Empty corpora publish a manifest-only generation so startup succeeds before any chunks exist.
+        if records:
+            retriever = bm25s.BM25()
+            retriever.index(bm25s.tokenize(texts, show_progress=False), show_progress=False)
+            retriever.save(generation_path, corpus=records, show_progress=False)
+        write_bm25_manifest(generation_path, manifest)
+        next_current_path.unlink(missing_ok=True)
+        next_current_path.symlink_to(generation_path, target_is_directory=True)
+        next_current_path.replace(current_path)
+    except Exception:
+        next_current_path.unlink(missing_ok=True)
+        shutil.rmtree(generation_path, ignore_errors=True)
+        raise
+
+
+def read_bm25_manifest(index_path: Path) -> BM25Manifest | None:
+    """Read the BM25 manifest if it exists and is valid."""
+    manifest_path = get_current_bm25_index(index_path) / MANIFEST_NAME
+    if not manifest_path.exists():
+        return None
+    body = json.loads(manifest_path.read_text(encoding="utf-8"))
+    return BM25Manifest(
+        created_at=datetime.fromisoformat(str(body["created_at"])),
+        db_updated_at=datetime.fromisoformat(str(body["db_updated_at"])) if body.get("db_updated_at") else None,
+        chunk_count=int(body["chunk_count"]),
+    )
+
+
+def write_bm25_manifest(index_path: Path, manifest: BM25Manifest) -> None:
+    """Write the BM25 manifest to an index directory."""
+    body = {
+        "created_at": manifest.created_at.isoformat(),
+        "db_updated_at": manifest.db_updated_at.isoformat() if manifest.db_updated_at else None,
+        "chunk_count": manifest.chunk_count,
+    }
+    (index_path / MANIFEST_NAME).write_text(json.dumps(body, indent=2, sort_keys=True), encoding="utf-8")
+
+
+def bm25_index_exists(index_path: Path, manifest: BM25Manifest | None) -> bool:
+    """Return whether a usable persisted BM25 index exists."""
+    active_index_path = get_current_bm25_index(index_path)
+    if manifest is None or not active_index_path.is_dir():
+        return False
+    if manifest.chunk_count == 0:
+        return True
+    return all((active_index_path / file_name).exists() for file_name in REQUIRED_INDEX_FILES)
+
+
+def next_bm25_generation_path(generations_path: Path, created_at: datetime) -> Path:
+    """Return an unused dated BM25 generation path."""
+    base_name = created_at.astimezone(UTC).strftime("%Y%m%dT%H%M%S.%fZ")
+    generation_path = generations_path / base_name
+    suffix = 1
+    while generation_path.exists():
+        generation_path = generations_path / f"{base_name}.{suffix}"
+        suffix += 1
+    return generation_path

python/ebook_search/config.py

@@ -0,0 +1,145 @@
+"""Configuration for the EPUB search app."""
+
+from __future__ import annotations
+
+from os import getenv
+from typing import Annotated, Self
+
+from pydantic import AliasChoices, Field, field_validator, model_validator
+from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict
+
+
+def normalize_embedding_alias(model: str) -> str:
+    """Normalize a supported embedding alias to its provider model name."""
+    aliases = {
+        "Qwen3-Embedding-0.6B": "qwen3-embedding-0.6b",
+        "Qwen3-Embedding-4B": "qwen3-embedding-4b",
+        "Qwen3-Embedding-8B": "qwen3-embedding-8b",
+        "Qwen/Qwen3-Embedding-0.6B": "qwen3-embedding-0.6b",
+        "Qwen/Qwen3-Embedding-4B": "qwen3-embedding-4b",
+        "Qwen/Qwen3-Embedding-8B": "qwen3-embedding-8b",
+        "qwen3-embedding:0.6b": "qwen3-embedding-0.6b",
+        "qwen3-embedding:4b": "qwen3-embedding-4b",
+        "qwen3-embedding:8b": "qwen3-embedding-8b",
+        "qwen3-embedding-0.6b": "qwen3-embedding-0.6b",
+        "qwen3-embedding-4b": "qwen3-embedding-4b",
+        "qwen3-embedding-8b": "qwen3-embedding-8b",
+    }
+    standard_model = aliases.get(model)
+    if standard_model is None:
+        error = f"Embedding model {model} is not supported. Supported models are {aliases.keys()}"
+        raise ValueError(error)
+    return standard_model
+
+
+def normalize_embedding_model(default: str = "qwen3-embedding-0.6b") -> str:
+    """Normalize the configured embedding alias to its provider model name."""
+    return normalize_embedding_alias(getenv("EBOOK_SEARCH_EMBEDDING_MODEL", default))
+
+
+class RerankConfig(BaseSettings):
+    """vLLM reranker settings."""
+
+    model_config = SettingsConfigDict(env_prefix="EBOOK_SEARCH_RERANK_", frozen=True, protected_namespaces=())
+
+    enabled: bool = True
+    base_url: str = "http://192.168.90.25:8001"
+    model: str = "qwen3-reranker-06b"
+    candidates: int = 24
+    timeout_seconds: float = 30.0
+    score_weight: float = 0.7
+    hybrid_weight: float = 0.3
+
+
+class EbookSearchConfig(BaseSettings):
+    """Runtime settings for EPUB search."""
+
+    model_config = SettingsConfigDict(
+        env_prefix="EBOOK_SEARCH_",
+        frozen=True,
+        populate_by_name=True,
+        protected_namespaces=(),
+    )
+
+    rerank: RerankConfig = Field(default_factory=RerankConfig)
+    top_k: int = 12
+    library_paths: Annotated[tuple[str, ...], NoDecode] = ()
+    chunk_tokens: int = 700
+    chunk_overlap: int = 100
+    vllm_base_url: str = "https://ollama.com/v1"
+    vllm_api_key: str = Field(
+        default="not-needed",
+        validation_alias=AliasChoices("EBOOK_SEARCH_VLLM_API_KEY", "OLLAMA_API_KEY"),
+    )
+    chat_model: str = "deepseek-v4-flash"
+    answer_enabled: bool = True
+    embedding_base_url: str = "http://192.168.90.25:8000/v1"
+    embedding_api_key: str = "not-needed"
+    embedding_model: str = "qwen3-embedding-0.6b"
+    embedding_batch_size: int = 32
+    embedding_timeout_seconds: float = 60.0
+    chat_timeout_seconds: float = 60.0
+    vector_candidate_multiplier: int = 4
+    bm25_candidate_limit: int = 120
+    rrf_rank_constant: int = 60
+    min_retrieval_confidence: float = 0.0
+    validate_citations_enabled: bool = True
+    bm25_index_dir: str = ".ebook_search_bm25"
+    bm25_refresh_delay_seconds: int = 60
+    protected_phrase_max_candidates_per_book: int = 5000
+    protected_phrase_llm_candidates_per_book: int = 500
+    protected_phrase_extraction_workers: int = 16
+    phrase_judge_book_workers: int = 20
+    phrase_judge_phrase_workers: int = 100
+    protected_phrase_confidence_threshold: float = 0.80
+    phrase_matching_enabled: bool = True
+    phrase_hit_boost: float = 0.25
+    phrase_min_tokens: int = 2
+    phrase_max_tokens: int = 5
+    phrase_max_entity_tokens: int = 8
+    phrase_raw_ngram_min_count: int = 2
+    phrase_raw_count_score_threshold: int = 3
+    phrase_raw_count_high_score_threshold: int = 10
+    phrase_chapter_count_score_threshold: int = 2
+    phrase_chapter_count_high_score_threshold: int = 5
+    phrase_target_protected_per_book: int = 100
+    phrase_default_allow_nested: bool = False
+    phrase_default_suppress_children: bool = True
+
+    @field_validator("library_paths", mode="before")
+    @classmethod
+    def split_library_paths(cls, value: object) -> object:
+        """Split a colon-separated library path string into a tuple of paths."""
+        if isinstance(value, str):
+            return tuple(path for path in value.split(":") if path)
+        return value
+
+    @field_validator("embedding_model")
+    @classmethod
+    def normalize_embedding(cls, value: str) -> str:
+        """Normalize the configured embedding alias to its provider model name."""
+        return normalize_embedding_alias(value)
+
+    @model_validator(mode="after")
+    def validate_runtime_consistency(self) -> Self:
+        """Reject configurations that cannot serve the features they enable."""
+        if not self.embedding_base_url.strip():
+            msg = "embedding_base_url must be set"
+            raise ValueError(msg)
+        if self.answer_enabled and (not self.vllm_base_url.strip() or not self.chat_model.strip()):
+            msg = "answer_enabled requires vllm_base_url and chat_model to be set"
+            raise ValueError(msg)
+        if self.rerank.enabled and not self.rerank.base_url.strip():
+            msg = "rerank.enabled requires rerank.base_url to be set"
+            raise ValueError(msg)
+        return self
+
+
+def load_rerank_config() -> RerankConfig:
+    """Load reranker config from environment variables."""
+    return RerankConfig()
+
+
+def load_config() -> EbookSearchConfig:
+    """Load EPUB search config from environment variables."""
+    return EbookSearchConfig()

python/ebook_search/docker/Dockerfile

@@ -0,0 +1,49 @@
+FROM python:3.14-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    APP_DIR=/home/richie/dotfiles \
+    EBOOK_SEARCH_HOST=0.0.0.0 \
+    EBOOK_SEARCH_PORT=8070 \
+    EBOOK_SEARCH_BM25_INDEX_DIR=/data/bm25
+
+WORKDIR ${APP_DIR}
+
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends build-essential curl \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY pyproject.toml README.md LICENSE ./
+COPY python ./python
+
+RUN python -m pip install --upgrade pip \
+    && python -m pip install \
+        "alembic" \
+        "beautifulsoup4" \
+        "bm25s" \
+        "ebooklib" \
+        "fastapi" \
+        "httpx" \
+        "jinja2" \
+        "pgvector" \
+        "psycopg[binary]" \
+        "pydantic" \
+        "pydantic-settings" \
+        "python-multipart" \
+        "sqlalchemy[asyncio]" \
+        "tiktoken" \
+        "typer" \
+        "uvicorn[standard]" \
+        "yake" \
+    && python -m pip install --no-deps --editable "${APP_DIR}"
+
+RUN useradd --create-home --uid 10001 app \
+    && mkdir -p /data \
+    && chown -R app:app /home/richie /data
+
+USER app
+
+EXPOSE 8070
+
+CMD ["sh", "-c", "exec python -m python.ebook_search.api.main --host \"${EBOOK_SEARCH_HOST}\" --port \"${EBOOK_SEARCH_PORT}\" --log-level \"${EBOOK_SEARCH_LOG_LEVEL:-INFO}\""]

python/ebook_search/docker/README.md

@@ -0,0 +1,40 @@
+# Ebook Search Docker
+
+Run the EPUB search app against the existing Postgres database on `jeeves`:
+
+```sh
+ebook-search-containers start --library-path /path/to/epubs --build
+```
+
+All ebook-search Docker files live in this directory:
+
+- `Dockerfile`
+- `docker-compose.yml`
+- `containers.py`
+- `container.py`
+
+The app listens on `http://localhost:8070`.
+
+Useful lifecycle commands:
+
+```sh
+ebook-search-containers build
+ebook-search-containers start --library-path /path/to/epubs
+ebook-search-containers logs
+ebook-search-containers ps
+ebook-search-containers stop
+```
+
+Direct compose usage from the repo root:
+
+```sh
+docker compose -f python/ebook_search/docker/docker-compose.yml ps
+```
+
+The compose service also loads the repo root `.env` into the container via `env_file`.
+
+Mount your EPUB directory by setting `EBOOK_LIBRARY_HOST_PATH` in an env file or on the command line. The container sees it as `/library`, and `EBOOK_SEARCH_LIBRARY_PATHS` is set to `/library` inside the container.
+
+Database connection settings are controlled by `RICHIE_DB`, `RICHIE_HOST`, `RICHIE_PORT`, `RICHIE_USER`, and `RICHIE_PASSWORD`. The default host is `jeeves`.
+
+Startup runs the Richie Alembic migrations automatically after creating the `main` schema and `vector` extension.

python/ebook_search/docker/__init__.py

@@ -0,0 +1 @@
+"""Docker packaging and lifecycle tooling for ebook search."""

python/ebook_search/docker/containers.py

@@ -0,0 +1,229 @@
+"""Docker container lifecycle management for ebook search."""
+
+from __future__ import annotations
+
+import logging
+import os
+import subprocess
+from pathlib import Path
+from typing import Annotated
+
+import typer
+
+from python.common import configure_logger, get_repo_dir
+
+logger = logging.getLogger(__name__)
+
+
+def get_compose_file() -> Path:
+    """Return the path to the docker-compose.yml file."""
+    return Path(__file__).resolve().with_name("docker-compose.yml")
+
+
+def compose_base_args() -> list[str]:
+    """Return the common docker compose arguments for the ebook search stack."""
+    return ["compose", "-f", str(get_compose_file())]
+
+
+def docker_run(
+    arguments: list[str],
+    *,
+    env: dict[str, str] | None = None,
+    capture_output: bool = False,
+) -> subprocess.CompletedProcess[str]:
+    """Run docker with repo-root cwd and consistent error handling."""
+    logger.info("docker %s", " ".join(arguments))
+    return subprocess.run(
+        ["docker", *arguments],
+        cwd=get_repo_dir(),
+        env=env,
+        text=True,
+        check=False,
+        capture_output=capture_output,
+    )
+
+
+def compose_env(*, library_path: Path | None = None, port: int | None = None) -> dict[str, str]:
+    """Return environment variables passed to docker compose."""
+    env = os.environ.copy()
+    if library_path is not None:
+        resolved_library = library_path.expanduser().resolve()
+        if not resolved_library.exists():
+            msg = f"EPUB library path does not exist: {resolved_library}"
+            raise FileNotFoundError(msg)
+        env["EBOOK_LIBRARY_HOST_PATH"] = str(resolved_library)
+    if port is not None:
+        env["EBOOK_SEARCH_PORT"] = str(port)
+    return env
+
+
+def ensure_compose_file() -> None:
+    """Raise if the ebook search compose file is missing."""
+    if not get_compose_file().is_file():
+        msg = f"Compose file not found: {get_compose_file()}"
+        raise FileNotFoundError(msg)
+
+
+def build_image() -> None:
+    """Build the ebook search app image."""
+    ensure_compose_file()
+    result = docker_run([*compose_base_args(), "build"])
+    if result.returncode != 0:
+        msg = "Failed to build ebook search image"
+        raise RuntimeError(msg)
+
+
+def start_stack(
+    *,
+    library_path: Path | None = None,
+    port: int | None = None,
+    build: bool = False,
+) -> None:
+    """Start the ebook search Docker compose stack."""
+    ensure_compose_file()
+    env = compose_env(library_path=library_path, port=port)
+    if build:
+        build_image()
+    result = docker_run(
+        [*compose_base_args(), "up", "-d"],
+        env=env,
+    )
+    if result.returncode != 0:
+        msg = f"Ebook search stack failed to start with code {result.returncode}"
+        raise RuntimeError(msg)
+    logger.info("Ebook search started.")
+
+
+def stop_stack(
+    *,
+    volumes: bool = False,
+) -> None:
+    """Stop and remove ebook search containers."""
+    ensure_compose_file()
+    command = [*compose_base_args(), "down"]
+    if volumes:
+        command.append("-v")
+    result = docker_run(command)
+    if result.returncode != 0:
+        msg = f"Ebook search stack failed to stop with code {result.returncode}"
+        raise RuntimeError(msg)
+
+
+def logs_stack(
+    *,
+    service: str | None = None,
+    tail: int = 100,
+    follow: bool = False,
+) -> str | None:
+    """Return recent logs from the ebook search stack."""
+    ensure_compose_file()
+    command = [*compose_base_args(), "logs", "--tail", str(tail)]
+    if follow:
+        command.append("--follow")
+    if service:
+        command.append(service)
+    result = docker_run(command, capture_output=not follow)
+    if result.returncode != 0:
+        return None
+    if follow:
+        return ""
+    return result.stdout + result.stderr
+
+
+def ps_stack() -> str | None:
+    """Return docker compose ps output for the ebook search stack."""
+    ensure_compose_file()
+    result = docker_run([*compose_base_args(), "ps"], capture_output=True)
+    if result.returncode != 0:
+        return None
+    return result.stdout + result.stderr
+
+
+app = typer.Typer(help="Ebook search Docker container management.", no_args_is_help=True)
+
+
+@app.command()
+def build() -> None:
+    """Build the ebook search Docker image."""
+    build_image()
+
+
+@app.command()
+def start(
+    library_path: Annotated[Path | None, typer.Option(help="Override host path containing EPUB files.")] = None,
+    port: Annotated[int | None, typer.Option(help="Override host port for the web UI.")] = None,
+    *,
+    build: Annotated[bool, typer.Option("--build", help="Build the image before starting.")] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+) -> None:
+    """Start the ebook search container."""
+    configure_logger(log_level)
+    start_stack(
+        library_path=library_path,
+        port=port,
+        build=build,
+    )
+
+
+@app.command()
+def stop(
+    *,
+    volumes: Annotated[bool, typer.Option("--volumes", help="Also remove ebook search data volumes.")] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+) -> None:
+    """Stop and remove ebook search containers."""
+    configure_logger(log_level)
+    stop_stack(volumes=volumes)
+
+
+@app.command()
+def restart(
+    library_path: Annotated[Path | None, typer.Option(help="Override host path containing EPUB files.")] = None,
+    port: Annotated[int | None, typer.Option(help="Override host port for the web UI.")] = None,
+    *,
+    build: Annotated[bool, typer.Option("--build", help="Build the image before starting.")] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+) -> None:
+    """Restart the ebook search stack."""
+    configure_logger(log_level)
+    stop_stack()
+    start_stack(
+        library_path=library_path,
+        port=port,
+        build=build,
+    )
+
+
+@app.command()
+def logs(
+    service: Annotated[str | None, typer.Option(help="Service name, or omit for all services.")] = None,
+    tail: Annotated[int, typer.Option(help="Number of recent log lines.")] = 100,
+    *,
+    follow: Annotated[bool, typer.Option("--follow", "-f", help="Follow logs.")] = False,
+) -> None:
+    """Show recent ebook search container logs."""
+    output = logs_stack(service=service, tail=tail, follow=follow)
+    if output is None:
+        typer.echo("No ebook search containers found.")
+        raise typer.Exit(code=1)
+    if output:
+        typer.echo(output)
+
+
+@app.command("ps")
+def ps() -> None:
+    """Show ebook search container status."""
+    output = ps_stack()
+    if output is None:
+        typer.echo("No ebook search containers found.")
+        raise typer.Exit(code=1)
+    typer.echo(output)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    app()
+
+
+if __name__ == "__main__":
+    cli()

python/ebook_search/docker/docker-compose.yml

@@ -0,0 +1,36 @@
+name: ebook-search
+
+services:
+  ebook-search:
+    build:
+      context: ../../..
+      dockerfile: python/ebook_search/docker/Dockerfile
+    image: ebook-search:latest
+    restart: unless-stopped
+    ports:
+      - "${EBOOK_SEARCH_PORT:-8070}:8070"
+    extra_hosts:
+      - "jeeves:192.168.90.40"
+    env_file:
+      - ../../../.env
+    environment:
+      EBOOK_SEARCH_HOST: "0.0.0.0"
+      EBOOK_SEARCH_PORT: "8070"
+      EBOOK_SEARCH_LIBRARY_PATHS: "/library"
+      EBOOK_SEARCH_BM25_INDEX_DIR: "/data/bm25"
+    volumes:
+      - "${EBOOK_LIBRARY_HOST_PATH:-/home/richie/ebooks}:/library:ro"
+      - ebook-search-data:/data
+    healthcheck:
+      test:
+        [
+          "CMD-SHELL",
+          "curl -fsS http://127.0.0.1:8070/health >/dev/null || exit 1",
+        ]
+      interval: 30s
+      timeout: 5s
+      retries: 5
+      start_period: 30s
+
+volumes:
+  ebook-search-data:

python/ebook_search/embeddings.py

@@ -0,0 +1,175 @@
+"""Embedding model helpers."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from sqlalchemy import func, select
+from sqlalchemy.dialects.postgresql import insert
+
+from python.ebook_search.llm_interface import request_embeddings
+from python.orm.richie import (
+    EbookChunk,
+    EbookChunkEmbedding1024,
+    EbookChunkEmbedding2560,
+    EbookChunkEmbedding4096,
+    EbookEmbeddingModel,
+)
+
+logger = logging.getLogger(__name__)
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    import httpx
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+
+MODEL_DIMENSIONS = {
+    "qwen3-embedding-0.6b": 1024,
+    "qwen3-embedding-4b": 2560,
+    "qwen3-embedding-8b": 4096,
+}
+
+
+def get_embedding_table(
+    dimension: int,
+) -> type[EbookChunkEmbedding1024 | EbookChunkEmbedding2560 | EbookChunkEmbedding4096]:
+    """Return the embedding table mapped to an embedding dimension."""
+    embedding_tables = {
+        1024: EbookChunkEmbedding1024,
+        2560: EbookChunkEmbedding2560,
+        4096: EbookChunkEmbedding4096,
+    }
+    table = embedding_tables.get(dimension)
+    if not table:
+        msg = f"Embedding dimension {dimension} is not supported"
+        raise ValueError(msg)
+    return table
+
+
+@dataclass(frozen=True)
+class EmbeddingModelStats:
+    """Embedding coverage for one model."""
+
+    model_name: str
+    dimension: int
+    embedded_chunks: int
+    total_chunks: int
+
+    @property
+    def missing_chunks(self) -> int:
+        """Return chunks missing this embedding model."""
+        return max(self.total_chunks - self.embedded_chunks, 0)
+
+
+async def embed_texts(
+    client: httpx.AsyncClient,
+    texts: Sequence[str],
+    config: EbookSearchConfig,
+) -> list[list[float]]:
+    """Embed text with the configured vLLM embedding model."""
+    logger.info(
+        "ebook_embed_request_start base_url=%s model=%s count=%s",
+        config.embedding_base_url,
+        config.embedding_model,
+        len(texts),
+    )
+    vectors = await request_embeddings(client, texts, config)
+    expected_dimension = MODEL_DIMENSIONS[config.embedding_model]
+    for vector in vectors:
+        if len(vector) != expected_dimension:
+            msg = f"Expected {expected_dimension} dimensions, got {len(vector)}"
+            raise ValueError(msg)
+    logger.info(
+        "ebook_embed_request_complete model=%s count=%s dimension=%s",
+        config.embedding_model,
+        len(vectors),
+        expected_dimension,
+    )
+    return vectors
+
+
+async def embed_query(client: httpx.AsyncClient, query: str, config: EbookSearchConfig) -> list[float]:
+    """Embed a search query with the Qwen retrieval instruction."""
+    instructed_query = f"Instruct: Retrieve relevant passages for the query.\nQuery: {query}"
+    return (await embed_texts(client, [instructed_query], config))[0]
+
+
+async def ensure_embedding_models(session: AsyncSession) -> None:
+    """Ensure supported embedding model rows exist."""
+    for name, dimension in MODEL_DIMENSIONS.items():
+        existing = await session.scalar(select(EbookEmbeddingModel).where(EbookEmbeddingModel.name == name))
+        if existing is None:
+            session.add(EbookEmbeddingModel(name=name, dimension=dimension, is_default=name == "qwen3-embedding-0.6b"))
+            logger.info("ebook_embedding_model_created model=%s dimension=%s", name, dimension)
+    await session.flush()
+
+
+async def embedding_model_stats(session: AsyncSession) -> list[EmbeddingModelStats]:
+    """Return embedding coverage counts for every supported model."""
+    total_chunks = await session.scalar(select(func.count(EbookChunk.id))) or 0
+    models = {
+        model.name: model
+        for model in await session.scalars(
+            select(EbookEmbeddingModel)
+            .where(EbookEmbeddingModel.name.in_(MODEL_DIMENSIONS))
+            .order_by(EbookEmbeddingModel.name)
+        )
+    }
+
+    stats: list[EmbeddingModelStats] = []
+    for model_name, dimension in MODEL_DIMENSIONS.items():
+        model = models.get(model_name)
+        embedded_chunks = 0
+        if model is not None:
+            table = get_embedding_table(dimension)
+            embedded_chunks = await session.scalar(select(func.count(table.id)).where(table.model_id == model.id)) or 0
+        stats.append(
+            EmbeddingModelStats(
+                model_name=model_name,
+                dimension=dimension,
+                embedded_chunks=embedded_chunks,
+                total_chunks=total_chunks,
+            )
+        )
+    return stats
+
+
+async def embed_missing_chunks(session: AsyncSession, client: httpx.AsyncClient, config: EbookSearchConfig) -> int:
+    """Embed chunks missing embeddings for the configured model."""
+    await ensure_embedding_models(session)
+    model = await session.scalar(select(EbookEmbeddingModel).where(EbookEmbeddingModel.name == config.embedding_model))
+    if model is None:
+        supported_models = ", ".join(MODEL_DIMENSIONS)
+        msg = f"Unknown embedding model: {config.embedding_model}. Supported models: {supported_models}"
+        raise ValueError(msg)
+
+    table = get_embedding_table(model.dimension)
+    chunks = list(
+        await session.scalars(
+            select(EbookChunk)
+            .outerjoin(table, (table.chunk_id == EbookChunk.id) & (table.model_id == model.id))
+            .where(table.id.is_(None))
+            .order_by(EbookChunk.id)
+            .limit(config.embedding_batch_size)
+        )
+    )
+    if not chunks:
+        logger.info("ebook_embed_missing_none model=%s", config.embedding_model)
+        return 0
+
+    logger.info("ebook_embed_missing_batch_start model=%s count=%s", config.embedding_model, len(chunks))
+    vectors = await embed_texts(client, [chunk.text for chunk in chunks], config)
+    rows = [
+        {"chunk_id": chunk.id, "model_id": model.id, "embedding": vector}
+        for chunk, vector in zip(chunks, vectors, strict=True)
+    ]
+    statement = insert(table).values(rows).on_conflict_do_nothing(index_elements=["chunk_id", "model_id"])
+    await session.execute(statement)
+    await session.flush()
+    logger.info("ebook_embed_missing_batch_complete model=%s count=%s", config.embedding_model, len(rows))
+    return len(rows)

python/ebook_search/epub_parse.py

@@ -0,0 +1,95 @@
+"""EPUB parsing helpers."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from bs4 import BeautifulSoup
+from ebooklib import ITEM_DOCUMENT, epub
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+WHITESPACE_RE = re.compile(r"\s+")
+
+
+@dataclass(frozen=True)
+class ParsedChapter:
+    """Text extracted from one EPUB spine document."""
+
+    title: str | None
+    href: str | None
+    text: str
+    page_labels: tuple[str, ...]
+
+
+@dataclass(frozen=True)
+class ParsedEpub:
+    """Parsed EPUB metadata and text."""
+
+    title: str
+    author: str | None
+    language: str | None
+    publisher: str | None
+    identifier: str | None
+    chapters: tuple[ParsedChapter, ...]
+
+
+def parse_epub(path: Path) -> ParsedEpub:
+    """Parse EPUB metadata and spine text."""
+    book = epub.read_epub(path)
+    chapters = []
+    for item in book.get_items_of_type(ITEM_DOCUMENT):
+        soup = BeautifulSoup(item.get_content(), "html.parser")
+        title = chapter_title(soup)
+        page_labels = tuple(extract_page_labels(soup))
+        text = clean_text(soup.get_text(" "))
+        if text:
+            chapters.append(ParsedChapter(title=title, href=item.get_name(), text=text, page_labels=page_labels))
+
+    return ParsedEpub(
+        title=metadata_value(book, "title") or path.stem,
+        author=metadata_value(book, "creator"),
+        language=metadata_value(book, "language"),
+        publisher=metadata_value(book, "publisher"),
+        identifier=metadata_value(book, "identifier"),
+        chapters=tuple(chapters),
+    )
+
+
+def metadata_value(book: epub.EpubBook, name: str) -> str | None:
+    """Return the first non-empty Dublin Core metadata value for a name."""
+    values = book.get_metadata("DC", name)
+    if not values:
+        return None
+    value = values[0][0]
+    return str(value).strip() or None
+
+
+def chapter_title(soup: BeautifulSoup) -> str | None:
+    """Extract the best available title from an EPUB document soup."""
+    heading = soup.find(["h1", "h2", "h3"])
+    if heading is None:
+        title = soup.find("title")
+        if title is None:
+            return None
+        return clean_text(title.get_text(" ")) or None
+    return clean_text(heading.get_text(" ")) or None
+
+
+def extract_page_labels(soup: BeautifulSoup) -> list[str]:
+    """Extract EPUB page-break labels from a document soup."""
+    labels: list[str] = []
+    for tag in soup.find_all(attrs={"epub:type": "pagebreak"}):
+        label = tag.get("title") or tag.get("aria-label") or tag.get_text(" ")
+        clean = clean_text(str(label))
+        if clean:
+            labels.append(clean)
+    return labels
+
+
+def clean_text(text: str) -> str:
+    """Normalize whitespace in extracted EPUB text."""
+    return WHITESPACE_RE.sub(" ", text).strip()

python/ebook_search/eval/__init__.py

@@ -0,0 +1 @@
+"""Offline evaluation tooling for the ebook search pipeline."""

python/ebook_search/eval/data/queries.jsonl

@@ -0,0 +1,71 @@
+{"query": "Who is Damien Montgomery and how does he become a Jump Mage?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What is a Rune Wright and why is Damien so rare?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "How does jump magic let starships travel faster than light?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What is the role of the Mage-King of Mars in the Protectorate?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What happened aboard the Blue Jay in the first Starship's Mage book?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "Who is Captain David Rice?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "How are amplifiers and simulacrums used to power a ship's jump?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What duties does a Hand of the Mage-King carry out?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "Explain the structure of the Royal Martian Navy.", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "How do mages carve runes to enchant a starship?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What threat do the Legatan rebels pose to the Protectorate?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "How does Damien handle his first command?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What is the significance of the simulacrum on a jump ship?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "Describe a mage duel in the Starship's Mage series.", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What moral conflicts does Damien face as a Hand of the Mage-King?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "How does the Protectorate keep peace among its member worlds?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "Who is the Keeper of Oaths and how does Damien work with them?", "answer": null, "answerable": true, "relevant_sources": ["Starship's Mage"]}
+{"query": "What event is known as the Onset and how does it change the world?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "Who is the main character at the start of the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How do survivors adapt after the Onset begins?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What new abilities emerge during the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "Describe the primary antagonist in the Onset series.", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How does society collapse and reorganize after the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What factions form in the aftermath of the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How does the protagonist gain power throughout the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What is the cause or origin of the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "Describe an early survival challenge faced after the Onset.", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How do the characters defend their stronghold during the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What relationships drive the protagonist's choices in the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How does the Onset escalate by the end of the first book?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What mysteries about the Onset remain unresolved?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How do the rules of the world change once the Onset takes hold?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "What weapons or tactics work best against the threats of the Onset?", "answer": null, "answerable": true, "relevant_sources": ["The Onset"]}
+{"query": "How does Bob Johansson become a von Neumann probe?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What is a replicant and why do Bob's copies have different personalities?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "Who are Riker, Homer, and Bill among the Bob clones?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What is GUPPI and how does Bob use it?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "Describe the threat posed by the Others.", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How does Bob protect and uplift the Deltans?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "Why do the replicants drift apart in personality over time?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What is the role of FAITH and the Brazilian Empire on Earth?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How does subspace communication work for the Bobs?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What happens to Bender after he goes missing?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How do the Bobs build self-replicating probes across the galaxy?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How does Bob evacuate humanity after Earth becomes uninhabitable?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "Describe the conflict between different factions of Bobs.", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What ethical dilemmas does Bob face when interfering with primitive species?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How does the original Bob differ from later generations of clones?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "How do the Bobs defeat the Others' system-harvesting fleets?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+{"query": "What role does Howard play in the human colonies?", "answer": null, "answerable": true, "relevant_sources": ["We Are Legion (We Are Bob)"]}
+// querys not it the dataset
+{"query": "How does Frodo destroy the One Ring in The Lord of the Rings?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "Who killed Dumbledore in Harry Potter and the Half-Blood Prince?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What house does Tyrion Lannister belong to in A Game of Thrones?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How does Paul Atreides control the spice on Arrakis in Dune?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What does the green light at the end of the dock mean in The Great Gatsby?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "Why does Hester Prynne wear a scarlet letter?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What does the white whale represent in Moby-Dick?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How does Elizabeth Bennet's view of Mr. Darcy change in Pride and Prejudice?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What crime does Raskolnikov commit in Crime and Punishment?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How does Katniss volunteer for the Hunger Games?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What is Winston Smith's job in Nineteen Eighty-Four?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "Who is Atticus Finch defending in To Kill a Mockingbird?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What is the capital of Australia?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How do I bake a sourdough loaf from scratch?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "Explain how photosynthesis converts sunlight into energy.", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What were the main causes of World War I?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How does compound interest work?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "How do I change a flat tire on a car?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What is the boiling point of water at sea level?", "answer": null, "answerable": false, "relevant_sources": []}
+{"query": "What is the recommended daily intake of vitamin D?", "answer": null, "answerable": false, "relevant_sources": []}

python/ebook_search/eval/dataset.py

@@ -0,0 +1,47 @@
+"""Shared query set loading for evaluation and load testing.
+
+Each JSONL record has a ``query`` and an optional reference ``answer``. ``answerable``
+marks whether the query should be answerable from the library (false for out-of-corpus
+"garbage" queries used to test the refusal path). Relevance for retrieval metrics is
+labeled at source (book) granularity in ``relevant_sources``; source titles must match
+``ebook_source.title`` values for the indexed corpus.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from pathlib import Path
+
+DEFAULT_QUERIES_PATH = Path(__file__).parent / "data" / "queries.jsonl"
+
+
+@dataclass(frozen=True)
+class GoldQuery:
+    """One labeled query shared by the eval and load-test tools."""
+
+    query: str
+    answer: str | None
+    answerable: bool
+    relevant_sources: tuple[str, ...]
+    relevant_substrings: tuple[str, ...]
+
+
+def load_gold_queries(path: Path = DEFAULT_QUERIES_PATH) -> list[GoldQuery]:
+    """Load labeled queries from a JSONL file. Blank lines and ``//`` comment lines are skipped."""
+    queries: list[GoldQuery] = []
+    for line in path.read_text(encoding="utf-8").splitlines():
+        stripped = line.strip()
+        if not stripped or stripped.startswith("//"):
+            continue
+        record = json.loads(stripped)
+        queries.append(
+            GoldQuery(
+                query=str(record["query"]),
+                answer=record.get("answer"),
+                answerable=bool(record.get("answerable", True)),
+                relevant_sources=tuple(record.get("relevant_sources", ())),
+                relevant_substrings=tuple(record.get("relevant_substrings", ())),
+            )
+        )
+    return queries

python/ebook_search/guardrails.py

@@ -0,0 +1,57 @@
+"""Serve-time output guardrails for retrieval confidence and answer citations."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.search import SearchResult
+
+CITATION_RE = re.compile(r"\[(\d+)\]")
+
+
+def retrieval_confidence(results: list[SearchResult]) -> float:
+    """Return the strongest interpretable relevance signal of the top result.
+
+    Reciprocal-rank-fusion scores are rank-based and not comparable across queries,
+    so the rerank relevance score is preferred, then vector cosine similarity, then
+    the final score.
+    """
+    if not results:
+        return 0.0
+    top = results[0]
+    if top.rerank_score is not None:
+        return top.rerank_score
+    if top.vector_score is not None:
+        return top.vector_score
+    return top.score
+
+
+def is_confident(results: list[SearchResult], config: EbookSearchConfig) -> bool:
+    """Return whether top-result confidence meets the configured threshold."""
+    return retrieval_confidence(results) >= config.min_retrieval_confidence
+
+
+@dataclass(frozen=True)
+class CitationReport:
+    """Validation summary for bracketed citation markers in a generated answer."""
+
+    cited: tuple[int, ...]
+    invalid: tuple[int, ...]
+    grounded: bool
+
+
+def validate_citations(answer: str, result_count: int) -> CitationReport:
+    """Validate bracketed citation markers against the number of shown sources.
+
+    A marker is valid when it points to a returned source (``1..result_count``).
+    ``grounded`` is true when the answer cites at least one valid source.
+    """
+    markers = sorted({int(match.group(1)) for match in CITATION_RE.finditer(answer)})
+    valid = range(1, result_count + 1)
+    cited = tuple(marker for marker in markers if marker in valid)
+    invalid = tuple(marker for marker in markers if marker not in valid)
+    return CitationReport(cited=cited, invalid=invalid, grounded=bool(cited))

python/ebook_search/ingest.py

@@ -0,0 +1,222 @@
+"""EPUB ingestion into Richie DB."""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import logging
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import tiktoken
+from sqlalchemy import or_, select
+
+from python.ebook_search.epub_parse import parse_epub
+from python.ebook_search.protected_phrases.matching import index_chunk_phrase_mentions_for_book
+from python.orm.richie import EbookChapter, EbookChunk, EbookSource
+
+logger = logging.getLogger(__name__)
+DEFAULT_CHUNK_TOKENS = 700
+DEFAULT_CHUNK_OVERLAP = 100
+
+if TYPE_CHECKING:
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.epub_parse import ParsedChapter
+
+
+@dataclass(frozen=True)
+class TextChunk:
+    """A token-bounded chunk of text."""
+
+    text: str
+    token_start: int
+    token_count: int
+
+
+def chunk_text(
+    text: str,
+    *,
+    chunk_tokens: int = DEFAULT_CHUNK_TOKENS,
+    overlap_tokens: int = DEFAULT_CHUNK_OVERLAP,
+) -> list[TextChunk]:
+    """Split text into overlapping token chunks."""
+    if chunk_tokens <= 0:
+        msg = "chunk_tokens must be positive"
+        raise ValueError(msg)
+    if overlap_tokens < 0 or overlap_tokens >= chunk_tokens:
+        msg = "overlap_tokens must be non-negative and smaller than chunk_tokens"
+        raise ValueError(msg)
+
+    encoding = tiktoken.get_encoding("cl100k_base")
+    tokens = encoding.encode(text)
+    if not tokens:
+        return []
+
+    chunks: list[TextChunk] = []
+    step = chunk_tokens - overlap_tokens
+    for start in range(0, len(tokens), step):
+        chunk = tokens[start : start + chunk_tokens]
+        if not chunk:
+            continue
+        chunks.append(
+            TextChunk(
+                text=encoding.decode(chunk).strip(),
+                token_start=start,
+                token_count=len(chunk),
+            )
+        )
+        if start + chunk_tokens >= len(tokens):
+            break
+    return [chunk for chunk in chunks if chunk.text]
+
+
+def find_library_epubs(library_path: str) -> tuple[Path, list[Path] | None]:
+    """Resolve one configured library path and collect its EPUB files (blocking filesystem walk).
+
+    Returns:
+        tuple[Path, list[Path] | None]: The expanded path and its EPUB files, or ``None`` when
+        the path is neither an EPUB file nor a directory.
+    """
+    path = Path(library_path).expanduser()
+    if path.is_file() and path.suffix.lower() == ".epub":
+        return path, [path]
+    if path.is_dir():
+        return path, sorted(path.rglob("*.epub"))
+    return path, None
+
+
+async def ingest_configured_paths(session: AsyncSession, config: EbookSearchConfig) -> int:
+    """Ingest every EPUB found under configured library paths."""
+    count = 0
+    for library_path in config.library_paths:
+        path, epub_paths = await asyncio.to_thread(find_library_epubs, library_path)
+        logger.info("ebook_ingest_path_start path=%s", path)
+        if epub_paths is None:
+            logger.warning("ebook_ingest_path_missing path=%s", path)
+            continue
+        for epub_path in epub_paths:
+            count += int(await ingest_file(session, epub_path, config))
+    logger.info("ebook_ingest_paths_complete changed_files=%s configured_paths=%s", count, len(config.library_paths))
+    return count
+
+
+def resolve_ingest_path(path: Path) -> Path:
+    """Expand and resolve an ingest path (blocking filesystem call)."""
+    return path.expanduser().resolve()
+
+
+async def ingest_file(session: AsyncSession, path: Path, config: EbookSearchConfig) -> bool:
+    """Ingest one EPUB file. Return True when the database changed."""
+    try:
+        resolved_path = await asyncio.to_thread(resolve_ingest_path, path)
+        logger.info("ebook_ingest_file_start path=%s", resolved_path)
+        file_hash = await asyncio.to_thread(sha256_file, resolved_path)
+        existing = await find_existing_source(session, resolved_path, file_hash)
+        if existing is not None and existing.file_sha256 == file_hash:
+            stat = resolved_path.stat()
+            existing.file_path = str(resolved_path)
+            existing.file_mtime = datetime.fromtimestamp(stat.st_mtime, tz=UTC)
+            existing.file_size = stat.st_size
+            await session.flush()
+            logger.info("ebook_ingest_file_unchanged source_id=%s path=%s", existing.id, resolved_path)
+            return False
+        if existing is not None:
+            logger.info("ebook_ingest_file_replacing source_id=%s path=%s", existing.id, resolved_path)
+            await session.delete(existing)
+            await session.flush()
+
+        stat = resolved_path.stat()
+        parsed = await asyncio.to_thread(parse_epub, resolved_path)
+        source = EbookSource(
+            title=parsed.title,
+            author=parsed.author,
+            language=parsed.language,
+            publisher=parsed.publisher,
+            identifier=parsed.identifier,
+            file_path=str(resolved_path),
+            file_sha256=file_hash,
+            file_mtime=datetime.fromtimestamp(stat.st_mtime, tz=UTC),
+            file_size=stat.st_size,
+        )
+        session.add(source)
+        await session.flush()
+
+        chunk_index = 0
+        for spine_index, parsed_chapter in enumerate(parsed.chapters):
+            chapter = EbookChapter(
+                source_id=source.id,
+                spine_index=spine_index,
+                title=parsed_chapter.title,
+                href=parsed_chapter.href,
+            )
+            session.add(chapter)
+            await session.flush()
+            chunk_index = add_chapter_chunks(session, source, chapter, parsed_chapter, chunk_index, config)
+
+        await session.commit()
+        mention_count = await index_chunk_phrase_mentions_for_book(session, source.id, config)
+        logger.info(
+            "ebook_ingest_file_complete source_id=%s path=%s chapters=%s chunks=%s phrase_mentions=%s",
+            source.id,
+            resolved_path,
+            len(parsed.chapters),
+            chunk_index,
+            mention_count,
+        )
+    except Exception:
+        logger.exception(f"ebook_ingest_file_error path={path}")
+        return False
+    else:
+        return True
+
+
+async def find_existing_source(session: AsyncSession, path: Path, file_hash: str) -> EbookSource | None:
+    """Find an existing source by canonical path or file hash."""
+    return await session.scalar(
+        select(EbookSource).where(or_(EbookSource.file_path == str(path), EbookSource.file_sha256 == file_hash))
+    )
+
+
+def add_chapter_chunks(
+    session: AsyncSession,
+    source: EbookSource,
+    chapter: EbookChapter,
+    parsed_chapter: ParsedChapter,
+    chunk_index: int,
+    config: EbookSearchConfig,
+) -> int:
+    """Add chunk rows for one parsed chapter and return the next chunk index."""
+    page_label = parsed_chapter.page_labels[0] if parsed_chapter.page_labels else None
+    for text_chunk in chunk_text(
+        parsed_chapter.text,
+        chunk_tokens=config.chunk_tokens,
+        overlap_tokens=config.chunk_overlap,
+    ):
+        session.add(
+            EbookChunk(
+                source_id=source.id,
+                chapter_id=chapter.id,
+                chunk_index=chunk_index,
+                text=text_chunk.text,
+                token_start=text_chunk.token_start,
+                token_count=text_chunk.token_count,
+                page_label=page_label,
+                content_sha256=hashlib.sha256(text_chunk.text.encode()).hexdigest(),
+                search_text=f"{source.title} {source.author or ''} {chapter.title or ''} {text_chunk.text}",
+            )
+        )
+        chunk_index += 1
+    return chunk_index
+
+
+def sha256_file(path: Path) -> str:
+    """Calculate the SHA-256 digest for a file."""
+    digest = hashlib.sha256()
+    with path.open("rb") as file:
+        for block in iter(lambda: file.read(1024 * 1024), b""):
+            digest.update(block)
+    return digest.hexdigest()

python/ebook_search/llm_interface.py

@@ -0,0 +1,223 @@
+"""LLM provider HTTP adapters."""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import httpx
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from python.ebook_search.config import EbookSearchConfig, RerankConfig
+
+logger = logging.getLogger(__name__)
+
+
+def auth_headers(api_key: str) -> dict[str, str]:
+    """Build authorization headers when an API key is configured."""
+    if api_key == "not-needed":
+        return {}
+    return {"Authorization": f"Bearer {api_key}"}
+
+
+async def request_embeddings(
+    client: httpx.AsyncClient,
+    texts: Sequence[str],
+    config: EbookSearchConfig,
+) -> list[list[float]]:
+    """Request embeddings from the configured OpenAI-compatible endpoint.
+
+    Args:
+        client (httpx.AsyncClient): Shared async client for LLM calls.
+        texts (Sequence[str]): Texts to embed.
+        config (EbookSearchConfig): Runtime settings supplying the endpoint, model, and auth.
+
+    Returns:
+        list[list[float]]: One embedding vector per input text.
+
+    Raises:
+        RuntimeError: If the request fails or the response cannot be parsed.
+    """
+    try:
+        response = await client.post(
+            f"{config.embedding_base_url.rstrip('/')}/embeddings",
+            headers=auth_headers(config.embedding_api_key),
+            json={"model": config.embedding_model, "input": list(texts)},
+            timeout=config.embedding_timeout_seconds,
+        )
+        response.raise_for_status()
+        return embedding_vectors_from_response(response.json())
+    except (httpx.HTTPError, ValueError, KeyError, TypeError) as error:
+        logger.exception(
+            "ebook_embed_request_failed base_url=%s model=%s count=%s",
+            config.embedding_base_url,
+            config.embedding_model,
+            len(texts),
+        )
+        msg = f"Embedding request failed. base_url={config.embedding_base_url} model={config.embedding_model}"
+        raise RuntimeError(msg) from error
+
+
+async def check_embedding_endpoint(
+    client: httpx.AsyncClient,
+    config: EbookSearchConfig,
+    *,
+    timeout_seconds: float = 5.0,
+) -> bool:
+    """Return whether the configured embedding endpoint answers a model listing."""
+    try:
+        response = await client.get(
+            f"{config.embedding_base_url.rstrip('/')}/models",
+            headers=auth_headers(config.embedding_api_key),
+            timeout=timeout_seconds,
+        )
+        response.raise_for_status()
+    except httpx.HTTPError as error:
+        logger.warning("ebook_embedding_endpoint_unreachable base_url=%s error=%s", config.embedding_base_url, error)
+        return False
+    return True
+
+
+async def check_chat_endpoint(
+    client: httpx.AsyncClient,
+    config: EbookSearchConfig,
+    *,
+    timeout_seconds: float = 5.0,
+) -> bool:
+    """Return whether the configured chat (answering) endpoint answers a model listing."""
+    try:
+        response = await client.get(
+            f"{config.vllm_base_url.rstrip('/')}/models",
+            headers=auth_headers(config.vllm_api_key),
+            timeout=timeout_seconds,
+        )
+        response.raise_for_status()
+    except httpx.HTTPError as error:
+        logger.warning("ebook_chat_endpoint_unreachable base_url=%s error=%s", config.vllm_base_url, error)
+        return False
+    return True
+
+
+def embedding_vectors_from_response(body: object) -> list[list[float]]:
+    """Extract embedding vectors from an OpenAI-compatible embedding response."""
+    if not isinstance(body, dict):
+        msg = "Embedding response is not an object"
+        raise TypeError(msg)
+
+    data = body["data"]
+    if not isinstance(data, list):
+        msg = "Embedding response data is not a list"
+        raise TypeError(msg)
+
+    vectors: list[list[float]] = []
+    for item in data:
+        if not isinstance(item, dict):
+            msg = "Embedding item is not an object"
+            raise TypeError(msg)
+        embedding = item["embedding"]
+        if not isinstance(embedding, list):
+            msg = "Embedding value is not a list"
+            raise TypeError(msg)
+        vectors.append([float(value) for value in embedding])
+    return vectors
+
+
+async def request_rerank(
+    client: httpx.AsyncClient,
+    query: str,
+    documents: Sequence[str],
+    config: RerankConfig,
+) -> object | None:
+    """Request rerank scores from the configured vLLM endpoint.
+
+    Args:
+        client (httpx.AsyncClient): Shared async client for LLM calls.
+        query (str): Query the documents are scored against.
+        documents (Sequence[str]): Candidate documents to score.
+        config (RerankConfig): Rerank endpoint settings.
+
+    Returns:
+        object | None: The decoded response body, or ``None`` when it is not valid JSON.
+    """
+    payload = {
+        "model": config.model,
+        "query": query,
+        "documents": list(documents),
+    }
+    response = await client.post(
+        f"{config.base_url.rstrip('/')}/rerank",
+        json=payload,
+        timeout=config.timeout_seconds,
+    )
+    response.raise_for_status()
+    try:
+        return response.json()
+    except ValueError:
+        logger.debug("ebook_rerank_response_invalid_json", extra={"response": response.text})
+        return None
+
+
+async def request_chat_completion(
+    client: httpx.AsyncClient,
+    config: EbookSearchConfig,
+    messages: Sequence[dict[str, str]],
+) -> str:
+    """Request a chat completion over a shared async client.
+
+    Args:
+        client (httpx.AsyncClient): Shared async client whose connection pool bounds concurrency.
+        config (EbookSearchConfig): Runtime settings supplying the endpoint, model, and auth.
+        messages (Sequence[dict[str, str]]): OpenAI-style chat messages.
+
+    Returns:
+        str: The assistant message text.
+
+    Raises:
+        RuntimeError: If the request fails or the response cannot be parsed.
+    """
+    try:
+        response = await client.post(
+            f"{config.vllm_base_url.rstrip('/')}/chat/completions",
+            headers=auth_headers(config.vllm_api_key),
+            json={
+                "model": config.chat_model,
+                "messages": list(messages),
+                "temperature": 0,
+            },
+            timeout=config.chat_timeout_seconds,
+        )
+        response.raise_for_status()
+        return chat_content_from_response(response.json())
+    except (httpx.HTTPError, ValueError, KeyError, TypeError) as error:
+        msg = f"Chat request failed. base_url={config.vllm_base_url} model={config.chat_model}"
+        raise RuntimeError(msg) from error
+
+
+def chat_content_from_response(body: object) -> str:
+    """Extract text content from an OpenAI-compatible chat response."""
+    if not isinstance(body, dict):
+        msg = "Chat response is not an object"
+        raise TypeError(msg)
+
+    choices = body["choices"]
+    if not isinstance(choices, list) or not choices:
+        msg = "Chat response has no choices"
+        raise ValueError(msg)
+
+    first = choices[0]
+    if not isinstance(first, dict):
+        msg = "Chat choice is not an object"
+        raise TypeError(msg)
+
+    message = first["message"]
+    if not isinstance(message, dict):
+        msg = "Chat message is not an object"
+        raise TypeError(msg)
+
+    content = message.get("content") or ""
+    if not isinstance(content, str):
+        msg = "Chat content is not text"
+        raise TypeError(msg)
+    return content

python/ebook_search/loadtest.py

@@ -0,0 +1,218 @@
+"""Load test for the EPUB search service.
+
+Drives ``POST /search`` on a running server at a configurable concurrency and reports
+latency percentiles, throughput, and HTTP status distribution. Queries are drawn from
+the shared JSONL set (see ``eval/data/queries.jsonl``) that the eval also uses, so load
+and evaluation exercise the same questions. Answer generation and reranking happen
+server-side, so this exercises the full retrieval pipeline.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import math
+import random
+import statistics
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Annotated
+
+import httpx
+import typer
+
+from python.common import configure_logger
+from python.ebook_search.eval.dataset import DEFAULT_QUERIES_PATH, load_gold_queries
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class RequestResult:
+    """Outcome of a single search request."""
+
+    status_code: int
+    latency_ms: float
+    ok: bool
+
+
+@dataclass(frozen=True)
+class LoadSummary:
+    """Aggregate results of a load test run."""
+
+    total: int
+    successes: int
+    failures: int
+    wall_seconds: float
+    throughput_rps: float
+    latency_p50_ms: float
+    latency_p90_ms: float
+    latency_p95_ms: float
+    latency_p99_ms: float
+    latency_mean_ms: float
+    latency_max_ms: float
+    status_counts: dict[int, int]
+
+
+def load_queries(queries_file: str | None) -> list[str]:
+    """Return the query strings from the shared JSONL set (or a custom JSONL file)."""
+    path = Path(queries_file) if queries_file else DEFAULT_QUERIES_PATH
+    queries = [gold.query for gold in load_gold_queries(path)]
+    if not queries:
+        msg = f"No queries found in {path}"
+        raise typer.BadParameter(msg)
+    return queries
+
+
+def pick_query(queries: list[str]) -> str:
+    """Return a uniformly random query from the pool (not a security context)."""
+    return random.choice(queries)  # noqa: S311 load-test query sampling is not security-sensitive
+
+
+def percentile(values_sorted: list[float], pct: float) -> float:
+    """Return the linearly-interpolated percentile of a sorted list."""
+    if not values_sorted:
+        return 0.0
+    rank = (pct / 100) * (len(values_sorted) - 1)
+    low = math.floor(rank)
+    high = math.ceil(rank)
+    if low == high:
+        return values_sorted[low]
+    return values_sorted[low] + (values_sorted[high] - values_sorted[low]) * (rank - low)
+
+
+def summarize(results: list[RequestResult], wall_seconds: float) -> LoadSummary:
+    """Aggregate per-request results into a load summary."""
+    latencies = sorted(result.latency_ms for result in results)
+    successes = sum(1 for result in results if result.ok)
+    status_counts: dict[int, int] = {}
+    for result in results:
+        status_counts[result.status_code] = status_counts.get(result.status_code, 0) + 1
+    return LoadSummary(
+        total=len(results),
+        successes=successes,
+        failures=len(results) - successes,
+        wall_seconds=wall_seconds,
+        throughput_rps=len(results) / wall_seconds if wall_seconds > 0 else 0.0,
+        latency_p50_ms=percentile(latencies, 50),
+        latency_p90_ms=percentile(latencies, 90),
+        latency_p95_ms=percentile(latencies, 95),
+        latency_p99_ms=percentile(latencies, 99),
+        latency_mean_ms=statistics.fmean(latencies) if latencies else 0.0,
+        latency_max_ms=latencies[-1] if latencies else 0.0,
+        status_counts=status_counts,
+    )
+
+
+async def send_search(client: httpx.AsyncClient, query: str, *, rerank: bool) -> RequestResult:
+    """Send one search request and record its status and latency."""
+    data = {"query": query, "rerank": "true"} if rerank else {"query": query}
+    start = time.perf_counter()
+    try:
+        response = await client.post("/search", data=data)
+    except httpx.HTTPError as error:
+        logger.warning("ebook_loadtest_request_failed error=%s", error)
+        return RequestResult(status_code=0, latency_ms=(time.perf_counter() - start) * 1000, ok=False)
+    return RequestResult(
+        status_code=response.status_code,
+        latency_ms=(time.perf_counter() - start) * 1000,
+        ok=response.is_success,
+    )
+
+
+async def worker(
+    client: httpx.AsyncClient,
+    queue: asyncio.Queue[str],
+    results: list[RequestResult],
+    *,
+    rerank: bool,
+) -> None:
+    """Pull queries off the queue and send requests until it is empty."""
+    while True:
+        try:
+            query = queue.get_nowait()
+        except asyncio.QueueEmpty:
+            return
+        results.append(await send_search(client, query, rerank=rerank))
+
+
+async def run_load(
+    *,
+    base_url: str,
+    queries: list[str],
+    request_count: int,
+    concurrency: int,
+    rerank: bool,
+    warmup: int,
+    timeout_seconds: float,
+) -> LoadSummary:
+    """Run the load test and return its aggregate summary."""
+    limits = httpx.Limits(max_connections=concurrency, max_keepalive_connections=concurrency)
+    async with httpx.AsyncClient(base_url=base_url, timeout=timeout_seconds, limits=limits) as client:
+        for _ in range(warmup):
+            await send_search(client, pick_query(queries), rerank=rerank)
+
+        queue: asyncio.Queue[str] = asyncio.Queue()
+        for _ in range(request_count):
+            queue.put_nowait(pick_query(queries))
+
+        results: list[RequestResult] = []
+        start = time.perf_counter()
+        workers = [asyncio.create_task(worker(client, queue, results, rerank=rerank)) for _ in range(concurrency)]
+        await asyncio.gather(*workers)
+        wall_seconds = time.perf_counter() - start
+    return summarize(results, wall_seconds)
+
+
+def print_summary(summary: LoadSummary) -> None:
+    """Print the load summary to stdout."""
+    typer.echo(f"requests={summary.total} successes={summary.successes} failures={summary.failures}")
+    typer.echo(f"wall={summary.wall_seconds:.2f}s throughput={summary.throughput_rps:.1f} req/s")
+    typer.echo(
+        f"latency_ms p50={summary.latency_p50_ms:.1f} p90={summary.latency_p90_ms:.1f} "
+        f"p95={summary.latency_p95_ms:.1f} p99={summary.latency_p99_ms:.1f} "
+        f"mean={summary.latency_mean_ms:.1f} max={summary.latency_max_ms:.1f}"
+    )
+    status_summary = " ".join(f"{code}={count}" for code, count in sorted(summary.status_counts.items()))
+    typer.echo(f"status {status_summary}")
+
+
+def main(
+    *,
+    base_url: Annotated[str, typer.Option(help="Base URL of the running service")] = "http://127.0.0.1:8070",
+    request_count: Annotated[int, typer.Option("--requests", help="Total requests to send")] = 200,
+    concurrency: Annotated[int, typer.Option(help="Concurrent in-flight requests")] = 10,
+    rerank: Annotated[bool, typer.Option(help="Request server-side reranking")] = False,
+    warmup: Annotated[int, typer.Option(help="Warmup requests, not measured")] = 5,
+    timeout_seconds: Annotated[float, typer.Option("--timeout", help="Per-request timeout seconds")] = 120.0,
+    queries_file: Annotated[str | None, typer.Option(help="Query JSONL file (defaults to the shared set)")] = None,
+    log_level: Annotated[str, typer.Option(help="Log level")] = "WARNING",
+) -> None:
+    """Load test the search endpoint and report latency and throughput."""
+    configure_logger(log_level)
+    queries = load_queries(queries_file)
+    logger.info(
+        "ebook_loadtest_start base_url=%s requests=%s concurrency=%s rerank=%s queries=%s",
+        base_url,
+        request_count,
+        concurrency,
+        rerank,
+        len(queries),
+    )
+    summary = asyncio.run(
+        run_load(
+            base_url=base_url,
+            queries=queries,
+            request_count=request_count,
+            concurrency=concurrency,
+            rerank=rerank,
+            warmup=warmup,
+            timeout_seconds=timeout_seconds,
+        )
+    )
+    print_summary(summary)
+
+
+if __name__ == "__main__":
+    typer.run(main)

python/ebook_search/protected_phrases/__init__.py

@@ -0,0 +1 @@
+"""Init."""

python/ebook_search/protected_phrases/config/__init__.py

@@ -0,0 +1,17 @@
+"""Protected phrase extraction, storage, and runtime matching."""
+
+from python.ebook_search.protected_phrases.config.lib import (
+    get_bad_ends,
+    get_bad_starts,
+    get_ignored_phrases,
+    get_junk_tokens,
+    get_most_common_words,
+)
+
+__all__ = [
+    "get_bad_ends",
+    "get_bad_starts",
+    "get_ignored_phrases",
+    "get_junk_tokens",
+    "get_most_common_words",
+]

python/ebook_search/protected_phrases/config/bad_ends.toml

@@ -0,0 +1,31 @@
+tokens = [
+  "a",
+  "an",
+  "and",
+  "any",
+  "as",
+  "at",
+  "be",
+  "because",
+  "but",
+  "by",
+  "can",
+  "could",
+  "do",
+  "for",
+  "from",
+  "have",
+  "if",
+  "of",
+  "or",
+  "some",
+  "than",
+  "the",
+  "these",
+  "this",
+  "to",
+  "will",
+  "with",
+  "would",
+  "did",
+]

python/ebook_search/protected_phrases/config/bad_starts.toml

@@ -0,0 +1,27 @@
+tokens = [
+    "a",
+    "an",
+    "did",
+    "didn't",
+    "he",
+    "here",
+    "how",
+    "i",
+    "it",
+    "she",
+    "that",
+    "the",
+    "there",
+    "they",
+    "this",
+    "we",
+    "what",
+    "when",
+    "where",
+    "which",
+    "who",
+    "whom",
+    "whose",
+    "why",
+    "you",
+]

python/ebook_search/protected_phrases/config/ignored_phrases.toml

@@ -0,0 +1,212 @@
+phrases = [
+  "a little",
+  "across the",
+  "and she",
+  "anyone in",
+  "are you",
+  "around him",
+  "around the",
+  "as much",
+  "as soon",
+  "at all",
+  "at least",
+  "before the",
+  "behind him",
+  "between the",
+  "but she",
+  "could not",
+  "did he",
+  "did i",
+  "did it",
+  "did not believe",
+  "did not care",
+  "did not even",
+  "did not know what",
+  "did not know",
+  "did not like",
+  "did not look",
+  "did not mean",
+  "did not move",
+  "did not need",
+  "did not see",
+  "did not seem",
+  "did not think",
+  "did not understand",
+  "did not want",
+  "did not",
+  "did she",
+  "did so",
+  "did that",
+  "did the",
+  "did they",
+  "did what",
+  "did you",
+  "didn't answer",
+  "didn't care",
+  "didn't even",
+  "didn't expect",
+  "didn't feel",
+  "didn't get",
+  "didn't i",
+  "didn't know",
+  "didn't like",
+  "didn't look",
+  "didn't make",
+  "didn't mean",
+  "didn't need",
+  "didn't really",
+  "didn't say",
+  "didn't see",
+  "didn't seem",
+  "didn't think",
+  "didn't want",
+  "didn't you",
+  "end up",
+  "ended up",
+  "had a",
+  "had been",
+  "have been",
+  "he asked",
+  "he concluded",
+  "he continued",
+  "he couldn't",
+  "he did",
+  "he didn't",
+  "he felt",
+  "he had",
+  "he hadn't",
+  "he knew",
+  "he noted",
+  "he pointed",
+  "he realized",
+  "he replied",
+  "he said",
+  "he saw",
+  "he tapped",
+  "he told",
+  "he was",
+  "he wasn't",
+  "his body",
+  "his chair",
+  "his feet",
+  "his hands",
+  "his head",
+  "his office",
+  "his own",
+  "his pc",
+  "his power",
+  "his shield",
+  "his sight",
+  "his voice",
+  "his wrist",
+  "how many",
+  "i am",
+  "i don't",
+  "i said",
+  "i was",
+  "i wouldn't",
+  "i'm not",
+  "if he",
+  "if they",
+  "is in",
+  "is not",
+  "is that",
+  "is the",
+  "it had",
+  "it had",
+  "it is",
+  "it was",
+  "it wasn't",
+  "it wasn't",
+  "no one",
+  "of course",
+  "of force",
+  "of it",
+  "of magic",
+  "of marines",
+  "of power",
+  "of those",
+  "old man",
+  "older man",
+  "one of",
+  "out of",
+  "set up",
+  "she admitted",
+  "she asked",
+  "she had",
+  "she replied",
+  "she said",
+  "she snapped",
+  "she told",
+  "she was",
+  "she'd been",
+  "shook his",
+  "sure he",
+  "tell you",
+  "that had",
+  "that is",
+  "that she",
+  "that was",
+  "the dark",
+  "the door",
+  "the first",
+  "the last",
+  "the man",
+  "the one",
+  "the only",
+  "the other",
+  "the rest",
+  "the room",
+  "the same",
+  "the two",
+  "the way",
+  "the world",
+  "there are",
+  "there was",
+  "there were",
+  "they are",
+  "they had",
+  "they were",
+  "they weren't",
+  "this is",
+  "this place",
+  "though he",
+  "through his",
+  "through the",
+  "to find",
+  "to get",
+  "to keep",
+  "to stay",
+  "to stop",
+  "to tell",
+  "to try",
+  "told her",
+  "told him",
+  "under his",
+  "was a",
+  "was enough",
+  "was going",
+  "was in",
+  "was no",
+  "was not",
+  "was now",
+  "was on",
+  "was one",
+  "was only",
+  "was still",
+  "was that",
+  "was the",
+  "was there",
+  "were in",
+  "what had",
+  "what happened",
+  "what was",
+  "where the",
+  "while i",
+  "you are",
+  "you can't",
+  "you don't",
+  "you know",
+  "you need",
+  "you were",
+]

python/ebook_search/protected_phrases/config/junk_tokens.toml

@@ -0,0 +1,71 @@
+tokens = [
+    "said",
+    "asked",
+    "replied",
+    "answered",
+    "looked",
+    "nodded",
+    "turned",
+    "shook",
+    "smiled",
+    "shrugged",
+    "pointed",
+    "continued",
+    "repeated",
+    "stared",
+    "agreed",
+    "glanced",
+    "walked",
+    "told",
+    "thought",
+    "knew",
+    "wanted",
+    "muttered",
+    "whispered",
+    "laughed",
+    "sighed",
+    "paused",
+    "gestured",
+    "waved",
+    "frowned",
+    "grinned",
+    "admitted",
+    "found",
+    "noted",
+    "murmured",
+    "ordered",
+    "i'm",
+    "i've",
+    "i'd",
+    "i'll",
+    "it's",
+    "that's",
+    "don't",
+    "didn't",
+    "doesn't",
+    "can't",
+    "won't",
+    "wouldn't",
+    "couldn't",
+    "shouldn't",
+    "isn't",
+    "wasn't",
+    "aren't",
+    "weren't",
+    "you're",
+    "you've",
+    "you'll",
+    "we're",
+    "we've",
+    "we'll",
+    "they're",
+    "they've",
+    "he's",
+    "she's",
+    "there's",
+    "what's",
+    "let's",
+    "who's",
+    "he'd",
+    "she'd",
+]

python/ebook_search/protected_phrases/config/lib.py

@@ -0,0 +1,60 @@
+"""Protected phrase extraction, storage, and runtime matching."""
+
+from __future__ import annotations
+
+import logging
+import tomllib
+from functools import cache
+from pathlib import Path
+
+from python.ebook_search.protected_phrases.text_normalization import normalize_text
+
+logger = logging.getLogger(__name__)
+
+
+def _load_toml_string_set(path: Path, key: str) -> frozenset[str]:
+    """Load and validate a TOML string list as a normalized immutable set."""
+    with path.open("rb") as file:
+        body = tomllib.load(file)
+
+    values = body.get(key)
+    if not isinstance(values, list) or not all(isinstance(item, str) for item in values):
+        msg = f"{path} must contain a {key!r} string list"
+        raise ValueError(msg)
+    return frozenset(normalize_text(value) for value in values if normalize_text(value))
+
+
+@cache
+def _get_phrase_config_dir() -> Path:
+    """Return the directory containing phrase configuration files."""
+    return Path(__file__).resolve().parent
+
+
+@cache
+def get_ignored_phrases() -> frozenset[str]:
+    """Return ignored phrase strings loaded from TOML."""
+    return _load_toml_string_set(_get_phrase_config_dir() / "ignored_phrases.toml", "phrases")
+
+
+@cache
+def get_bad_ends() -> frozenset[str]:
+    """Return bad phrase-ending tokens loaded from TOML."""
+    return _load_toml_string_set(_get_phrase_config_dir() / "bad_ends.toml", "tokens")
+
+
+@cache
+def get_bad_starts() -> frozenset[str]:
+    """Return bad phrase-starting tokens loaded from TOML."""
+    return _load_toml_string_set(_get_phrase_config_dir() / "bad_starts.toml", "tokens")
+
+
+@cache
+def get_most_common_words() -> frozenset[str]:
+    """Return the most common English words loaded from TOML."""
+    return _load_toml_string_set(_get_phrase_config_dir() / "most_common_words.toml", "words")
+
+
+@cache
+def get_junk_tokens() -> frozenset[str]:
+    """Return junk tokens (dialogue verbs and pronoun contractions) loaded from TOML."""
+    return _load_toml_string_set(_get_phrase_config_dir() / "junk_tokens.toml", "tokens")

python/ebook_search/protected_phrases/config/most_common_words.toml

@@ -0,0 +1,102 @@
+words = [
+    "the",
+    "be",
+    "to",
+    "of",
+    "and",
+    "a",
+    "in",
+    "that",
+    "have",
+    "I",
+    "it",
+    "for",
+    "not",
+    "on",
+    "with",
+    "he",
+    "as",
+    "you",
+    "do",
+    "at",
+    "this",
+    "but",
+    "his",
+    "by",
+    "from",
+    "they",
+    "we",
+    "say",
+    "her",
+    "she",
+    "or",
+    "an",
+    "will",
+    "my",
+    "one",
+    "all",
+    "would",
+    "there",
+    "their",
+    "what",
+    "so",
+    "up",
+    "out",
+    "if",
+    "about",
+    "who",
+    "get",
+    "which",
+    "go",
+    "me",
+    "when",
+    "make",
+    "can",
+    "like",
+    "time",
+    "no",
+    "just",
+    "him",
+    "know",
+    "take",
+    "people",
+    "into",
+    "year",
+    "your",
+    "good",
+    "some",
+    "could",
+    "them",
+    "see",
+    "other",
+    "than",
+    "then",
+    "now",
+    "look",
+    "only",
+    "come",
+    "its",
+    "over",
+    "think",
+    "also",
+    "back",
+    "after",
+    "use",
+    "two",
+    "how",
+    "our",
+    "work",
+    "first",
+    "well",
+    "way",
+    "even",
+    "new",
+    "want",
+    "because",
+    "any",
+    "these",
+    "give",
+    "day",
+    "most",
+    "us",
+]

python/ebook_search/protected_phrases/extraction.py

@@ -0,0 +1,853 @@
+"""Candidate phrase extraction and scoring for protected phrases."""
+
+from __future__ import annotations
+
+import logging
+import re
+from collections import Counter, defaultdict
+from functools import lru_cache
+from time import perf_counter
+from typing import TYPE_CHECKING, Protocol
+
+from yake import KeywordExtractor
+
+from python.ebook_search.protected_phrases.config import (
+    get_bad_ends,
+    get_bad_starts,
+    get_ignored_phrases,
+    get_junk_tokens,
+    get_most_common_words,
+)
+from python.ebook_search.protected_phrases.models import PhraseCandidate
+from python.ebook_search.protected_phrases.text_normalization import tokenize, tokenize_with_offsets
+
+if TYPE_CHECKING:
+    from collections.abc import Iterable, Mapping, Sequence
+
+    from python.ebook_search.config import EbookSearchConfig
+
+logger = logging.getLogger(__name__)
+
+
+BAD_START_SCORE_PENALTY = 10.0
+BAD_END_SCORE_PENALTY = 10.0
+MULTI_SOURCE_SCORE_BONUS = 2.0
+MULTI_SOURCE_MIN_SOURCES = 2
+CAPITALIZED_PHRASE_RE = re.compile(r"\b(?:[A-Z][a-zA-Z']+)(?:\s+(?:of|the|and|in|on|for|[A-Z][a-zA-Z']+)){0,6}")
+
+
+class SpacySpan(Protocol):
+    """Small protocol for the spaCy span attributes used by this module."""
+
+    text: str
+
+
+class SpacyEntity(SpacySpan, Protocol):
+    """Small protocol for the spaCy entity attributes used by this module."""
+
+    label_: str
+
+
+class SpacyDoc(Protocol):
+    """Small protocol for the spaCy doc attributes used by this module."""
+
+    ents: Iterable[SpacyEntity]
+    noun_chunks: Iterable[SpacySpan]
+
+
+class SpacyLanguage(Protocol):
+    """Small protocol for a callable spaCy language pipeline."""
+
+    def __call__(self, text: str) -> SpacyDoc:
+        """Parse text into a spaCy-like doc."""
+
+
+class YakeExtractor(Protocol):
+    """Small protocol for the YAKE extractor used by this module."""
+
+    def extract_keywords(self, text: str) -> Iterable[tuple[str, float]]:
+        """Return YAKE keyword tuples."""
+
+
+class YakeExtractorFactory(Protocol):
+    """Callable constructor protocol for YAKE keyword extractors."""
+
+    def __call__(self, *, lan: str, n: int, dedupLim: float, top: int) -> YakeExtractor:  # noqa: N803
+        """Create a YAKE keyword extractor.
+
+        Args:
+            lan (str): Language code passed to YAKE.
+            n (int): Maximum n-gram size to extract.
+            dedupLim (float): Deduplication similarity threshold.
+            top (int): Maximum number of keyphrases to return.
+
+        Returns:
+            YakeExtractor: The constructed keyword extractor.
+        """
+
+
+def normalize_candidate_phrase(
+    phrase_text: str,
+    config: EbookSearchConfig,
+    *,
+    min_tokens: int | None = None,
+    max_tokens: int | None = None,
+    strip_leading_article: bool = False,
+) -> tuple[str, str, int] | None:
+    """Normalize a candidate phrase and validate token bounds.
+
+    Args:
+        phrase_text (str): Raw phrase text to normalize.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        min_tokens (int | None): Minimum token count override; defaults to ``config.phrase_min_tokens``.
+        max_tokens (int | None): Maximum token count override; defaults to ``config.phrase_max_tokens``.
+        strip_leading_article (bool): Whether to drop a single leading English article.
+
+    Returns:
+        tuple[str, str, int] | None: Display text, normalized phrase, and token count, or ``None``
+        when the phrase falls outside the token bounds or is ignored.
+    """
+    normalized_tokens = tokenize_with_offsets(phrase_text)
+    start = 0
+    if strip_leading_article and normalized_tokens and normalized_tokens[0].text in {"the", "a", "an"}:
+        start = 1
+
+    selected_tokens = normalized_tokens[start:]
+    min_count = config.phrase_min_tokens if min_tokens is None else min_tokens
+    max_count = config.phrase_max_tokens if max_tokens is None else max_tokens
+    if len(selected_tokens) < min_count or len(selected_tokens) > max_count:
+        return None
+
+    phrase_norm = " ".join(token.text for token in selected_tokens)
+    if phrase_norm in get_ignored_phrases():
+        return None
+
+    display_text = phrase_text[selected_tokens[0].start_char : selected_tokens[-1].end_char].strip()
+    return display_text or phrase_norm, phrase_norm, len(selected_tokens)
+
+
+def count_raw_ngrams(tokens: Sequence[str], config: EbookSearchConfig) -> Counter[str]:
+    """Count every n-gram window in one normalized token block.
+
+    ``tokens`` are already normalized (see :func:`tokenize`), so each window's normalized form
+    is the joined tokens directly. Counting into a plain :class:`Counter` rather than
+    :class:`PhraseCandidate` objects keeps this hot loop cheap; callers filter ignored phrases
+    and materialize candidates per unique phrase afterwards, which is far fewer operations than
+    doing either per window.
+
+    Args:
+        tokens (Sequence[str]): Normalized tokens for one text block.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        Counter[str]: Raw occurrence counts keyed by normalized phrase.
+    """
+    return Counter(
+        " ".join(tokens[start : start + ngram_size])
+        for ngram_size in range(config.phrase_min_tokens, config.phrase_max_tokens + 1)
+        for start in range(len(tokens) - ngram_size + 1)
+    )
+
+
+def extract_raw_ngrams_by_chapter(
+    chapters: Sequence[str],
+    config: EbookSearchConfig,
+) -> dict[str, PhraseCandidate]:
+    """Extract raw n-grams across chapters, tracking both raw counts and chapter spread.
+
+    Counting each chapter separately makes chapter spread fall out of dict membership: a phrase's
+    ``chapter_count`` is simply how many per-chapter count maps contain it, so no per-window seen
+    tracking is needed. This also lets the enrichment step skip re-sliding the same n-gram sizes.
+
+    Phrases below the minimum raw count are dropped here rather than materialized: most unique
+    n-grams occur once, and :func:`filter_storable_candidates` would discard them as too rare
+    anyway, so building ``PhraseCandidate`` objects for them is wasted work.
+
+    Args:
+        chapters (Sequence[str]): Chapter-like text blocks to slide n-gram windows over.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates meeting the minimum raw count, keyed by normalized
+        phrase, with raw and chapter counts.
+    """
+    chapter_count_maps = [count_raw_ngrams(tokenize(chapter), config) for chapter in chapters]
+    total_counts: Counter[str] = Counter()
+    chapter_spread: Counter[str] = Counter()
+    for chapter_counts in chapter_count_maps:
+        total_counts.update(chapter_counts)
+        chapter_spread.update(chapter_counts.keys())
+    min_raw_count = minimum_candidate_raw_count(config)
+    ignored = get_ignored_phrases()
+    return {
+        phrase_norm: PhraseCandidate(
+            phrase_text=phrase_norm,
+            phrase_norm=phrase_norm,
+            token_count=phrase_norm.count(" ") + 1,
+            source_raw_ngram=True,
+            raw_count=raw_count,
+            chapter_count=chapter_spread[phrase_norm],
+        )
+        for phrase_norm, raw_count in total_counts.items()
+        if raw_count >= min_raw_count and phrase_norm not in ignored
+    }
+
+
+@lru_cache(maxsize=2)
+def get_yake_extractor(max_ngram: int, top_k: int) -> KeywordExtractor:
+    """Return a cached YAKE extractor for the given settings.
+
+    Constructing a ``KeywordExtractor`` loads the language's stopword list from disk, so it is
+    cached and reused across books rather than rebuilt on every call.
+
+    Args:
+        max_ngram (int): Maximum n-gram size to extract.
+        top_k (int): Maximum number of keyphrases to request.
+
+    Returns:
+        KeywordExtractor: A shared extractor instance for the given settings.
+    """
+    return KeywordExtractor(lan="en", n=max_ngram, dedupLim=0.85, top=top_k)
+
+
+def extract_yake_candidates(
+    book_text: str,
+    config: EbookSearchConfig,
+    top_k: int = 1000,
+) -> dict[str, PhraseCandidate]:
+    """Extract YAKE keyphrases when the optional YAKE package is installed.
+
+    Args:
+        book_text (str): Full book text to extract keyphrases from.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        top_k (int): Maximum number of YAKE keyphrases to request.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates keyed by normalized phrase, with YAKE scores.
+    """
+    extractor = get_yake_extractor(config.phrase_max_tokens, top_k)
+    out: dict[str, PhraseCandidate] = {}
+    for phrase_text, yake_score in extractor.extract_keywords(book_text):
+        normalized = normalize_candidate_phrase(phrase_text, config)
+        if normalized is None:
+            continue
+        display_text, phrase_norm, token_count = normalized
+        out[phrase_norm] = PhraseCandidate(
+            phrase_text=display_text,
+            phrase_norm=phrase_norm,
+            token_count=token_count,
+            source_yake=True,
+            yake_score=float(yake_score),
+        )
+    return out
+
+
+def extract_spacy_candidates(
+    book_text: str,
+    nlp: SpacyLanguage,
+    config: EbookSearchConfig,
+) -> dict[str, PhraseCandidate]:
+    """Extract spaCy named entities and noun chunks from one text block.
+
+    Args:
+        book_text (str): Text block to parse with spaCy.
+        nlp (SpacyLanguage): Callable spaCy language pipeline.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates keyed by normalized phrase from entities and noun chunks.
+    """
+    out: dict[str, PhraseCandidate] = {}
+    doc = nlp(book_text)
+
+    for ent in doc.ents:
+        normalized = normalize_candidate_phrase(
+            ent.text,
+            config,
+            max_tokens=config.phrase_max_entity_tokens,
+        )
+        if normalized is None:
+            continue
+        phrase_text, phrase_norm, token_count = normalized
+        out[phrase_norm] = PhraseCandidate(
+            phrase_text=phrase_text,
+            phrase_norm=phrase_norm,
+            token_count=token_count,
+            source_spacy_ner=True,
+            spacy_label=ent.label_,
+        )
+
+    for chunk in doc.noun_chunks:
+        normalized = normalize_candidate_phrase(chunk.text, config, strip_leading_article=True)
+        if normalized is None:
+            continue
+        phrase_text, phrase_norm, token_count = normalized
+        out[phrase_norm] = PhraseCandidate(
+            phrase_text=phrase_text,
+            phrase_norm=phrase_norm,
+            token_count=token_count,
+            source_spacy_noun_chunk=True,
+        )
+    return out
+
+
+def extract_capitalized_phrases(original_text: str, config: EbookSearchConfig) -> dict[str, PhraseCandidate]:
+    """Extract capitalized phrase runs that often carry fictional terms.
+
+    Args:
+        original_text (str): Original-case book text to scan for capitalized runs.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates keyed by normalized phrase from capitalized runs.
+    """
+    out: dict[str, PhraseCandidate] = {}
+    for match in CAPITALIZED_PHRASE_RE.finditer(original_text):
+        phrase_text = match.group(0).strip()
+        normalized = normalize_candidate_phrase(
+            phrase_text,
+            config,
+            max_tokens=config.phrase_max_entity_tokens,
+        )
+        if normalized is None:
+            continue
+        display_text, phrase_norm, token_count = normalized
+        out[phrase_norm] = PhraseCandidate(
+            phrase_text=display_text,
+            phrase_norm=phrase_norm,
+            token_count=token_count,
+            source_capitalized=True,
+        )
+    return out
+
+
+def extract_metadata_candidates(
+    metadata: Mapping[str, object] | None,
+    config: EbookSearchConfig,
+) -> dict[str, PhraseCandidate]:
+    """Extract phrases from book metadata values such as title, author, and series.
+
+    Args:
+        metadata (Mapping[str, object] | None): Book metadata values, or ``None`` when unavailable.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates keyed by normalized phrase from metadata values.
+    """
+    if metadata is None:
+        return {}
+
+    out: dict[str, PhraseCandidate] = {}
+    for value in metadata.values():
+        if value is None:
+            continue
+        phrase_text = str(value).strip()
+        normalized = normalize_candidate_phrase(
+            phrase_text,
+            config,
+            max_tokens=config.phrase_max_entity_tokens,
+        )
+        if normalized is None:
+            continue
+        display_text, phrase_norm, token_count = normalized
+        out[phrase_norm] = PhraseCandidate(
+            phrase_text=display_text,
+            phrase_norm=phrase_norm,
+            token_count=token_count,
+            source_metadata=True,
+        )
+    return out
+
+
+def merge_candidate_sources(*sources: Mapping[str, PhraseCandidate]) -> dict[str, PhraseCandidate]:
+    """Merge candidate dictionaries by normalized phrase.
+
+    Args:
+        *sources (Mapping[str, PhraseCandidate]): Candidate maps to combine, keyed by normalized phrase.
+
+    Returns:
+        dict[str, PhraseCandidate]: One merged candidate per normalized phrase.
+    """
+    merged: dict[str, PhraseCandidate] = {}
+    for source in sources:
+        for phrase_norm, item in source.items():
+            existing = merged.setdefault(
+                phrase_norm,
+                PhraseCandidate(
+                    phrase_text=item.phrase_text,
+                    phrase_norm=phrase_norm,
+                    token_count=item.token_count,
+                ),
+            )
+            merge_candidate(existing, item)
+    return merged
+
+
+def merge_candidate(existing: PhraseCandidate, item: PhraseCandidate) -> None:
+    """Merge one candidate into an existing candidate object.
+
+    Args:
+        existing (PhraseCandidate): Candidate mutated in place to absorb ``item``.
+        item (PhraseCandidate): Candidate whose sources, counts, and scores are merged in.
+    """
+    existing.source_raw_ngram = existing.source_raw_ngram or item.source_raw_ngram
+    existing.source_yake = existing.source_yake or item.source_yake
+    existing.source_spacy_ner = existing.source_spacy_ner or item.source_spacy_ner
+    existing.source_spacy_noun_chunk = existing.source_spacy_noun_chunk or item.source_spacy_noun_chunk
+    existing.source_capitalized = existing.source_capitalized or item.source_capitalized
+    existing.source_metadata = existing.source_metadata or item.source_metadata
+    existing.raw_count += item.raw_count
+    existing.chapter_count = max(existing.chapter_count, item.chapter_count)
+    if item.yake_score is not None:
+        existing.yake_score = item.yake_score
+    if item.spacy_label:
+        existing.spacy_label = item.spacy_label
+
+
+def enrich_with_frequency_and_chapter_counts(
+    candidates: Mapping[str, PhraseCandidate],
+    chapters: Sequence[str],
+    *,
+    counted_sizes: Iterable[int] = (),
+) -> dict[str, PhraseCandidate]:
+    """Add raw occurrence and chapter-spread counts to candidates.
+
+    Candidates whose ``token_count`` is in ``counted_sizes`` are left untouched: those counts
+    were already computed while sliding the chapters in :func:`extract_raw_ngrams_by_chapter`,
+    so re-sliding those n-gram sizes here would just duplicate that work.
+
+    Args:
+        candidates (Mapping[str, PhraseCandidate]): Candidates to enrich, keyed by normalized phrase.
+        chapters (Sequence[str]): Chapter-like text blocks used to count occurrences and spread.
+        counted_sizes (Iterable[int]): Token counts whose counts are already populated and should be skipped.
+
+    Returns:
+        dict[str, PhraseCandidate]: Candidates with updated ``raw_count`` and ``chapter_count`` values.
+    """
+    if not candidates:
+        return {}
+
+    already_counted = set(counted_sizes)
+    candidate_sets_by_size: dict[int, set[str]] = defaultdict(set)
+    for phrase_norm, candidate in candidates.items():
+        if candidate.token_count in already_counted:
+            continue
+        candidate_sets_by_size[candidate.token_count].add(phrase_norm)
+
+    enriched = dict(candidates)
+    if not candidate_sets_by_size:
+        return enriched
+
+    total_counts, chapter_counts = count_candidate_occurrences(candidate_sets_by_size, chapters)
+    for phrase_norm, candidate in enriched.items():
+        if candidate.token_count in already_counted:
+            continue
+        candidate.raw_count = max(candidate.raw_count, total_counts[phrase_norm])
+        candidate.chapter_count = chapter_counts[phrase_norm]
+    return enriched
+
+
+def count_candidate_occurrences(
+    candidate_sets_by_size: Mapping[int, set[str]],
+    chapters: Sequence[str],
+) -> tuple[dict[str, int], dict[str, int]]:
+    """Count total occurrences and chapter spread for candidate phrases across chapters.
+
+    Args:
+        candidate_sets_by_size (Mapping[int, set[str]]): Candidate normalized phrases grouped by token count.
+        chapters (Sequence[str]): Chapter-like text blocks to slide n-gram windows over.
+
+    Returns:
+        tuple[dict[str, int], dict[str, int]]: Total occurrence counts and chapter-spread counts,
+        each keyed by normalized phrase.
+    """
+    total_counts: defaultdict[str, int] = defaultdict(int)
+    chapter_counts: defaultdict[str, int] = defaultdict(int)
+    for chapter in chapters:
+        seen_in_chapter: set[str] = set()
+        chapter_tokens = tokenize(chapter)
+        for ngram_size, candidate_norms in candidate_sets_by_size.items():
+            for start in range(len(chapter_tokens) - ngram_size + 1):
+                phrase_norm = " ".join(chapter_tokens[start : start + ngram_size])
+                if phrase_norm not in candidate_norms:
+                    continue
+                total_counts[phrase_norm] += 1
+                seen_in_chapter.add(phrase_norm)
+        for phrase_norm in seen_in_chapter:
+            chapter_counts[phrase_norm] += 1
+    return total_counts, chapter_counts
+
+
+def filter_storable_candidates(
+    candidates: Mapping[str, PhraseCandidate],
+    config: EbookSearchConfig,
+) -> tuple[dict[str, PhraseCandidate], int, int, int, int]:
+    """Remove candidates that should not be persisted.
+
+    Args:
+        candidates (Mapping[str, PhraseCandidate]): Candidates to filter, keyed by normalized phrase.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        tuple[dict[str, PhraseCandidate], int, int, int, int]: The storable candidates followed by the
+        counts dropped for being too short, too rare, too common, and junk.
+    """
+    min_raw_count = minimum_candidate_raw_count(config)
+    filtered: dict[str, PhraseCandidate] = {}
+    too_short = 0
+    too_rare = 0
+    too_common = 0
+    junk = 0
+    for phrase_norm, candidate in candidates.items():
+        if candidate.token_count < config.phrase_min_tokens:
+            too_short += 1
+            continue
+        if candidate.raw_count < min_raw_count:
+            too_rare += 1
+            continue
+        phrase_tokens = phrase_norm.split()
+        if is_most_common_word_phrase(phrase_tokens):
+            too_common += 1
+            continue
+        if is_junk_phrase(phrase_tokens):
+            junk += 1
+            continue
+        filtered[phrase_norm] = candidate
+    return filtered, too_short, too_rare, too_common, junk
+
+
+def minimum_candidate_raw_count(config: EbookSearchConfig) -> int:
+    """Return the minimum occurrence count required before storing a candidate.
+
+    Args:
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        int: The minimum raw occurrence count, never less than 1.
+    """
+    return max(config.phrase_raw_ngram_min_count, 1)
+
+
+def is_most_common_word_phrase(phrase_tokens: list[str]) -> bool:
+    """Return whether every token in a normalized phrase is a common word.
+
+    Args:
+        phrase_tokens (list[str]): Normalized phrase tokens to inspect.
+
+    Returns:
+        bool: True when the phrase is non-empty and every token is a common word.
+    """
+    common_words = get_most_common_words()
+    return bool(phrase_tokens) and all(token in common_words for token in phrase_tokens)
+
+
+def is_junk_phrase(phrase_tokens: list[str]) -> bool:
+    """Return whether a normalized phrase is lexical junk not worth LLM judging.
+
+    Judged data shows phrases containing a dialogue/action verb or a pronoun contraction are
+    never kept, and phrases whose tokens are mostly common words almost never are. Possessives
+    of proper nouns (``chapman's death``) pass because matching is by exact token, and
+    exactly-half-common bigrams (``data feed``) pass because the common-word rule is strict.
+
+    Args:
+        phrase_tokens (list[str]): Normalized phrase tokens to inspect.
+
+    Returns:
+        bool: True when the phrase contains a junk token or is majority common words.
+    """
+    if not phrase_tokens:
+        return False
+    junk_tokens = get_junk_tokens()
+    if any(token in junk_tokens for token in phrase_tokens):
+        return True
+    common_words = get_most_common_words()
+    half_phrase_len = len(phrase_tokens) // 2
+    return sum(token in common_words for token in phrase_tokens) > half_phrase_len
+
+
+def score_candidate(candidate: PhraseCandidate, config: EbookSearchConfig) -> float:
+    """Score a phrase candidate before LLM judging.
+
+    Args:
+        candidate (PhraseCandidate): Candidate to score.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        float: Combined score from sources, frequency, and length, less any penalties.
+    """
+    score = source_score(candidate) + frequency_score(candidate, config) + token_count_score(candidate, config)
+    if non_raw_source_count(candidate) >= MULTI_SOURCE_MIN_SOURCES:
+        score += MULTI_SOURCE_SCORE_BONUS
+    if candidate.phrase_norm in get_ignored_phrases():
+        score -= 100.0
+    if has_bad_start(candidate.phrase_norm):
+        score -= BAD_START_SCORE_PENALTY
+    if has_bad_end(candidate.phrase_norm):
+        score -= BAD_END_SCORE_PENALTY
+    return score
+
+
+def non_raw_source_count(candidate: PhraseCandidate) -> int:
+    """Count the non-raw-ngram extraction sources that produced a candidate.
+
+    Args:
+        candidate (PhraseCandidate): Candidate whose enabled sources are counted.
+
+    Returns:
+        int: Number of enabled sources other than the raw n-gram slide.
+    """
+    return sum(
+        (
+            candidate.source_yake,
+            candidate.source_spacy_ner,
+            candidate.source_spacy_noun_chunk,
+            candidate.source_capitalized,
+            candidate.source_metadata,
+        )
+    )
+
+
+def has_bad_start(phrase_norm: str) -> bool:
+    """Return whether a normalized phrase starts with a bad starting token.
+
+    Args:
+        phrase_norm (str): Normalized phrase text to inspect.
+
+    Returns:
+        bool: True when the first token is a known bad starting token.
+    """
+    phrase_tokens = phrase_norm.split()
+    return bool(phrase_tokens and phrase_tokens[0] in get_bad_starts())
+
+
+def has_bad_end(phrase_norm: str) -> bool:
+    """Return whether a normalized phrase ends with a bad ending token.
+
+    Args:
+        phrase_norm (str): Normalized phrase text to inspect.
+
+    Returns:
+        bool: True when the last token is a known bad ending token.
+    """
+    phrase_tokens = phrase_norm.split()
+    return bool(phrase_tokens and phrase_tokens[-1] in get_bad_ends())
+
+
+def source_score(candidate: PhraseCandidate) -> float:
+    """Return the score contribution from extraction sources.
+
+    Args:
+        candidate (PhraseCandidate): Candidate whose enabled sources are weighted.
+
+    Returns:
+        float: Summed weight of the candidate's enabled extraction sources.
+    """
+    return sum(
+        weight
+        for enabled, weight in (
+            (candidate.source_yake, 2.0),
+            (candidate.source_spacy_ner, 2.5),
+            (candidate.source_spacy_noun_chunk, 1.5),
+            (candidate.source_capitalized, 2.0),
+            (candidate.source_metadata, 2.0),
+            (candidate.source_raw_ngram, 0.5),
+        )
+        if enabled
+    )
+
+
+def frequency_score(candidate: PhraseCandidate, config: EbookSearchConfig) -> float:
+    """Return the score contribution from frequency and chapter spread.
+
+    Args:
+        candidate (PhraseCandidate): Candidate whose counts are scored.
+        config (EbookSearchConfig): Runtime phrase-tuning settings holding score thresholds.
+
+    Returns:
+        float: Summed weight for each frequency and chapter-spread threshold the candidate meets.
+    """
+    return sum(
+        weight
+        for count, threshold, weight in (
+            (candidate.raw_count, config.phrase_raw_count_score_threshold, 0.5),
+            (candidate.raw_count, config.phrase_raw_count_high_score_threshold, 0.5),
+            (candidate.chapter_count, config.phrase_chapter_count_score_threshold, 0.5),
+            (candidate.chapter_count, config.phrase_chapter_count_high_score_threshold, 0.5),
+        )
+        if count >= threshold
+    )
+
+
+def token_count_score(candidate: PhraseCandidate, config: EbookSearchConfig) -> float:
+    """Return the score contribution from phrase length.
+
+    Args:
+        candidate (PhraseCandidate): Candidate whose token count is scored.
+        config (EbookSearchConfig): Runtime phrase-tuning settings holding the max token bound.
+
+    Returns:
+        float: Length-based score contribution, which may be negative for over- or under-length phrases.
+    """
+    if candidate.token_count == 1:
+        return -0.5
+    if candidate.token_count in {2, 3, 4}:
+        return 0.5
+    if candidate.token_count > config.phrase_max_tokens:
+        return -1.0
+    return 0.0
+
+
+def get_sample_contexts(normalized_book_text: str, phrase_norm: str, max_contexts: int = 5) -> list[str]:
+    """Return normalized context snippets containing a candidate phrase.
+
+    ``normalized_book_text`` is expected to already be ``normalize_text``-ed by the caller
+    so the whole book is not re-normalized for every phrase.
+
+    Args:
+        normalized_book_text (str): Whole book text, already normalized, to search.
+        phrase_norm (str): Normalized phrase to find contexts around.
+        max_contexts (int): Maximum number of context snippets to return.
+
+    Returns:
+        list[str]: Up to ``max_contexts`` normalized snippets surrounding the phrase.
+    """
+    contexts: list[str] = []
+    start = 0
+    while len(contexts) < max_contexts:
+        index = normalized_book_text.find(phrase_norm, start)
+        if index == -1:
+            break
+        left = max(0, index - 300)
+        right = min(len(normalized_book_text), index + len(phrase_norm) + 300)
+        contexts.append(normalized_book_text[left:right])
+        start = index + len(phrase_norm)
+    return contexts
+
+
+def candidate_source_names(candidate: PhraseCandidate) -> list[str]:
+    """Return enabled source names for an extracted candidate.
+
+    Args:
+        candidate (PhraseCandidate): Candidate whose enabled sources are listed.
+
+    Returns:
+        list[str]: Names of the extraction sources that produced the candidate.
+    """
+    names: list[str] = []
+    if candidate.source_raw_ngram:
+        names.append("raw_ngram")
+    if candidate.source_yake:
+        names.append("yake")
+    if candidate.source_spacy_ner:
+        names.append("spacy_ner")
+    if candidate.source_spacy_noun_chunk:
+        names.append("spacy_noun_chunk")
+    if candidate.source_capitalized:
+        names.append("capitalized")
+    if candidate.source_metadata:
+        names.append("metadata")
+    return names
+
+
+def extract_phrase_candidates_for_book(
+    book_text: str,
+    chapters: Sequence[str],
+    config: EbookSearchConfig,
+    *,
+    nlp: SpacyLanguage | None = None,
+    metadata: Mapping[str, object] | None = None,
+) -> list[PhraseCandidate]:
+    """Extract, score, and limit phrase candidates for one book.
+
+    Args:
+        book_text (str): Full book text used for most extraction sources.
+        chapters (Sequence[str]): Chapter-like text blocks used for spaCy and frequency counts.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        nlp (SpacyLanguage | None): Optional spaCy pipeline for entity and noun-chunk sources.
+        metadata (Mapping[str, object] | None): Optional book metadata used as a candidate source.
+
+    Returns:
+        list[PhraseCandidate]: Scored candidates sorted best-first and capped per book.
+    """
+    started_at = perf_counter()
+    logger.info(
+        "ebook_phrase_candidate_extract_start chapters=%s chars=%s min_tokens=%s max_tokens=%s max_candidates=%s",
+        len(chapters),
+        len(book_text),
+        config.phrase_min_tokens,
+        config.phrase_max_tokens,
+        config.protected_phrase_max_candidates_per_book,
+    )
+    raw_started_at = perf_counter()
+    raw = extract_raw_ngrams_by_chapter(chapters, config)
+    logger.info(
+        "ebook_phrase_candidate_extract_raw_complete candidates=%s duration_ms=%.1f",
+        len(raw),
+        (perf_counter() - raw_started_at) * 1000,
+    )
+    yake_started_at = perf_counter()
+    yake_candidates = extract_yake_candidates(book_text, config)
+    logger.info(
+        "ebook_phrase_candidate_extract_yake_complete candidates=%s duration_ms=%.1f",
+        len(yake_candidates),
+        (perf_counter() - yake_started_at) * 1000,
+    )
+    spacy_candidates: dict[str, PhraseCandidate] = {}
+    if nlp is not None:
+        spacy_started_at = perf_counter()
+        for chapter in chapters:
+            spacy_candidates = merge_candidate_sources(spacy_candidates, extract_spacy_candidates(chapter, nlp, config))
+        logger.info(
+            "ebook_phrase_candidate_extract_spacy_complete candidates=%s duration_ms=%.1f",
+            len(spacy_candidates),
+            (perf_counter() - spacy_started_at) * 1000,
+        )
+    capitalized_started_at = perf_counter()
+    capitalized = extract_capitalized_phrases(book_text, config)
+    logger.info(
+        "ebook_phrase_candidate_extract_capitalized_complete candidates=%s duration_ms=%.1f",
+        len(capitalized),
+        (perf_counter() - capitalized_started_at) * 1000,
+    )
+    metadata_candidates = extract_metadata_candidates(metadata, config)
+
+    candidates = merge_candidate_sources(raw, yake_candidates, spacy_candidates, capitalized, metadata_candidates)
+    enriched_started_at = perf_counter()
+    # Raw n-gram sizes were already counted per chapter above, so only enrich the remaining
+    # (entity-length) sizes here instead of re-sliding every size over the whole book.
+    candidates = enrich_with_frequency_and_chapter_counts(
+        candidates,
+        chapters,
+        counted_sizes=range(config.phrase_min_tokens, config.phrase_max_tokens + 1),
+    )
+    pre_filter_count = len(candidates)
+    candidates, filtered_too_short, filtered_too_rare, filtered_too_common, filtered_junk = filter_storable_candidates(
+        candidates, config
+    )
+    for candidate in candidates.values():
+        candidate.candidate_score = score_candidate(candidate, config)
+
+    limited = sorted(candidates.values(), key=lambda item: item.candidate_score, reverse=True)[
+        : config.protected_phrase_max_candidates_per_book
+    ]
+    logger.info(
+        "ebook_phrase_candidate_extract_complete raw=%s yake=%s spacy=%s capitalized=%s metadata=%s "
+        "merged=%s filtered_too_short=%s filtered_too_rare=%s filtered_too_common=%s filtered_junk=%s "
+        "min_uses=%s storable=%s limited=%s enrich_score_ms=%.1f duration_ms=%.1f",
+        len(raw),
+        len(yake_candidates),
+        len(spacy_candidates),
+        len(capitalized),
+        len(metadata_candidates),
+        pre_filter_count,
+        filtered_too_short,
+        filtered_too_rare,
+        filtered_too_common,
+        filtered_junk,
+        minimum_candidate_raw_count(config),
+        len(candidates),
+        len(limited),
+        (perf_counter() - enriched_started_at) * 1000,
+        (perf_counter() - started_at) * 1000,
+    )
+    return limited

python/ebook_search/protected_phrases/generate_ngrams.py

@@ -0,0 +1,370 @@
+"""Book-level orchestration for candidate n-gram generation and recalculation."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from collections import deque
+from time import perf_counter
+from typing import TYPE_CHECKING
+
+from sqlalchemy import select
+
+from python.ebook_search.protected_phrases.extraction import extract_phrase_candidates_for_book
+from python.ebook_search.protected_phrases.models import (
+    BookCandidateResult,
+    PhraseCandidateGenerationResult,
+    PhraseRecalculationResult,
+)
+from python.ebook_search.protected_phrases.pool import extract_phrase_candidates_in_pool, get_extraction_pool
+from python.ebook_search.protected_phrases.store import (
+    bulk_upsert_unjudged_candidates,
+    delete_phrase_data_for_book,
+    load_book_chapter_texts,
+    metadata_for_source,
+    new_candidate_row,
+    prune_unstorable_unjudged_candidate_phrases,
+)
+from python.orm.richie import EbookCandidatePhrase, EbookSource
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping, Sequence
+    from concurrent.futures import Future
+
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.protected_phrases.extraction import SpacyLanguage
+    from python.ebook_search.protected_phrases.models import PhraseCandidate
+
+logger = logging.getLogger(__name__)
+
+
+async def generate_candidate_phrases_for_books(
+    session: AsyncSession,
+    config: EbookSearchConfig,
+    *,
+    only_missing: bool = False,
+) -> PhraseCandidateGenerationResult:
+    """Create or refresh candidate phrases for indexed books without calling the LLM judge.
+
+    Extraction always runs concurrently in the shared process pool so a full backfill uses
+    multiple cores.
+
+    Args:
+        session (Session): Active database session.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        only_missing (bool): When True, only generate for books that have no candidate phrases
+            yet instead of refreshing every book.
+
+    Returns:
+        PhraseCandidateGenerationResult: Per-corpus counts of books seen, built, and candidates stored.
+    """
+    source_query = select(EbookSource).order_by(EbookSource.id)
+    if only_missing:
+        has_candidates = select(EbookCandidatePhrase.id).where(EbookCandidatePhrase.book_id == EbookSource.id)
+        source_query = source_query.where(~has_candidates.exists())
+    sources = (await session.scalars(source_query)).all()
+    books_seen = len(sources)
+    logger.info(
+        "ebook_candidate_phrase_generation_start books_seen=%s min_tokens=%s max_tokens=%s max_candidates_per_book=%s",
+        books_seen,
+        config.phrase_min_tokens,
+        config.phrase_max_tokens,
+        config.protected_phrase_max_candidates_per_book,
+    )
+
+    outcomes = await generate_candidates_for_sources_pooled(session, sources, config)
+
+    result = PhraseCandidateGenerationResult(
+        books_seen=books_seen,
+        books_built=sum(1 for outcome in outcomes if outcome.built),
+        candidate_phrases=sum(outcome.candidates for outcome in outcomes),
+    )
+    logger.info(
+        "ebook_candidate_phrase_generation_complete books_seen=%s books_built=%s candidate_total=%s",
+        result.books_seen,
+        result.books_built,
+        result.candidate_phrases,
+    )
+    return result
+
+
+async def generate_candidates_for_sources_pooled(
+    session: AsyncSession,
+    sources: Sequence[EbookSource],
+    config: EbookSearchConfig,
+) -> list[BookCandidateResult]:
+    """Generate candidate phrases for many books, extracting them concurrently in worker processes.
+
+    Chapter loading and row persistence stay on the caller's session (serial), while the CPU-bound
+    extraction runs in the shared process pool. A bounded window of in-flight books overlaps
+    extraction across cores without loading every book's candidates into memory at once.
+
+    Args:
+        session (Session): Active database session.
+        sources (Sequence[EbookSource]): Indexed books to generate candidates for.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        list[BookCandidateResult]: One result per book.
+    """
+    pool = get_extraction_pool(config.protected_phrase_extraction_workers)
+    max_in_flight = max(1, config.protected_phrase_extraction_workers) * 2
+    pending: deque[tuple[EbookSource, Future[list[PhraseCandidate]]]] = deque()
+    outcomes: list[BookCandidateResult] = []
+
+    async def drain_one() -> None:
+        source, future = pending.popleft()
+        extracted = await asyncio.wrap_future(future)
+        outcomes.append(await store_source_candidates(session, source, extracted, config))
+
+    try:
+        for source in sources:
+            chapters = await load_book_chapter_texts(session, source.id)
+            if not chapters:
+                logger.warning("ebook_candidate_phrase_generation_book_empty source_id=%s", source.id)
+                outcomes.append(BookCandidateResult())
+                continue
+            future = pool.submit(
+                extract_phrase_candidates_for_book,
+                "\n\n".join(chapters),
+                chapters,
+                config,
+                metadata=metadata_for_source(source),
+            )
+            pending.append((source, future))
+            if len(pending) >= max_in_flight:
+                await drain_one()
+        while pending:
+            await drain_one()
+    except Exception:
+        for _, future in pending:
+            future.cancel()
+        await session.rollback()
+        logger.exception("ebook_candidate_phrase_generation_pooled_failed")
+        raise
+    return outcomes
+
+
+async def store_source_candidates(
+    session: AsyncSession,
+    source: EbookSource,
+    limited_candidates: list[PhraseCandidate],
+    config: EbookSearchConfig,
+) -> BookCandidateResult:
+    """Persist and commit one book's already-extracted candidates.
+
+    Args:
+        session (AsyncSession): Active database session.
+        source (EbookSource): Book the candidates belong to.
+        limited_candidates (list[PhraseCandidate]): Scored candidates to persist.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        BookCandidateResult: Candidate count and that the book was committed.
+    """
+    book_started_at = perf_counter()
+    saved_count = await store_candidate_phrases_for_book(session, source.id, None, limited_candidates, config)
+    await session.commit()
+    logger.info(
+        "ebook_candidate_phrase_generation_book_committed source_id=%s candidates=%s duration_ms=%.1f",
+        source.id,
+        saved_count,
+        (perf_counter() - book_started_at) * 1000,
+    )
+    return BookCandidateResult(candidates=saved_count, built=True)
+
+
+async def recalculate_candidate_phrases_for_book(
+    session: AsyncSession,
+    source: EbookSource,
+    config: EbookSearchConfig,
+    *,
+    nlp: SpacyLanguage | None = None,
+    use_process_pool: bool = False,
+) -> PhraseRecalculationResult:
+    """Remove all book phrase data, regenerate candidates, and commit the completed book.
+
+    Args:
+        session (Session): Active database session.
+        source (EbookSource): Indexed book to recalculate.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        nlp (SpacyLanguage | None): Optional spaCy pipeline for entity and noun-chunk sources.
+        use_process_pool (bool): Run the CPU-bound extraction in a worker process so concurrent
+            recalculations do not serialize behind the GIL. Defaults to in-process for callers
+            (tests, backfills) that do not need it.
+
+    Returns:
+        PhraseRecalculationResult: Deleted-row counts and the number of candidates regenerated.
+    """
+    started_at = perf_counter()
+    logger.info(
+        "ebook_candidate_phrase_recalculation_start source_id=%s title=%r",
+        source.id,
+        source.title,
+    )
+    try:
+        deleted = await delete_phrase_data_for_book(session, source.id)
+        chapters = await load_book_chapter_texts(session, source.id)
+        if not chapters:
+            logger.warning("ebook_candidate_phrase_recalculation_book_empty source_id=%s", source.id)
+            await session.commit()
+            return PhraseRecalculationResult(
+                book_id=source.id,
+                deleted_candidates=deleted.deleted_candidates,
+                deleted_protected_phrases=deleted.deleted_protected_phrases,
+                deleted_aliases=deleted.deleted_aliases,
+                deleted_mentions=deleted.deleted_mentions,
+                candidate_phrases=0,
+            )
+
+        candidate_count = await generate_candidate_phrases_for_book(
+            session,
+            source.id,
+            series_id=None,
+            chapters=chapters,
+            config=config,
+            nlp=nlp,
+            metadata=metadata_for_source(source),
+            replace_all=True,
+            use_process_pool=use_process_pool,
+        )
+        await session.commit()
+    except Exception:
+        await session.rollback()
+        logger.exception("ebook_candidate_phrase_recalculation_failed source_id=%s", source.id)
+        raise
+
+    result = PhraseRecalculationResult(
+        book_id=source.id,
+        deleted_candidates=deleted.deleted_candidates,
+        deleted_protected_phrases=deleted.deleted_protected_phrases,
+        deleted_aliases=deleted.deleted_aliases,
+        deleted_mentions=deleted.deleted_mentions,
+        candidate_phrases=candidate_count,
+    )
+    logger.info(
+        "ebook_candidate_phrase_recalculation_complete source_id=%s deleted_candidates=%s "
+        "deleted_protected=%s deleted_aliases=%s deleted_mentions=%s candidates=%s duration_ms=%.1f",
+        source.id,
+        result.deleted_candidates,
+        result.deleted_protected_phrases,
+        result.deleted_aliases,
+        result.deleted_mentions,
+        result.candidate_phrases,
+        (perf_counter() - started_at) * 1000,
+    )
+    return result
+
+
+async def generate_candidate_phrases_for_book(
+    session: AsyncSession,
+    book_id: int,
+    series_id: int | None,
+    chapters: Sequence[str],
+    config: EbookSearchConfig,
+    *,
+    nlp: SpacyLanguage | None = None,
+    metadata: Mapping[str, object] | None = None,
+    replace_all: bool = False,
+    use_process_pool: bool = False,
+) -> int:
+    """Extract and store candidate phrases for one book without LLM judging.
+
+    Args:
+        session (Session): Active database session.
+        book_id (int): Book the candidates belong to.
+        series_id (int | None): Series scope for the stored candidates.
+        chapters (Sequence[str]): Chapter-like text blocks used for extraction and frequency counts.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        nlp (SpacyLanguage | None): Optional spaCy pipeline for entity and noun-chunk sources.
+        metadata (Mapping[str, object] | None): Optional book metadata used as a candidate source.
+        replace_all (bool): When the caller has already cleared this book's candidates (e.g. a
+            recalculation), skip the per-candidate existence lookup and bulk-insert new rows.
+        use_process_pool (bool): Run the CPU-bound extraction in a worker process to avoid
+            serializing concurrent requests behind the GIL. Ignored when ``nlp`` is set, since
+            the spaCy pipeline cannot be sent to a worker process.
+
+    Returns:
+        int: Number of candidate phrase rows stored.
+    """
+    started_at = perf_counter()
+    book_text = "\n\n".join(chapters)
+    if use_process_pool and nlp is None:
+        limited_candidates = await extract_phrase_candidates_in_pool(book_text, chapters, config, metadata=metadata)
+    else:
+        limited_candidates = extract_phrase_candidates_for_book(
+            book_text,
+            chapters,
+            config,
+            nlp=nlp,
+            metadata=metadata,
+        )
+    saved_count = await store_candidate_phrases_for_book(
+        session,
+        book_id,
+        series_id,
+        limited_candidates,
+        config,
+        replace_all=replace_all,
+    )
+    logger.info(
+        "ebook_candidate_phrase_generation_book_duration book_id=%s candidates=%s duration_ms=%.1f",
+        book_id,
+        saved_count,
+        (perf_counter() - started_at) * 1000,
+    )
+    return saved_count
+
+
+async def store_candidate_phrases_for_book(
+    session: AsyncSession,
+    book_id: int,
+    series_id: int | None,
+    limited_candidates: list[PhraseCandidate],
+    config: EbookSearchConfig,
+    *,
+    replace_all: bool = False,
+) -> int:
+    """Persist already-extracted candidate phrase rows for one book without committing.
+
+    Args:
+        session (Session): Active database session.
+        book_id (int): Book the candidates belong to.
+        series_id (int | None): Series scope for the stored candidates.
+        limited_candidates (list[PhraseCandidate]): Scored candidates to persist.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        replace_all (bool): When the caller has already cleared this book's candidates, skip the
+            per-candidate existence lookup and bulk-insert new rows.
+
+    Returns:
+        int: Number of candidate phrase rows stored.
+    """
+    save_started_at = perf_counter()
+    if replace_all:
+        rows = [new_candidate_row(book_id, series_id, candidate) for candidate in limited_candidates]
+        session.add_all(rows)
+        await session.flush()
+        saved_count = len(rows)
+        logger.info(
+            "ebook_candidate_phrase_save_start book_id=%s candidates=%s mode=bulk_insert",
+            book_id,
+            len(limited_candidates),
+        )
+    else:
+        pruned_count = await prune_unstorable_unjudged_candidate_phrases(session, book_id, config)
+        logger.info(
+            "ebook_candidate_phrase_save_start book_id=%s candidates=%s pruned_unstorable=%s",
+            book_id,
+            len(limited_candidates),
+            pruned_count,
+        )
+        saved_count = await bulk_upsert_unjudged_candidates(session, book_id, series_id, limited_candidates)
+    logger.info(
+        "ebook_candidate_phrase_save_complete book_id=%s candidates=%s save_ms=%.1f",
+        book_id,
+        saved_count,
+        (perf_counter() - save_started_at) * 1000,
+    )
+    return saved_count

python/ebook_search/protected_phrases/judge_ngrams.py

@@ -0,0 +1,511 @@
+"""Book-level orchestration for LLM judging and promotion of candidate phrases."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import re
+from time import perf_counter
+from typing import TYPE_CHECKING
+
+import httpx
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from python.ebook_search.llm_interface import request_chat_completion
+from python.ebook_search.protected_phrases.extraction import (
+    candidate_source_names,
+    get_sample_contexts,
+    is_junk_phrase,
+    is_most_common_word_phrase,
+    score_candidate,
+)
+from python.ebook_search.protected_phrases.matching import index_chunk_phrase_mentions_for_book
+from python.ebook_search.protected_phrases.models import BookJudgmentResult, LLMJudgment, PhraseJudgmentBackfillResult
+from python.ebook_search.protected_phrases.store import (
+    count_protected_phrases,
+    count_unjudged_candidates,
+    load_book_text,
+    load_candidates_for_judgment,
+    phrase_candidate_from_row,
+    save_candidate_to_db,
+    upsert_protected_phrase,
+)
+from python.ebook_search.protected_phrases.text_normalization import normalize_text
+from python.orm.richie import EbookSource
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from sqlalchemy.ext.asyncio import AsyncEngine
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.protected_phrases.models import PhraseCandidate
+    from python.orm.richie import EbookProtectedPhrase
+
+JSON_OBJECT_RE = re.compile(r"\{.*\}", re.DOTALL)
+
+
+logger = logging.getLogger(__name__)
+
+
+async def judge_candidate_phrases_for_books(
+    engine: AsyncEngine,
+    config: EbookSearchConfig,
+    *,
+    source_ids: Sequence[int] | None = None,
+) -> PhraseJudgmentBackfillResult:
+    """Judge candidate phrases for books, fanning LLM calls out across books and phrases.
+
+    Up to ``phrase_judge_book_workers`` books are judged at once, and within each book candidates
+    are judged in concurrent chunks of ``phrase_judge_phrase_workers``. Each book uses its own
+    short-lived sessions for reads and writes; no database connection is held while LLM calls are
+    in flight. For a pseudo-single-threaded run (solo testing, debugging), set both worker
+    settings to 1.
+
+    Args:
+        engine (AsyncEngine): Engine used to open one session per book.
+        config (EbookSearchConfig): Runtime phrase-tuning settings and chat configuration.
+        source_ids (Sequence[int] | None): Books to judge; ``None`` judges every indexed book.
+
+    Returns:
+        PhraseJudgmentBackfillResult: Per-corpus counts of books judged, failures, candidates,
+        protected phrases, and mentions.
+    """
+    if source_ids is None:
+        async with AsyncSession(engine) as session:
+            source_ids = list((await session.scalars(select(EbookSource.id).order_by(EbookSource.id))).all())
+    books_seen = len(source_ids)
+    book_workers = max(1, config.phrase_judge_book_workers)
+    phrase_workers = max(1, config.phrase_judge_phrase_workers)
+    logger.info(
+        "ebook_candidate_phrase_judgment_start books_seen=%s book_workers=%s phrase_workers=%s "
+        "confidence_threshold=%.2f",
+        books_seen,
+        book_workers,
+        phrase_workers,
+        config.protected_phrase_confidence_threshold,
+    )
+
+    book_semaphore = asyncio.Semaphore(book_workers)
+    max_connections = book_workers * phrase_workers
+    limits = httpx.Limits(max_connections=max_connections, max_keepalive_connections=max_connections)
+    async with httpx.AsyncClient(limits=limits) as client:
+        outcomes = await asyncio.gather(
+            *(judge_one_book_async(engine, source_id, config, client, book_semaphore) for source_id in source_ids)
+        )
+
+    result = PhraseJudgmentBackfillResult(
+        books_seen=books_seen,
+        books_judged=sum(1 for outcome in outcomes if outcome.committed),
+        books_failed=sum(1 for outcome in outcomes if outcome.failed),
+        candidates_judged=sum(outcome.judged for outcome in outcomes),
+        protected_phrases=sum(outcome.protected for outcome in outcomes),
+        phrase_mentions=sum(outcome.mentions for outcome in outcomes),
+    )
+    logger.info(
+        "ebook_candidate_phrase_judgment_complete books_seen=%s books_judged=%s books_failed=%s "
+        "candidates_judged=%s protected=%s mentions=%s",
+        result.books_seen,
+        result.books_judged,
+        result.books_failed,
+        result.candidates_judged,
+        result.protected_phrases,
+        result.phrase_mentions,
+    )
+    return result
+
+
+async def judge_one_book_async(
+    engine: AsyncEngine,
+    source_id: int,
+    config: EbookSearchConfig,
+    client: httpx.AsyncClient,
+    book_semaphore: asyncio.Semaphore,
+) -> BookJudgmentResult:
+    """Judge one book concurrently and persist the outcome, honoring the book-level limit.
+
+    Args:
+        engine (AsyncEngine): Engine used to open the book's read and write sessions.
+        source_id (int): Book to judge candidates for.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        client (httpx.AsyncClient): Shared async client for LLM calls.
+        book_semaphore (asyncio.Semaphore): Caps how many books judge at once.
+
+    Returns:
+        BookJudgmentResult: The book's judgment outcome.
+    """
+    async with book_semaphore:
+        try:
+            prepared = await prepare_book_judgment(engine, source_id, config)
+            if prepared is None:
+                return BookJudgmentResult()
+            work_items, target_remaining = prepared
+            judged = await judge_book_candidates_async(client, config, source_id, work_items, target_remaining)
+            if not judged:
+                return BookJudgmentResult()
+            return await persist_book_judgments(engine, source_id, config, judged)
+        except Exception:
+            logger.exception("ebook_candidate_phrase_judgment_book_failed source_id=%s", source_id)
+            return BookJudgmentResult(failed=True)
+
+
+async def prepare_book_judgment(
+    engine: AsyncEngine,
+    source_id: int,
+    config: EbookSearchConfig,
+) -> tuple[list[tuple[int, PhraseCandidate]], int | None] | None:
+    """Load one book's candidates to judge, with sample contexts, on a short-lived read session.
+
+    Args:
+        engine (AsyncEngine): Engine used to open the read session.
+        source_id (int): Book to load candidates for.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+
+    Returns:
+        tuple[list[tuple[int, PhraseCandidate]], int | None] | None: Candidate rows paired with
+        in-memory candidates and the remaining protected-phrase target, or ``None`` when the book
+        has nothing to judge.
+    """
+    judgment_limit = config.protected_phrase_llm_candidates_per_book
+    if judgment_limit <= 0:
+        return None
+    async with AsyncSession(engine) as session:
+        if not await count_unjudged_candidates(session, source_id, config):
+            logger.info("ebook_candidate_phrase_judgment_book_skip_no_unjudged source_id=%s", source_id)
+            return None
+        existing_protected = await count_protected_phrases(session, source_id)
+        target_remaining: int | None = None
+        if config.phrase_target_protected_per_book > 0:
+            target_remaining = max(config.phrase_target_protected_per_book - existing_protected, 0)
+            if target_remaining == 0:
+                logger.info(
+                    "ebook_candidate_phrase_judgment_skipped_target_met source_id=%s existing_protected=%s target=%s",
+                    source_id,
+                    existing_protected,
+                    config.phrase_target_protected_per_book,
+                )
+                return None
+        book_text = await load_book_text(session, source_id)
+        if not book_text:
+            logger.warning("ebook_candidate_phrase_judgment_book_empty source_id=%s", source_id)
+            return None
+        normalized_book_text = normalize_text(book_text)
+        # Stored rows may predate the current junk filters and score weights, so re-filter and
+        # rescore every unjudged row here instead of trusting the persisted candidate_score.
+        rows = await load_candidates_for_judgment(session, source_id, config)
+        scored_items: list[tuple[int, PhraseCandidate]] = []
+        skipped_junk = 0
+        for row in rows:
+            candidate = phrase_candidate_from_row(row)
+            if is_junk_phrase(candidate.phrase_norm.split()):
+                skipped_junk += 1
+                continue
+            candidate.candidate_score = score_candidate(candidate, config)
+            scored_items.append((row.id, candidate))
+        scored_items.sort(key=lambda item: item[1].candidate_score, reverse=True)
+        work_items = scored_items[:judgment_limit]
+        for _, candidate in work_items:
+            candidate.sample_contexts = candidate.sample_contexts or get_sample_contexts(
+                normalized_book_text, candidate.phrase_norm
+            )
+    logger.info(
+        "ebook_candidate_phrase_judgment_candidates_loaded source_id=%s candidates=%s skipped_junk=%s "
+        "unjudged_rows=%s existing_protected=%s target_remaining=%s judgment_limit=%s",
+        source_id,
+        len(work_items),
+        skipped_junk,
+        len(rows),
+        existing_protected,
+        target_remaining,
+        judgment_limit,
+    )
+    return work_items, target_remaining
+
+
+async def judge_book_candidates_async(
+    client: httpx.AsyncClient,
+    config: EbookSearchConfig,
+    source_id: int,
+    work_items: list[tuple[int, PhraseCandidate]],
+    target_remaining: int | None,
+) -> list[tuple[int, PhraseCandidate, LLMJudgment, bool]]:
+    """Judge a book's candidates in concurrent chunks, stopping once the target is reached.
+
+    Promotion decisions are made in memory so judging can stop early without any database writes.
+
+    Args:
+        client (httpx.AsyncClient): Shared async client for LLM calls.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        source_id (int): Book being judged, for logging.
+        work_items (list[tuple[int, PhraseCandidate]]): Candidate row ids paired with candidates,
+            in best-first score order.
+        target_remaining (int | None): Remaining protected-phrase target, or ``None`` for no cap.
+
+    Returns:
+        list[tuple[int, PhraseCandidate, LLMJudgment, bool]]: Judged rows with their judgment and
+        whether each should be promoted.
+    """
+    chunk_size = max(1, config.phrase_judge_phrase_workers)
+    judged: list[tuple[int, PhraseCandidate, LLMJudgment, bool]] = []
+    promoted = 0
+    for start in range(0, len(work_items), chunk_size):
+        chunk = work_items[start : start + chunk_size]
+        judgments = await asyncio.gather(*(judge_candidate_async(client, config, candidate) for _, candidate in chunk))
+        for (candidate_id, candidate), judgment in zip(chunk, judgments, strict=True):
+            promote = (target_remaining is None or promoted < target_remaining) and should_protect_judged_candidate(
+                candidate, judgment, source_id, config, candidate_id=candidate_id
+            )
+            if promote:
+                promoted += 1
+            judged.append((candidate_id, candidate, judgment, promote))
+        if target_remaining is not None and promoted >= target_remaining:
+            break
+    return judged
+
+
+async def judge_candidate_async(
+    client: httpx.AsyncClient,
+    config: EbookSearchConfig,
+    candidate: PhraseCandidate,
+) -> LLMJudgment:
+    """Judge one candidate with the LLM over the shared async client.
+
+    Args:
+        client (httpx.AsyncClient): Shared async client for LLM calls.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        candidate (PhraseCandidate): Candidate to judge.
+
+    Returns:
+        LLMJudgment: The parsed judgment.
+    """
+    content = await request_chat_completion(client, config, build_judge_messages(candidate))
+    return parse_llm_judgment(content, config)
+
+
+async def persist_book_judgments(
+    engine: AsyncEngine,
+    source_id: int,
+    config: EbookSearchConfig,
+    judged: list[tuple[int, PhraseCandidate, LLMJudgment, bool]],
+) -> BookJudgmentResult:
+    """Persist one book's judgments and promotions in a single committed transaction.
+
+    Args:
+        engine (AsyncEngine): Engine used to open the write session.
+        source_id (int): Book being persisted.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        judged (list[tuple[int, PhraseCandidate, LLMJudgment, bool]]): Judged candidates with their
+            judgment and promotion flag.
+
+    Returns:
+        BookJudgmentResult: The book's committed counts, or a failed result on error.
+    """
+    book_started_at = perf_counter()
+    async with AsyncSession(engine, expire_on_commit=False) as session:
+        try:
+            protected: list[EbookProtectedPhrase] = []
+            for candidate_id, candidate, judgment, promote in judged:
+                candidate_row = await save_candidate_to_db(session, source_id, None, candidate, judgment=judgment)
+                if promote:
+                    protected.append(
+                        await upsert_protected_phrase(session, source_id, None, candidate, judgment, candidate_row)
+                    )
+                logger.info(
+                    "ebook_candidate_phrase_judgment_candidate_complete source_id=%s candidate_id=%s phrase=%r "
+                    "keep=%s confidence=%.3f category=%r promoted=%s",
+                    source_id,
+                    candidate_id,
+                    candidate.phrase_norm,
+                    judgment.keep,
+                    judgment.confidence,
+                    judgment.category,
+                    promote,
+                )
+            await session.flush()
+            mentions = await index_chunk_phrase_mentions_for_book(session, source_id, config) if protected else 0
+            await session.commit()
+        except Exception:
+            await session.rollback()
+            logger.exception("ebook_candidate_phrase_judgment_book_persist_failed source_id=%s", source_id)
+            return BookJudgmentResult(failed=True)
+    logger.info(
+        "ebook_candidate_phrase_judgment_book_committed source_id=%s judged=%s protected=%s mentions=%s "
+        "duration_ms=%.1f",
+        source_id,
+        len(judged),
+        len(protected),
+        mentions,
+        (perf_counter() - book_started_at) * 1000,
+    )
+    return BookJudgmentResult(judged=len(judged), protected=len(protected), mentions=mentions, committed=True)
+
+
+def should_protect_judged_candidate(
+    candidate: PhraseCandidate,
+    judgment: LLMJudgment,
+    book_id: int,
+    config: EbookSearchConfig,
+    *,
+    candidate_id: int,
+) -> bool:
+    """Report whether a judged candidate qualifies to become a protected phrase.
+
+    Args:
+        candidate (PhraseCandidate): In-memory candidate that was judged.
+        judgment (LLMJudgment): Judge decision for the candidate.
+        book_id (int): Book the candidate belongs to, for logging.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        candidate_id (int): Stored candidate row id the judgment came from, for logging.
+
+    Returns:
+        bool: True when the judged candidate should be promoted to a protected phrase.
+    """
+    if not judgment.keep or judgment.confidence < config.protected_phrase_confidence_threshold:
+        return False
+    accepted_norm = normalize_text(judgment.canonical or candidate.phrase_text)
+    accepted_tokens = accepted_norm.split()
+    accepted_token_count = len(accepted_tokens)
+    if accepted_token_count < config.phrase_min_tokens:
+        logger.info(
+            "ebook_candidate_phrase_judgment_candidate_skip_short_canonical book_id=%s candidate_id=%s "
+            "phrase=%r canonical=%r token_count=%s min_tokens=%s",
+            book_id,
+            candidate_id,
+            candidate.phrase_norm,
+            accepted_norm,
+            accepted_token_count,
+            config.phrase_min_tokens,
+        )
+        return False
+    if is_most_common_word_phrase(accepted_tokens):
+        logger.info(
+            "ebook_candidate_phrase_judgment_candidate_skip_common_canonical book_id=%s candidate_id=%s "
+            "phrase=%r canonical=%r",
+            book_id,
+            candidate_id,
+            candidate.phrase_norm,
+            accepted_norm,
+        )
+        return False
+    return True
+
+
+def build_judge_messages(candidate: PhraseCandidate) -> list[dict[str, str]]:
+    """Build the chat messages used to judge one candidate phrase.
+
+    Args:
+        candidate (PhraseCandidate): Candidate to describe for the judge.
+
+    Returns:
+        list[dict[str, str]]: OpenAI-style system and user messages.
+    """
+    payload = {
+        "phrase": candidate.phrase_norm,
+        "token_count": candidate.token_count,
+        "sources": candidate_source_names(candidate),
+        "raw_count": candidate.raw_count,
+        "chapter_count": candidate.chapter_count,
+        "contexts": candidate.sample_contexts,
+    }
+    return [
+        {
+            "role": "system",
+            "content": (
+                "Judge whether a candidate phrase from a book should be protected for RAG retrieval. "
+                "Do not extract new phrases. Reject common grammar fragments, ordinary nonspecific phrases, "
+                "unstable fragments, and phrases kept only because they are frequent. Keep people, places, "
+                "organizations, factions, events, technologies, fictional conditions, magic systems, formal titles, "
+                "named concepts, and recurring world-specific terms. Return only a JSON object with keys: keep, "
+                "canonical, category, aliases, confidence, importance, allow_nested, suppress_children, reason."
+            ),
+        },
+        {"role": "user", "content": json.dumps(payload, ensure_ascii=True)},
+    ]
+
+
+def parse_llm_judgment(content: str, config: EbookSearchConfig) -> LLMJudgment:
+    """Parse and validate an LLM phrase-judge response.
+
+    Args:
+        content (str): Raw model response text.
+        config (EbookSearchConfig): Runtime phrase-tuning settings supplying nesting defaults.
+
+    Returns:
+        LLMJudgment: The parsed and validated judgment.
+
+    Raises:
+        TypeError: If the decoded JSON body is not an object.
+    """
+    body = json.loads(extract_json_object(content))
+    if not isinstance(body, dict):
+        msg = "LLM phrase judge response is not a JSON object"
+        raise TypeError(msg)
+
+    aliases = body.get("aliases", ())
+    if not isinstance(aliases, list | tuple):
+        aliases = ()
+    return LLMJudgment(
+        keep=bool(body.get("keep", False)),
+        canonical=optional_text(body.get("canonical")),
+        category=optional_text(body.get("category")),
+        aliases=tuple(str(alias) for alias in aliases if isinstance(alias, str) and alias.strip()),
+        confidence=clamped_float(body.get("confidence"), default=0.0),
+        importance=clamped_float(body.get("importance"), default=0.5),
+        allow_nested=bool(body.get("allow_nested", config.phrase_default_allow_nested)),
+        suppress_children=bool(body.get("suppress_children", config.phrase_default_suppress_children)),
+        reason=optional_text(body.get("reason")),
+    )
+
+
+def extract_json_object(content: str) -> str:
+    """Extract a JSON object from plain or fenced model output.
+
+    Args:
+        content (str): Raw model response text.
+
+    Returns:
+        str: The substring spanning the first JSON object.
+
+    Raises:
+        ValueError: If no JSON object is found in the response.
+    """
+    stripped = content.strip()
+    if stripped.startswith("{") and stripped.endswith("}"):
+        return stripped
+    match = JSON_OBJECT_RE.search(stripped)
+    if match is None:
+        msg = "LLM phrase judge response did not contain a JSON object"
+        raise ValueError(msg)
+    return match.group(0)
+
+
+def optional_text(value: object) -> str | None:
+    """Return stripped text for a nullable JSON value.
+
+    Args:
+        value (object): Decoded JSON value that may or may not be a string.
+
+    Returns:
+        str | None: The stripped string, or ``None`` when it is not a non-empty string.
+    """
+    if not isinstance(value, str):
+        return None
+    stripped = value.strip()
+    return stripped or None
+
+
+def clamped_float(value: object, *, default: float) -> float:
+    """Coerce a JSON number into the 0.0 to 1.0 range.
+
+    Args:
+        value (object): Decoded JSON value that may or may not be a number.
+        default (float): Fallback returned when ``value`` is not numeric.
+
+    Returns:
+        float: The value clamped to ``[0.0, 1.0]``, or ``default`` when non-numeric.
+    """
+    if not isinstance(value, int | float):
+        return default
+    return min(max(float(value), 0.0), 1.0)

python/ebook_search/protected_phrases/matching.py

@@ -0,0 +1,491 @@
+"""Runtime protected-phrase matching and chunk mention indexing."""
+
+from __future__ import annotations
+
+import logging
+from collections import defaultdict
+from typing import TYPE_CHECKING
+
+from sqlalchemy import and_, delete, func, or_, select
+
+from python.ebook_search.protected_phrases.config import get_ignored_phrases
+from python.ebook_search.protected_phrases.models import (
+    ChunkPhraseHit,
+    HydratedPhraseMatch,
+    PhraseLookup,
+    PhraseMatch,
+)
+from python.ebook_search.protected_phrases.text_normalization import tokenize_with_offsets
+from python.orm.richie import (
+    EbookChunk,
+    EbookChunkPhraseMention,
+    EbookPhraseAlias,
+    EbookProtectedPhrase,
+)
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator, Sequence
+
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.protected_phrases.text_normalization import NormalizedToken
+
+logger = logging.getLogger(__name__)
+
+
+async def load_phrase_lookup(
+    session: AsyncSession,
+    config: EbookSearchConfig,
+    *,
+    book_id: int | None = None,
+    series_id: int | None = None,
+) -> PhraseLookup:
+    """Load protected phrases and aliases into RAM lookup maps.
+
+    Args:
+        session (AsyncSession): Active database session.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        book_id (int | None): Optional book scope to restrict loaded phrases.
+        series_id (int | None): Optional series scope to restrict loaded phrases.
+
+    Returns:
+        PhraseLookup: Normalized phrase and alias maps with the token-window bounds to test.
+    """
+    norm_to_ids: defaultdict[str, list[int]] = defaultdict(list)
+    alias_to_ids: defaultdict[str, list[int]] = defaultdict(list)
+    max_tokens = config.phrase_max_tokens
+
+    phrase_statement = select(
+        EbookProtectedPhrase.id,
+        EbookProtectedPhrase.phrase_norm,
+        EbookProtectedPhrase.token_count,
+    )
+    scope_filter = protected_phrase_scope_filter(book_id=book_id, series_id=series_id)
+    if scope_filter is not None:
+        phrase_statement = phrase_statement.where(scope_filter)
+
+    for row in await session.execute(phrase_statement):
+        phrase_id = int(row.id)
+        phrase_norm = str(row.phrase_norm)
+        norm_to_ids[phrase_norm].append(phrase_id)
+        max_tokens = max(max_tokens, int(row.token_count))
+
+    alias_statement = select(
+        EbookPhraseAlias.alias_norm,
+        EbookPhraseAlias.phrase_id,
+    ).join(EbookProtectedPhrase, EbookProtectedPhrase.id == EbookPhraseAlias.phrase_id)
+    if scope_filter is not None:
+        alias_statement = alias_statement.where(scope_filter)
+
+    for row in await session.execute(alias_statement):
+        alias_norm = str(row.alias_norm)
+        alias_to_ids[alias_norm].append(int(row.phrase_id))
+        max_tokens = max(max_tokens, len(alias_norm.split()))
+
+    return PhraseLookup(
+        norm_to_phrase_ids={key: tuple(values) for key, values in norm_to_ids.items()},
+        alias_to_phrase_ids={key: tuple(values) for key, values in alias_to_ids.items()},
+        min_tokens=config.phrase_min_tokens,
+        max_tokens=max_tokens,
+    )
+
+
+def protected_phrase_scope_filter(*, book_id: int | None, series_id: int | None) -> object | None:
+    """Build a SQLAlchemy filter for optional phrase book and series scope.
+
+    Args:
+        book_id (int | None): Optional book scope to include alongside global phrases.
+        series_id (int | None): Optional series scope to include alongside global phrases.
+
+    Returns:
+        object | None: A combined SQLAlchemy filter clause, or ``None`` when no scope is given.
+    """
+    conditions = []
+    if book_id is not None:
+        conditions.append(or_(EbookProtectedPhrase.book_id.is_(None), EbookProtectedPhrase.book_id == book_id))
+    if series_id is not None:
+        conditions.append(or_(EbookProtectedPhrase.series_id.is_(None), EbookProtectedPhrase.series_id == series_id))
+    if not conditions:
+        return None
+    return and_(*conditions)
+
+
+def generate_query_ngrams(
+    tokens_: Sequence[str],
+    min_n: int,
+    max_n: int,
+) -> Iterator[tuple[str, int, int]]:
+    """Generate normalized query windows from longest to shortest.
+
+    Args:
+        tokens_ (Sequence[str]): Normalized query tokens.
+        min_n (int): Smallest window size to yield.
+        max_n (int): Largest window size to yield, capped at the token count.
+
+    Yields:
+        tuple[str, int, int]: Normalized window text with its start and end token indices.
+    """
+    capped_max_n = min(max_n, len(tokens_))
+    for ngram_size in range(capped_max_n, min_n - 1, -1):
+        for start in range(len(tokens_) - ngram_size + 1):
+            end = start + ngram_size
+            phrase_norm = " ".join(tokens_[start:end])
+            if phrase_norm in get_ignored_phrases():
+                continue
+            yield phrase_norm, start, end
+
+
+def detect_phrase_candidates(query_text: str, lookup: PhraseLookup) -> list[PhraseMatch]:
+    """Detect protected phrase windows in a user query using RAM hash lookups.
+
+    Args:
+        query_text (str): User query text to scan.
+        lookup (PhraseLookup): In-memory phrase and alias lookup maps.
+
+    Returns:
+        list[PhraseMatch]: Unhydrated phrase matches found in the query.
+    """
+    return detect_phrase_candidates_from_tokens(tokenize_with_offsets(query_text), lookup)
+
+
+def detect_phrase_candidates_in_text(text: str, lookup: PhraseLookup) -> list[PhraseMatch]:
+    """Detect protected phrase windows in arbitrary text with character offsets.
+
+    Args:
+        text (str): Arbitrary text, such as a chunk, to scan.
+        lookup (PhraseLookup): In-memory phrase and alias lookup maps.
+
+    Returns:
+        list[PhraseMatch]: Unhydrated phrase matches found in the text.
+    """
+    return detect_phrase_candidates_from_tokens(tokenize_with_offsets(text), lookup)
+
+
+def detect_phrase_candidates_from_tokens(tokens_: Sequence[NormalizedToken], lookup: PhraseLookup) -> list[PhraseMatch]:
+    """Detect protected phrase windows from already-normalized tokens.
+
+    Args:
+        tokens_ (Sequence[NormalizedToken]): Normalized tokens with character offsets.
+        lookup (PhraseLookup): In-memory phrase and alias lookup maps.
+
+    Returns:
+        list[PhraseMatch]: Deduplicated unhydrated phrase matches with token and character spans.
+    """
+    matches: list[PhraseMatch] = []
+    seen: set[tuple[int | None, str, int, int]] = set()
+    token_texts = [token.text for token in tokens_]
+    for phrase_norm, start, end in generate_query_ngrams(token_texts, min_n=lookup.min_tokens, max_n=lookup.max_tokens):
+        phrase_ids = lookup.norm_to_phrase_ids.get(phrase_norm, ())
+        alias_ids = lookup.alias_to_phrase_ids.get(phrase_norm, ())
+        for phrase_id in (*phrase_ids, *alias_ids):
+            key = (phrase_id, phrase_norm, start, end)
+            if key in seen:
+                continue
+            seen.add(key)
+            matches.append(
+                PhraseMatch(
+                    phrase_norm=phrase_norm,
+                    phrase_id=phrase_id,
+                    start_token=start,
+                    end_token=end,
+                    token_count=end - start,
+                    start_char=tokens_[start].start_char,
+                    end_char=tokens_[end - 1].end_char,
+                )
+            )
+    return matches
+
+
+async def hydrate_matches(session: AsyncSession, matches: Sequence[PhraseMatch]) -> list[HydratedPhraseMatch]:
+    """Fetch protected phrase metadata for raw phrase matches.
+
+    Args:
+        session (AsyncSession): Active database session.
+        matches (Sequence[PhraseMatch]): Unhydrated matches to enrich.
+
+    Returns:
+        list[HydratedPhraseMatch]: Matches with protected-phrase metadata attached.
+    """
+    if not matches:
+        return []
+
+    phrase_ids = sorted({match.phrase_id for match in matches if match.phrase_id is not None})
+    if not phrase_ids:
+        return []
+
+    rows = {
+        row.id: row
+        for row in await session.scalars(select(EbookProtectedPhrase).where(EbookProtectedPhrase.id.in_(phrase_ids)))
+    }
+    hydrated: list[HydratedPhraseMatch] = []
+    for match in matches:
+        if match.phrase_id is None:
+            continue
+        phrase = rows.get(match.phrase_id)
+        if phrase is None:
+            continue
+        hydrated.append(
+            HydratedPhraseMatch(
+                phrase_id=phrase.id,
+                matched_norm=match.phrase_norm,
+                phrase_text=phrase.phrase_text,
+                phrase_norm=phrase.phrase_norm,
+                canonical_id=phrase.canonical_id,
+                phrase_type=phrase.phrase_type,
+                token_count=match.token_count,
+                confidence=phrase.confidence,
+                importance=phrase.importance,
+                allow_nested=phrase.allow_nested,
+                suppress_children=phrase.suppress_children,
+                start_token=match.start_token,
+                end_token=match.end_token,
+                start_char=match.start_char,
+                end_char=match.end_char,
+                book_id=phrase.book_id,
+                series_id=phrase.series_id,
+            )
+        )
+    return hydrated
+
+
+def overlaps(first: HydratedPhraseMatch, second: HydratedPhraseMatch) -> bool:
+    """Return whether two token spans overlap.
+
+    Args:
+        first (HydratedPhraseMatch): First match to compare.
+        second (HydratedPhraseMatch): Second match to compare.
+
+    Returns:
+        bool: True when the two token spans share at least one token position.
+    """
+    return not (first.end_token <= second.start_token or first.start_token >= second.end_token)
+
+
+def is_inside(child: HydratedPhraseMatch, parent: HydratedPhraseMatch) -> bool:
+    """Return whether one token span is strictly inside another.
+
+    Args:
+        child (HydratedPhraseMatch): Candidate nested match.
+        parent (HydratedPhraseMatch): Candidate enclosing match.
+
+    Returns:
+        bool: True when ``child`` lies within ``parent`` and is not the same span.
+    """
+    return (
+        child.start_token >= parent.start_token
+        and child.end_token <= parent.end_token
+        and (child.start_token, child.end_token, child.phrase_id)
+        != (parent.start_token, parent.end_token, parent.phrase_id)
+    )
+
+
+def rank_match(match: HydratedPhraseMatch) -> tuple[float, float, int]:
+    """Rank phrase matches by importance, confidence, then token count.
+
+    Args:
+        match (HydratedPhraseMatch): Match to build a sort key for.
+
+    Returns:
+        tuple[float, float, int]: A comparable key of importance, confidence, and token count.
+    """
+    return (match.importance, match.confidence, match.token_count)
+
+
+def should_suppress(candidate: HydratedPhraseMatch, kept: HydratedPhraseMatch) -> bool:
+    """Return whether an already-kept match should suppress a candidate.
+
+    Args:
+        candidate (HydratedPhraseMatch): Match being considered for keeping.
+        kept (HydratedPhraseMatch): Match already kept that may suppress the candidate.
+
+    Returns:
+        bool: True when the candidate should be dropped in favor of the kept match.
+    """
+    if not overlaps(candidate, kept):
+        return False
+    if candidate.canonical_id == kept.canonical_id:
+        return rank_match(kept) >= rank_match(candidate)
+    if is_inside(candidate, kept) and kept.suppress_children and not candidate.allow_nested:
+        return True
+    return not candidate.allow_nested and rank_match(kept) > rank_match(candidate)
+
+
+def resolve_overlaps(matches: Sequence[HydratedPhraseMatch]) -> list[HydratedPhraseMatch]:
+    """Resolve overlapping phrase matches without relying only on longest match.
+
+    Args:
+        matches (Sequence[HydratedPhraseMatch]): Hydrated matches that may overlap.
+
+    Returns:
+        list[HydratedPhraseMatch]: The kept, non-suppressed matches.
+    """
+    sorted_matches = sorted(
+        matches,
+        key=lambda match: (match.start_token, -match.token_count, -match.importance, -match.confidence),
+    )
+    kept: list[HydratedPhraseMatch] = []
+    for candidate in sorted_matches:
+        if any(should_suppress(candidate, existing) for existing in kept):
+            continue
+        kept.append(candidate)
+    return kept
+
+
+async def detect_protected_phrases_for_query(
+    session: AsyncSession,
+    query_text: str,
+    config: EbookSearchConfig,
+    *,
+    lookup: PhraseLookup | None = None,
+    book_id: int | None = None,
+    series_id: int | None = None,
+) -> list[HydratedPhraseMatch]:
+    """Run the full online protected-phrase query-detection pipeline.
+
+    Args:
+        session (AsyncSession): Active database session.
+        query_text (str): User query text to detect phrases in.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        lookup (PhraseLookup | None): Optional preloaded lookup; loaded on demand when ``None``.
+        book_id (int | None): Optional book scope for lookup loading.
+        series_id (int | None): Optional series scope for lookup loading.
+
+    Returns:
+        list[HydratedPhraseMatch]: Hydrated, overlap-resolved phrase matches for the query.
+    """
+    active_lookup = (
+        lookup
+        if lookup is not None
+        else await load_phrase_lookup(session, config, book_id=book_id, series_id=series_id)
+    )
+    return resolve_overlaps(await hydrate_matches(session, detect_phrase_candidates(query_text, active_lookup)))
+
+
+async def index_chunk_phrase_mentions_for_book(
+    session: AsyncSession,
+    book_id: int,
+    config: EbookSearchConfig,
+    *,
+    series_id: int | None = None,
+    lookup: PhraseLookup | None = None,
+) -> int:
+    """Rebuild chunk phrase mentions for all chunks in one book.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose chunk mentions are rebuilt.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        series_id (int | None): Optional series scope for lookup loading.
+        lookup (PhraseLookup | None): Optional preloaded lookup; loaded on demand when ``None``.
+
+    Returns:
+        int: Total number of chunk phrase mentions indexed for the book.
+    """
+    active_lookup = (
+        lookup
+        if lookup is not None
+        else await load_phrase_lookup(session, config, book_id=book_id, series_id=series_id)
+    )
+    await session.execute(delete(EbookChunkPhraseMention).where(EbookChunkPhraseMention.book_id == book_id))
+    chunks = await session.scalars(select(EbookChunk).where(EbookChunk.source_id == book_id).order_by(EbookChunk.id))
+    count = 0
+    for chunk in chunks:
+        count += await index_chunk_phrase_mentions(session, chunk, lookup=active_lookup)
+    await session.flush()
+    logger.info("ebook_chunk_phrase_mentions_indexed book_id=%s mentions=%s", book_id, count)
+    return count
+
+
+async def index_chunk_phrase_mentions(session: AsyncSession, chunk: EbookChunk, *, lookup: PhraseLookup) -> int:
+    """Store protected phrase mentions for one chunk.
+
+    Args:
+        session (AsyncSession): Active database session.
+        chunk (EbookChunk): Chunk whose text is scanned for phrase mentions.
+        lookup (PhraseLookup): In-memory phrase and alias lookup maps.
+
+    Returns:
+        int: Number of phrase mentions stored for the chunk.
+    """
+    raw_matches = detect_phrase_candidates_in_text(chunk.text, lookup)
+    hydrated = resolve_overlaps(await hydrate_matches(session, raw_matches))
+    for match in hydrated:
+        session.add(
+            EbookChunkPhraseMention(
+                chunk_id=chunk.id,
+                phrase_id=match.phrase_id,
+                book_id=match.book_id if match.book_id is not None else chunk.source_id,
+                series_id=match.series_id,
+                start_char=match.start_char if match.start_char is not None else 0,
+                end_char=match.end_char,
+            )
+        )
+    return len(hydrated)
+
+
+async def phrase_hits_for_chunks(
+    session: AsyncSession,
+    *,
+    chunk_ids: Sequence[int],
+    phrase_ids: Sequence[int],
+) -> dict[int, tuple[ChunkPhraseHit, ...]]:
+    """Return matched protected phrases with mention counts by chunk id using indexed chunk mentions.
+
+    Args:
+        session (AsyncSession): Active database session.
+        chunk_ids (Sequence[int]): Chunk ids to look up mentions for.
+        phrase_ids (Sequence[int]): Protected phrase ids to restrict the results to.
+
+    Returns:
+        dict[int, tuple[ChunkPhraseHit, ...]]: Phrase hits per chunk id, ordered by mention count.
+    """
+    if not chunk_ids or not phrase_ids:
+        return {}
+
+    mention_count = func.count(EbookChunkPhraseMention.phrase_id).label("mention_count")
+    statement = (
+        select(
+            EbookChunkPhraseMention.chunk_id,
+            EbookProtectedPhrase.id.label("phrase_id"),
+            EbookProtectedPhrase.phrase_text,
+            mention_count,
+        )
+        .join(EbookProtectedPhrase, EbookProtectedPhrase.id == EbookChunkPhraseMention.phrase_id)
+        .where(
+            EbookChunkPhraseMention.chunk_id.in_(chunk_ids),
+            EbookChunkPhraseMention.phrase_id.in_(phrase_ids),
+        )
+        .group_by(EbookChunkPhraseMention.chunk_id, EbookProtectedPhrase.id, EbookProtectedPhrase.phrase_text)
+        .order_by(EbookChunkPhraseMention.chunk_id, mention_count.desc(), EbookProtectedPhrase.phrase_text)
+    )
+    hits: defaultdict[int, list[ChunkPhraseHit]] = defaultdict(list)
+    for row in await session.execute(statement):
+        hits[int(row.chunk_id)].append(
+            ChunkPhraseHit(
+                phrase_id=int(row.phrase_id),
+                phrase_text=str(row.phrase_text),
+                mention_count=int(row.mention_count),
+            )
+        )
+    return {chunk_id: tuple(chunk_hits) for chunk_id, chunk_hits in hits.items()}
+
+
+async def phrase_hit_counts_for_chunks(
+    session: AsyncSession,
+    *,
+    chunk_ids: Sequence[int],
+    phrase_ids: Sequence[int],
+) -> dict[int, int]:
+    """Return phrase-hit counts by chunk id using indexed chunk mentions.
+
+    Args:
+        session (AsyncSession): Active database session.
+        chunk_ids (Sequence[int]): Chunk ids to count mentions for.
+        phrase_ids (Sequence[int]): Protected phrase ids to restrict the counts to.
+
+    Returns:
+        dict[int, int]: Total mention count per chunk id.
+    """
+    hits = await phrase_hits_for_chunks(session, chunk_ids=chunk_ids, phrase_ids=phrase_ids)
+    return {chunk_id: sum(hit.mention_count for hit in chunk_hits) for chunk_id, chunk_hits in hits.items()}

python/ebook_search/protected_phrases/models.py

@@ -0,0 +1,285 @@
+"""Dataclasses shared by protected phrase extraction, judging, matching, and backfills."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping
+
+
+@dataclass(slots=True)
+class PhraseCandidate:
+    """A phrase candidate with merged extraction-source metadata.
+
+    Attributes:
+        phrase_text (str): Display text for the phrase.
+        phrase_norm (str): Normalized phrase used as the merge key.
+        token_count (int): Number of normalized tokens in the phrase.
+        source_raw_ngram (bool): Whether the raw n-gram extractor produced the phrase.
+        source_yake (bool): Whether YAKE keyword extraction produced the phrase.
+        source_spacy_ner (bool): Whether spaCy named-entity recognition produced the phrase.
+        source_spacy_noun_chunk (bool): Whether spaCy noun chunking produced the phrase.
+        source_capitalized (bool): Whether the capitalized-run extractor produced the phrase.
+        source_metadata (bool): Whether book metadata produced the phrase.
+        spacy_label (str | None): spaCy entity label when NER produced the phrase.
+        raw_count (int): Occurrences counted across the book text.
+        chapter_count (int): Number of chapters containing the phrase.
+        yake_score (float | None): Raw YAKE score when available; lower is better.
+        candidate_score (float): Combined pre-judging score.
+        sample_contexts (list[str]): Normalized context snippets around occurrences.
+    """
+
+    phrase_text: str
+    phrase_norm: str
+    token_count: int
+    source_raw_ngram: bool = False
+    source_yake: bool = False
+    source_spacy_ner: bool = False
+    source_spacy_noun_chunk: bool = False
+    source_capitalized: bool = False
+    source_metadata: bool = False
+    spacy_label: str | None = None
+    raw_count: int = 0
+    chapter_count: int = 0
+    yake_score: float | None = None
+    candidate_score: float = 0.0
+    sample_contexts: list[str] = field(default_factory=list)
+
+
+@dataclass(frozen=True, slots=True)
+class LLMJudgment:
+    """A structured phrase judgment returned by the LLM judge.
+
+    Attributes:
+        keep (bool): Whether the judge accepted the phrase for protection.
+        canonical (str | None): Canonical phrase text chosen by the judge.
+        category (str | None): Phrase category such as person, place, or event.
+        aliases (tuple[str, ...]): Alternate surface forms for the phrase.
+        confidence (float): Judge confidence between 0.0 and 1.0.
+        importance (float): Judge importance between 0.0 and 1.0.
+        allow_nested (bool): Whether the phrase may match inside a larger kept match.
+        suppress_children (bool): Whether the phrase suppresses matches nested inside it.
+        reason (str | None): Free-text explanation from the judge.
+    """
+
+    keep: bool
+    canonical: str | None
+    category: str | None
+    aliases: tuple[str, ...]
+    confidence: float
+    importance: float = 0.5
+    allow_nested: bool = False
+    suppress_children: bool = True
+    reason: str | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class PhraseLookup:
+    """In-memory lookup maps used for constant-time phrase-window checks.
+
+    Attributes:
+        norm_to_phrase_ids (Mapping[str, tuple[int, ...]]): Normalized phrase to protected phrase ids.
+        alias_to_phrase_ids (Mapping[str, tuple[int, ...]]): Normalized alias to protected phrase ids.
+        min_tokens (int): Smallest token-window size to test.
+        max_tokens (int): Largest token-window size to test.
+    """
+
+    norm_to_phrase_ids: Mapping[str, tuple[int, ...]]
+    alias_to_phrase_ids: Mapping[str, tuple[int, ...]]
+    min_tokens: int
+    max_tokens: int
+
+
+@dataclass(frozen=True, slots=True)
+class PhraseMatch:
+    """An unhydrated query or chunk phrase match.
+
+    Attributes:
+        phrase_norm (str): Normalized text of the matched window.
+        start_token (int): Index of the first matched token.
+        end_token (int): Index one past the last matched token.
+        token_count (int): Number of tokens in the match.
+        phrase_id (int | None): Matched protected phrase id when known.
+        start_char (int | None): Start character offset in the source text.
+        end_char (int | None): End character offset in the source text.
+    """
+
+    phrase_norm: str
+    start_token: int
+    end_token: int
+    token_count: int
+    phrase_id: int | None = None
+    start_char: int | None = None
+    end_char: int | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class HydratedPhraseMatch:
+    """A phrase match with protected-phrase metadata attached.
+
+    Attributes:
+        phrase_id (int): Protected phrase id.
+        matched_norm (str): Normalized window text that matched.
+        phrase_text (str): Display text of the protected phrase.
+        phrase_norm (str): Normalized text of the protected phrase.
+        canonical_id (str): Deterministic ``category:slug`` identifier.
+        phrase_type (str | None): Phrase category.
+        token_count (int): Number of tokens in the match.
+        confidence (float): Stored judge confidence.
+        importance (float): Stored judge importance.
+        allow_nested (bool): Whether the phrase may match inside a larger kept match.
+        suppress_children (bool): Whether the phrase suppresses matches nested inside it.
+        start_token (int): Index of the first matched token.
+        end_token (int): Index one past the last matched token.
+        start_char (int | None): Start character offset in the source text.
+        end_char (int | None): End character offset in the source text.
+        book_id (int | None): Book scope of the phrase.
+        series_id (int | None): Series scope of the phrase.
+    """
+
+    phrase_id: int
+    matched_norm: str
+    phrase_text: str
+    phrase_norm: str
+    canonical_id: str
+    phrase_type: str | None
+    token_count: int
+    confidence: float
+    importance: float
+    allow_nested: bool
+    suppress_children: bool
+    start_token: int
+    end_token: int
+    start_char: int | None = None
+    end_char: int | None = None
+    book_id: int | None = None
+    series_id: int | None = None
+
+
+@dataclass(frozen=True, slots=True)
+class ChunkPhraseHit:
+    """One protected phrase with its mention count inside one retrieved chunk.
+
+    Attributes:
+        phrase_id (int): Protected phrase id.
+        phrase_text (str): Display text of the protected phrase.
+        mention_count (int): Indexed mentions of the phrase in the chunk.
+    """
+
+    phrase_id: int
+    phrase_text: str
+    mention_count: int
+
+
+@dataclass(frozen=True, slots=True)
+class PhraseCandidateGenerationResult:
+    """Summary of candidate phrase extraction for indexed books.
+
+    Attributes:
+        books_seen (int): Indexed books examined.
+        books_built (int): Books that had candidates generated and committed.
+        candidate_phrases (int): Candidate phrases stored across all books.
+    """
+
+    books_seen: int
+    books_built: int
+    candidate_phrases: int
+
+
+@dataclass(frozen=True, slots=True)
+class CorpusPhraseStats:
+    """Corpus-wide candidate and protected phrase counts for the admin page.
+
+    Attributes:
+        total_books (int): Indexed books in the corpus.
+        books_with_candidates (int): Books that have candidate phrases generated.
+        books_fully_judged (int): Books with candidates where every candidate has been judged.
+        candidate_phrases (int): Candidate phrases stored across all books.
+        judged_candidates (int): Candidate phrases that have been LLM judged.
+        unjudged_candidates (int): Candidate phrases still waiting for judgment.
+        protected_phrases (int): Protected phrases promoted across all books.
+    """
+
+    total_books: int
+    books_with_candidates: int
+    books_fully_judged: int
+    candidate_phrases: int
+    judged_candidates: int
+    unjudged_candidates: int
+    protected_phrases: int
+
+
+@dataclass(frozen=True, slots=True)
+class PhraseJudgmentBackfillResult:
+    """Summary of LLM judging for stored candidate phrases.
+
+    Attributes:
+        books_seen (int): Indexed books examined.
+        books_judged (int): Books with judgments committed.
+        books_failed (int): Books rolled back after an error.
+        candidates_judged (int): Candidate phrases sent to the LLM judge.
+        protected_phrases (int): Protected phrases promoted from candidates.
+        phrase_mentions (int): Chunk phrase mentions indexed across all books.
+    """
+
+    books_seen: int
+    books_judged: int
+    books_failed: int
+    candidates_judged: int
+    protected_phrases: int
+    phrase_mentions: int
+
+
+@dataclass(frozen=True, slots=True)
+class BookJudgmentResult:
+    """Outcome of judging one book's candidate phrases.
+
+    Attributes:
+        judged (int): Candidate phrases sent to the LLM judge.
+        protected (int): Protected phrases promoted from candidates.
+        mentions (int): Chunk phrase mentions indexed for the book.
+        committed (bool): Whether the book's judgments were committed.
+        failed (bool): Whether the book was rolled back after an error.
+    """
+
+    judged: int = 0
+    protected: int = 0
+    mentions: int = 0
+    committed: bool = False
+    failed: bool = False
+
+
+@dataclass(frozen=True, slots=True)
+class BookCandidateResult:
+    """Outcome of generating one book's candidate phrases.
+
+    Attributes:
+        candidates (int): Candidate phrases stored for the book.
+        built (bool): Whether candidate generation was committed.
+    """
+
+    candidates: int = 0
+    built: bool = False
+
+
+@dataclass(frozen=True, slots=True)
+class PhraseRecalculationResult:
+    """Summary of phrase cleanup and candidate regeneration for one book.
+
+    Attributes:
+        book_id (int): Book the recalculation ran against.
+        deleted_candidates (int): Candidate phrase rows deleted.
+        deleted_protected_phrases (int): Protected phrase rows deleted.
+        deleted_aliases (int): Phrase alias rows deleted.
+        deleted_mentions (int): Chunk phrase mention rows deleted.
+        candidate_phrases (int): Candidate phrases regenerated after cleanup.
+    """
+
+    book_id: int
+    deleted_candidates: int
+    deleted_protected_phrases: int
+    deleted_aliases: int
+    deleted_mentions: int
+    candidate_phrases: int

python/ebook_search/protected_phrases/pool.py

@@ -0,0 +1,101 @@
+"""Process pool for offloading CPU-bound phrase extraction off the request thread.
+
+Phrase extraction is pure-Python CPU work (n-gram sliding, YAKE), so running it inline in a
+sync request handler serializes concurrent recalculations behind the GIL. Submitting it to a
+``ProcessPoolExecutor`` lets concurrent extractions run in parallel across cores instead. A
+``spawn`` context is used so workers do not inherit the parent's database engine, connections,
+or server threads.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import multiprocessing
+import os
+from concurrent.futures import ProcessPoolExecutor
+from threading import Lock
+from typing import TYPE_CHECKING
+
+from python.ebook_search.protected_phrases.extraction import extract_phrase_candidates_for_book
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping, Sequence
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.protected_phrases.models import PhraseCandidate
+
+logger = logging.getLogger(__name__)
+
+
+class _ExtractionPool:
+    """Lazily created process-wide extraction pool and the lock guarding it."""
+
+    def __init__(self) -> None:
+        self.lock = Lock()
+        self.pool: ProcessPoolExecutor | None = None
+
+
+_extraction_pool = _ExtractionPool()
+
+
+def get_extraction_pool(max_workers: int) -> ProcessPoolExecutor:
+    """Return the shared extraction process pool, creating it on first use.
+
+    Args:
+        max_workers (int): Desired worker count; values below 1 fall back to the CPU count.
+
+    Returns:
+        ProcessPoolExecutor: The shared pool for phrase extraction.
+    """
+    with _extraction_pool.lock:
+        if _extraction_pool.pool is None:
+            workers = max_workers if max_workers > 0 else (os.cpu_count() or 1)
+            _extraction_pool.pool = ProcessPoolExecutor(
+                max_workers=workers,
+                mp_context=multiprocessing.get_context("spawn"),
+            )
+            logger.info("ebook_phrase_extraction_pool_started workers=%s", workers)
+        return _extraction_pool.pool
+
+
+def shutdown_extraction_pool() -> None:
+    """Shut down the shared extraction pool if it was started."""
+    with _extraction_pool.lock:
+        if _extraction_pool.pool is not None:
+            _extraction_pool.pool.shutdown(wait=False, cancel_futures=True)
+            _extraction_pool.pool = None
+            logger.info("ebook_phrase_extraction_pool_shutdown")
+
+
+async def extract_phrase_candidates_in_pool(
+    book_text: str,
+    chapters: Sequence[str],
+    config: EbookSearchConfig,
+    *,
+    metadata: Mapping[str, object] | None,
+) -> list[PhraseCandidate]:
+    """Run book phrase extraction in a worker process and await the result.
+
+    Only the CPU-bound extraction runs in the worker; the caller keeps all database work in the
+    request process. The spaCy pipeline is not supported here because it is not picklable, so
+    this always runs the non-spaCy extraction path.
+
+    Args:
+        book_text (str): Full book text used for extraction.
+        chapters (Sequence[str]): Chapter-like text blocks used for frequency counts.
+        config (EbookSearchConfig): Runtime phrase-tuning settings.
+        metadata (Mapping[str, object] | None): Optional book metadata used as a candidate source.
+
+    Returns:
+        list[PhraseCandidate]: Scored candidates sorted best-first and capped per book.
+    """
+    pool = get_extraction_pool(config.protected_phrase_extraction_workers)
+    future = pool.submit(
+        extract_phrase_candidates_for_book,
+        book_text,
+        list(chapters),
+        config,
+        metadata=dict(metadata) if metadata is not None else None,
+    )
+    return await asyncio.wrap_future(future)

python/ebook_search/protected_phrases/store.py

@@ -0,0 +1,700 @@
+"""Database persistence for candidate and protected phrase rows."""
+
+from __future__ import annotations
+
+import logging
+import re
+from typing import TYPE_CHECKING
+
+from sqlalchemy import delete, func, or_, select
+from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.dialects.sqlite import insert as sqlite_insert
+
+from python.ebook_search.protected_phrases.extraction import minimum_candidate_raw_count
+from python.ebook_search.protected_phrases.models import (
+    CorpusPhraseStats,
+    PhraseCandidate,
+    PhraseRecalculationResult,
+)
+from python.ebook_search.protected_phrases.text_normalization import normalize_text
+from python.orm.richie import (
+    EbookCandidatePhrase,
+    EbookChunk,
+    EbookChunkPhraseMention,
+    EbookPhraseAlias,
+    EbookProtectedPhrase,
+    EbookSource,
+)
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from sqlalchemy.dialects.postgresql.dml import Insert as PostgresInsert
+    from sqlalchemy.dialects.sqlite.dml import Insert as SqliteInsert
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    from python.ebook_search.config import EbookSearchConfig
+    from python.ebook_search.protected_phrases.models import LLMJudgment
+    from python.orm.richie.base import TableBase
+
+logger = logging.getLogger(__name__)
+
+
+def dialect_insert(session: AsyncSession, table: type[TableBase]) -> PostgresInsert | SqliteInsert:
+    """Return a dialect-specific INSERT construct that supports ``ON CONFLICT DO UPDATE``.
+
+    Production runs on PostgreSQL while tests run on SQLite; both support upserts with
+    compatible SQLAlchemy constructs, so the correct one is chosen from the bound dialect.
+
+    Args:
+        session (AsyncSession): Active database session whose bind selects the dialect.
+        table (type[TableBase]): Mapped table to insert into.
+
+    Returns:
+        PostgresInsert | SqliteInsert: A dialect insert exposing ``on_conflict_do_update``.
+    """
+    if session.get_bind().dialect.name == "sqlite":
+        return sqlite_insert(table)
+    return pg_insert(table)
+
+
+async def load_book_text(session: AsyncSession, book_id: int) -> str:
+    """Load a book's indexed chunk text as one string for phrase extraction.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose chunk text is loaded.
+
+    Returns:
+        str: The book's chunk text joined into a single string.
+    """
+    texts = await session.scalars(
+        select(EbookChunk.text).where(EbookChunk.source_id == book_id).order_by(EbookChunk.chunk_index)
+    )
+    return "\n\n".join(stripped for text in texts if (stripped := text.strip()))
+
+
+async def load_book_chapter_texts(session: AsyncSession, book_id: int) -> list[str]:
+    """Reconstruct chapter-like text blocks from indexed chunks for phrase extraction.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose chunks are grouped into chapters.
+
+    Returns:
+        list[str]: Non-empty chapter-like text blocks in chunk order.
+    """
+    rows = await session.execute(
+        select(EbookChunk.chapter_id, EbookChunk.text)
+        .where(EbookChunk.source_id == book_id)
+        .order_by(EbookChunk.chunk_index)
+    )
+    chapters: list[str] = []
+    current_chapter_id: int | None = None
+    current_parts: list[str] = []
+    have_current = False
+
+    for chapter_id, text in rows:
+        if have_current and chapter_id != current_chapter_id:
+            chapter_text = "\n\n".join(current_parts).strip()
+            if chapter_text:
+                chapters.append(chapter_text)
+            current_parts = []
+        current_chapter_id = chapter_id
+        current_parts.append(str(text))
+        have_current = True
+
+    if current_parts:
+        chapter_text = "\n\n".join(current_parts).strip()
+        if chapter_text:
+            chapters.append(chapter_text)
+    return chapters
+
+
+def metadata_for_source(source: EbookSource) -> dict[str, object | None]:
+    """Return phrase extraction metadata for one indexed source.
+
+    Args:
+        source (EbookSource): Indexed source to read metadata from.
+
+    Returns:
+        dict[str, object | None]: Title, author, language, publisher, and identifier values.
+    """
+    return {
+        "title": source.title,
+        "author": source.author,
+        "language": source.language,
+        "publisher": source.publisher,
+        "identifier": source.identifier,
+    }
+
+
+async def metadata_for_source_id(session: AsyncSession, source_id: int) -> dict[str, object | None]:
+    """Return phrase extraction metadata for one indexed source by id.
+
+    Args:
+        session (AsyncSession): Active database session.
+        source_id (int): Id of the indexed source to read metadata from.
+
+    Returns:
+        dict[str, object | None]: Title, author, language, publisher, and identifier values.
+
+    Raises:
+        ValueError: If no source exists with the given id.
+    """
+    source = await session.get(EbookSource, source_id)
+    if source is None:
+        msg = f"No indexed source with id {source_id}"
+        raise ValueError(msg)
+    return metadata_for_source(source)
+
+
+async def count_protected_phrases(session: AsyncSession, book_id: int) -> int:
+    """Count stored protected phrases for one book.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose protected phrases are counted.
+
+    Returns:
+        int: Number of protected phrases stored for the book.
+    """
+    return (
+        await session.scalars(
+            select(func.count(EbookProtectedPhrase.id)).where(EbookProtectedPhrase.book_id == book_id)
+        )
+    ).one()
+
+
+async def count_unjudged_candidates(session: AsyncSession, book_id: int, config: EbookSearchConfig) -> int:
+    """Count storable candidate rows for a book that have not yet been judged.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose unjudged candidates are counted.
+        config (EbookSearchConfig): Runtime phrase-tuning settings supplying storage thresholds.
+
+    Returns:
+        int: Number of storable, unjudged candidate rows for the book.
+    """
+    return (
+        await session.scalars(
+            select(func.count(EbookCandidatePhrase.id)).where(
+                EbookCandidatePhrase.book_id == book_id,
+                EbookCandidatePhrase.llm_judged.is_(False),
+                EbookCandidatePhrase.token_count >= config.phrase_min_tokens,
+                EbookCandidatePhrase.raw_count >= minimum_candidate_raw_count(config),
+            )
+        )
+    ).one()
+
+
+async def corpus_phrase_stats(session: AsyncSession) -> CorpusPhraseStats:
+    """Summarize candidate and protected phrase coverage across the whole corpus.
+
+    Args:
+        session (AsyncSession): Active database session.
+
+    Returns:
+        CorpusPhraseStats: Corpus-wide phrase counts and per-book coverage counts.
+    """
+    total_books = (await session.scalars(select(func.count(EbookSource.id)))).one()
+    candidate_phrases, judged_candidates, books_with_candidates, books_with_unjudged = (
+        await session.execute(
+            select(
+                func.count(EbookCandidatePhrase.id),
+                func.count(EbookCandidatePhrase.id).filter(EbookCandidatePhrase.llm_judged.is_(True)),
+                func.count(func.distinct(EbookCandidatePhrase.book_id)),
+                func.count(func.distinct(EbookCandidatePhrase.book_id)).filter(
+                    EbookCandidatePhrase.llm_judged.is_(False)
+                ),
+            )
+        )
+    ).one()
+    protected_phrases = (await session.scalars(select(func.count(EbookProtectedPhrase.id)))).one()
+    return CorpusPhraseStats(
+        total_books=total_books,
+        books_with_candidates=books_with_candidates,
+        books_fully_judged=books_with_candidates - books_with_unjudged,
+        candidate_phrases=candidate_phrases,
+        judged_candidates=judged_candidates,
+        unjudged_candidates=candidate_phrases - judged_candidates,
+        protected_phrases=protected_phrases,
+    )
+
+
+async def book_ids_pending_first_judgment(session: AsyncSession) -> list[int]:
+    """Return books that have candidate phrases but no judged candidates yet.
+
+    Args:
+        session (AsyncSession): Active database session.
+
+    Returns:
+        list[int]: Book ids with candidates where judging has never run, ordered by id.
+    """
+    judged_books = select(EbookCandidatePhrase.book_id).where(EbookCandidatePhrase.llm_judged.is_(True)).distinct()
+    return list(
+        (
+            await session.scalars(
+                select(EbookCandidatePhrase.book_id)
+                .where(EbookCandidatePhrase.book_id.not_in(judged_books))
+                .distinct()
+                .order_by(EbookCandidatePhrase.book_id)
+            )
+        ).all()
+    )
+
+
+async def load_candidates_for_judgment(
+    session: AsyncSession,
+    book_id: int,
+    config: EbookSearchConfig,
+) -> Sequence[EbookCandidatePhrase]:
+    """Load every storable unjudged candidate row for a book.
+
+    Rows may have been stored before the current junk filters and score weights existed, so
+    callers re-check :func:`is_junk_phrase` and rescore before selecting what to judge.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose candidates are loaded.
+        config (EbookSearchConfig): Runtime phrase-tuning settings supplying storage thresholds.
+
+    Returns:
+        Sequence[EbookCandidatePhrase]: Storable, unjudged candidate rows ordered by stored score.
+    """
+    query = (
+        select(EbookCandidatePhrase)
+        .where(
+            EbookCandidatePhrase.book_id == book_id,
+            EbookCandidatePhrase.llm_judged.is_(False),
+            EbookCandidatePhrase.token_count >= config.phrase_min_tokens,
+            EbookCandidatePhrase.raw_count >= minimum_candidate_raw_count(config),
+        )
+        .order_by(
+            EbookCandidatePhrase.candidate_score.desc(),
+            EbookCandidatePhrase.raw_count.desc(),
+            EbookCandidatePhrase.id,
+        )
+    )
+    return (await session.scalars(query)).all()
+
+
+def phrase_candidate_from_row(row: EbookCandidatePhrase) -> PhraseCandidate:
+    """Recreate an in-memory candidate from a persisted candidate row.
+
+    Args:
+        row (EbookCandidatePhrase): Stored candidate row to convert.
+
+    Returns:
+        PhraseCandidate: An in-memory candidate mirroring the row's fields.
+    """
+    return PhraseCandidate(
+        phrase_text=row.phrase_text,
+        phrase_norm=row.phrase_norm,
+        token_count=row.token_count,
+        source_raw_ngram=row.source_raw_ngram,
+        source_yake=row.source_yake,
+        source_spacy_ner=row.source_spacy_ner,
+        source_spacy_noun_chunk=row.source_spacy_noun_chunk,
+        source_capitalized=row.source_capitalized,
+        source_metadata=row.source_metadata,
+        spacy_label=row.spacy_label,
+        raw_count=row.raw_count,
+        chapter_count=row.chapter_count,
+        yake_score=row.yake_score,
+        candidate_score=row.candidate_score,
+        sample_contexts=row.sample_contexts or [],
+    )
+
+
+def candidate_row_values(
+    book_id: int,
+    series_id: int | None,
+    candidate: PhraseCandidate,
+    *,
+    judgment: LLMJudgment | None,
+) -> dict[str, object]:
+    """Build the column values for one candidate phrase upsert.
+
+    Args:
+        book_id (int): Book the candidate belongs to.
+        series_id (int | None): Series scope stored on the row.
+        candidate (PhraseCandidate): Candidate whose fields are written to the row.
+        judgment (LLMJudgment | None): Judgment to record, or ``None`` to leave the row unjudged.
+
+    Returns:
+        dict[str, object]: Column values keyed by column name.
+    """
+    values: dict[str, object] = {
+        "book_id": book_id,
+        "phrase_norm": candidate.phrase_norm,
+        "series_id": series_id,
+        "phrase_text": candidate.phrase_text,
+        "token_count": candidate.token_count,
+        "source_raw_ngram": candidate.source_raw_ngram,
+        "source_yake": candidate.source_yake,
+        "source_spacy_ner": candidate.source_spacy_ner,
+        "source_spacy_noun_chunk": candidate.source_spacy_noun_chunk,
+        "source_capitalized": candidate.source_capitalized,
+        "source_metadata": candidate.source_metadata,
+        "spacy_label": candidate.spacy_label,
+        "raw_count": candidate.raw_count,
+        "chapter_count": candidate.chapter_count,
+        "yake_score": candidate.yake_score,
+        "candidate_score": candidate.candidate_score,
+        "llm_judged": judgment is not None,
+    }
+    if candidate.sample_contexts:
+        values["sample_contexts"] = list(candidate.sample_contexts)
+    if judgment is not None:
+        values.update(
+            llm_keep=judgment.keep,
+            llm_confidence=judgment.confidence,
+            llm_category=judgment.category,
+            llm_reason=judgment.reason,
+        )
+    return values
+
+
+async def save_candidate_to_db(
+    session: AsyncSession,
+    book_id: int,
+    series_id: int | None,
+    candidate: PhraseCandidate,
+    *,
+    judgment: LLMJudgment | None,
+) -> EbookCandidatePhrase:
+    """Insert or update one candidate phrase row.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book the candidate belongs to.
+        series_id (int | None): Series scope stored on the row.
+        candidate (PhraseCandidate): Candidate whose fields are written to the row.
+        judgment (LLMJudgment | None): Judgment to record, or ``None`` to leave the row unjudged.
+
+    Returns:
+        EbookCandidatePhrase: The inserted or updated candidate row.
+    """
+    values = candidate_row_values(book_id, series_id, candidate, judgment=judgment)
+
+    # Preserve an existing judgment when this call is only refreshing candidate fields.
+    skip_update = {"book_id", "phrase_norm"}
+    if judgment is None:
+        skip_update.add("llm_judged")
+    insert_statement = dialect_insert(session, EbookCandidatePhrase).values(**values)
+    statement = insert_statement.on_conflict_do_update(
+        index_elements=["book_id", "phrase_norm"],
+        set_={column: insert_statement.excluded[column] for column in values if column not in skip_update},
+    ).returning(EbookCandidatePhrase)
+    return (await session.scalars(statement, execution_options={"populate_existing": True})).one()
+
+
+BULK_CANDIDATE_UPSERT_CHUNK = 1000
+
+
+async def bulk_upsert_unjudged_candidates(
+    session: AsyncSession,
+    book_id: int,
+    series_id: int | None,
+    candidates: Sequence[PhraseCandidate],
+) -> int:
+    """Insert or update many freshly extracted candidate rows in chunked multi-row upserts.
+
+    Saving one row per statement costs one database round trip per candidate, which dominated
+    generation time for full books, so candidates are written ``BULK_CANDIDATE_UPSERT_CHUNK``
+    rows per statement instead. Existing judgments and sample contexts are never overwritten:
+    fresh extractions carry no contexts, and ``llm_judged`` plus the ``llm_*`` columns are left
+    out of the conflict update. Candidates must have unique ``phrase_norm`` values, as produced
+    by extraction, since one multi-row upsert cannot touch the same row twice.
+
+    Args:
+        session (Session): Active database session.
+        book_id (int): Book the candidates belong to.
+        series_id (int | None): Series scope stored on the rows.
+        candidates (Sequence[PhraseCandidate]): Freshly extracted candidates to persist.
+
+    Returns:
+        int: Number of candidate rows written.
+    """
+    values = [
+        candidate_row_values(book_id, series_id, candidate, judgment=None)
+        for candidate in candidates
+        if not candidate.sample_contexts
+    ]
+    if len(values) != len(candidates):
+        msg = "bulk_upsert_unjudged_candidates only accepts freshly extracted candidates without sample contexts"
+        raise ValueError(msg)
+    skip_update = {"book_id", "phrase_norm", "llm_judged"}
+    for chunk_start in range(0, len(values), BULK_CANDIDATE_UPSERT_CHUNK):
+        chunk = values[chunk_start : chunk_start + BULK_CANDIDATE_UPSERT_CHUNK]
+        insert_statement = dialect_insert(session, EbookCandidatePhrase).values(chunk)
+        statement = insert_statement.on_conflict_do_update(
+            index_elements=["book_id", "phrase_norm"],
+            set_={column: insert_statement.excluded[column] for column in chunk[0] if column not in skip_update},
+        )
+        await session.execute(statement)
+    return len(values)
+
+
+def new_candidate_row(book_id: int, series_id: int | None, candidate: PhraseCandidate) -> EbookCandidatePhrase:
+    """Build a fresh unjudged candidate row without checking for an existing one.
+
+    Unlike :func:`save_candidate_to_db`, this does no lookup, so it is only safe when the caller
+    guarantees there is no existing row for ``(book_id, candidate.phrase_norm)`` — for example
+    right after :func:`delete_phrase_data_for_book` has cleared the book.
+
+    Args:
+        book_id (int): Book the candidate belongs to.
+        series_id (int | None): Series scope stored on the row.
+        candidate (PhraseCandidate): Candidate whose fields are written to the row.
+
+    Returns:
+        EbookCandidatePhrase: A new, unattached candidate row.
+    """
+    row = EbookCandidatePhrase(book_id=book_id, phrase_norm=candidate.phrase_norm)
+    row.llm_judged = False
+    row.series_id = series_id
+    row.phrase_text = candidate.phrase_text
+    row.token_count = candidate.token_count
+    row.source_raw_ngram = candidate.source_raw_ngram
+    row.source_yake = candidate.source_yake
+    row.source_spacy_ner = candidate.source_spacy_ner
+    row.source_spacy_noun_chunk = candidate.source_spacy_noun_chunk
+    row.source_capitalized = candidate.source_capitalized
+    row.source_metadata = candidate.source_metadata
+    row.spacy_label = candidate.spacy_label
+    row.raw_count = candidate.raw_count
+    row.chapter_count = candidate.chapter_count
+    row.yake_score = candidate.yake_score
+    row.candidate_score = candidate.candidate_score
+    if candidate.sample_contexts:
+        row.sample_contexts = list(candidate.sample_contexts)
+    return row
+
+
+async def upsert_protected_phrase(
+    session: AsyncSession,
+    book_id: int,
+    series_id: int | None,
+    candidate: PhraseCandidate,
+    judgment: LLMJudgment,
+    source_candidate: EbookCandidatePhrase,
+) -> EbookProtectedPhrase:
+    """Insert or update one accepted protected phrase and its aliases.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book the protected phrase belongs to.
+        series_id (int | None): Series scope stored on the phrase.
+        candidate (PhraseCandidate): Candidate the phrase was promoted from.
+        judgment (LLMJudgment): Accepted judgment supplying canonical text, category, and aliases.
+        source_candidate (EbookCandidatePhrase): Candidate row the phrase was promoted from.
+
+    Returns:
+        EbookProtectedPhrase: The inserted or updated protected phrase row.
+
+    Raises:
+        ValueError: If the chosen phrase text normalizes to empty.
+    """
+    phrase_text = judgment.canonical or candidate.phrase_text
+    phrase_norm = normalize_text(phrase_text)
+    if not phrase_norm:
+        msg = f"Protected phrase normalized to empty text: {phrase_text!r}"
+        raise ValueError(msg)
+
+    values = {
+        "book_id": book_id,
+        "phrase_norm": phrase_norm,
+        "series_id": series_id,
+        "phrase_text": phrase_text,
+        "canonical_id": make_canonical_id(judgment, phrase_norm),
+        "phrase_type": judgment.category,
+        "token_count": len(phrase_norm.split()),
+        "confidence": judgment.confidence,
+        "importance": judgment.importance,
+        "allow_nested": judgment.allow_nested,
+        "suppress_children": judgment.suppress_children,
+        "source_candidate_id": source_candidate.id,
+    }
+    insert_statement = dialect_insert(session, EbookProtectedPhrase).values(**values)
+    statement = insert_statement.on_conflict_do_update(
+        index_elements=["book_id", "phrase_norm"],
+        set_={
+            column: insert_statement.excluded[column] for column in values if column not in {"book_id", "phrase_norm"}
+        },
+    ).returning(EbookProtectedPhrase)
+    row = (await session.scalars(statement, execution_options={"populate_existing": True})).one()
+
+    for alias_text in judgment.aliases:
+        await upsert_phrase_alias(session, row, alias_text)
+    return row
+
+
+async def upsert_phrase_alias(
+    session: AsyncSession,
+    phrase: EbookProtectedPhrase,
+    alias_text: str,
+) -> EbookPhraseAlias | None:
+    """Insert or update one protected phrase alias.
+
+    Args:
+        session (AsyncSession): Active database session.
+        phrase (EbookProtectedPhrase): Protected phrase the alias points to.
+        alias_text (str): Alias surface form to store.
+
+    Returns:
+        EbookPhraseAlias | None: The alias row, or ``None`` when the alias is empty or equals the phrase.
+    """
+    alias_norm = normalize_text(alias_text)
+    if not alias_norm or alias_norm == phrase.phrase_norm:
+        return None
+
+    insert_statement = dialect_insert(session, EbookPhraseAlias).values(
+        phrase_id=phrase.id,
+        alias_norm=alias_norm,
+        alias_text=alias_text,
+        confidence=1.0,
+    )
+    statement = insert_statement.on_conflict_do_update(
+        index_elements=["phrase_id", "alias_norm"],
+        set_={
+            "alias_text": insert_statement.excluded.alias_text,
+            "confidence": insert_statement.excluded.confidence,
+        },
+    ).returning(EbookPhraseAlias)
+    return (await session.scalars(statement, execution_options={"populate_existing": True})).one()
+
+
+def make_canonical_id(judgment: LLMJudgment, phrase_norm: str) -> str:
+    """Create a deterministic canonical id from a judgment category and phrase.
+
+    Args:
+        judgment (LLMJudgment): Judgment supplying the phrase category.
+        phrase_norm (str): Normalized phrase text to slugify.
+
+    Returns:
+        str: A ``category:slug`` canonical identifier.
+    """
+    category = slugify_identifier(judgment.category or "phrase")
+    phrase_slug = slugify_identifier(phrase_norm)
+    return f"{category}:{phrase_slug}"
+
+
+def slugify_identifier(value: str) -> str:
+    """Normalize text for use inside a canonical id.
+
+    Args:
+        value (str): Text to slugify.
+
+    Returns:
+        str: A lowercase underscore slug, or ``"unknown"`` when empty.
+    """
+    slug = re.sub(r"[^a-z0-9]+", "_", normalize_text(value).replace("'", ""))
+    return slug.strip("_") or "unknown"
+
+
+async def prune_unstorable_unjudged_candidate_phrases(
+    session: AsyncSession,
+    book_id: int,
+    config: EbookSearchConfig,
+) -> int:
+    """Delete old unjudged candidate rows that no longer satisfy storage filters.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose stale candidates are pruned.
+        config (EbookSearchConfig): Runtime phrase-tuning settings supplying storage thresholds.
+
+    Returns:
+        int: Number of candidate rows deleted.
+    """
+    deleted = rowcount(
+        await session.execute(
+            delete(EbookCandidatePhrase).where(
+                EbookCandidatePhrase.book_id == book_id,
+                EbookCandidatePhrase.llm_judged.is_(False),
+                or_(
+                    EbookCandidatePhrase.token_count < config.phrase_min_tokens,
+                    EbookCandidatePhrase.raw_count < minimum_candidate_raw_count(config),
+                ),
+            )
+        )
+    )
+    if deleted:
+        logger.info(
+            "ebook_candidate_phrase_unstorable_pruned book_id=%s deleted=%s min_tokens=%s min_uses=%s",
+            book_id,
+            deleted,
+            config.phrase_min_tokens,
+            minimum_candidate_raw_count(config),
+        )
+    return deleted
+
+
+async def delete_phrase_data_for_book(session: AsyncSession, book_id: int) -> PhraseRecalculationResult:
+    """Delete all candidate, protected, alias, and mention phrase data for one book.
+
+    Args:
+        session (AsyncSession): Active database session.
+        book_id (int): Book whose phrase data is deleted.
+
+    Returns:
+        PhraseRecalculationResult: Deleted-row counts with ``candidate_phrases`` set to 0.
+    """
+    protected_ids = (
+        await session.scalars(select(EbookProtectedPhrase.id).where(EbookProtectedPhrase.book_id == book_id))
+    ).all()
+    deleted_aliases = 0
+    if protected_ids:
+        deleted_aliases = rowcount(
+            await session.execute(delete(EbookPhraseAlias).where(EbookPhraseAlias.phrase_id.in_(protected_ids)))
+        )
+
+    deleted_mentions = rowcount(
+        await session.execute(delete(EbookChunkPhraseMention).where(EbookChunkPhraseMention.book_id == book_id))
+    )
+    if protected_ids:
+        deleted_mentions += rowcount(
+            await session.execute(
+                delete(EbookChunkPhraseMention).where(EbookChunkPhraseMention.phrase_id.in_(protected_ids))
+            )
+        )
+
+    deleted_protected = rowcount(
+        await session.execute(delete(EbookProtectedPhrase).where(EbookProtectedPhrase.book_id == book_id))
+    )
+    deleted_candidates = rowcount(
+        await session.execute(delete(EbookCandidatePhrase).where(EbookCandidatePhrase.book_id == book_id))
+    )
+    await session.flush()
+    logger.info(
+        "ebook_candidate_phrase_data_deleted book_id=%s candidates=%s protected=%s aliases=%s mentions=%s",
+        book_id,
+        deleted_candidates,
+        deleted_protected,
+        deleted_aliases,
+        deleted_mentions,
+    )
+    return PhraseRecalculationResult(
+        book_id=book_id,
+        deleted_candidates=deleted_candidates,
+        deleted_protected_phrases=deleted_protected,
+        deleted_aliases=deleted_aliases,
+        deleted_mentions=deleted_mentions,
+        candidate_phrases=0,
+    )
+
+
+def rowcount(result: object) -> int:
+    """Return a safe integer rowcount from a SQLAlchemy execution result.
+
+    Args:
+        result (object): SQLAlchemy execution result that may expose ``rowcount``.
+
+    Returns:
+        int: The result's rowcount, or 0 when it is missing or negative.
+    """
+    count = getattr(result, "rowcount", 0)
+    return int(count if count is not None and count >= 0 else 0)

python/ebook_search/protected_phrases/text_normalization.py

@@ -0,0 +1,91 @@
+"""Protected phrase extraction, storage, and runtime matching."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+JSON_OBJECT_RE = re.compile(r"\{.*\}", re.DOTALL)
+
+
+@dataclass(frozen=True, slots=True)
+class NormalizedToken:
+    """A normalized token plus its source character span."""
+
+    text: str
+    start_char: int
+    end_char: int
+
+
+def normalize_text(text: str) -> str:
+    """Normalize text for phrase storage and lookup.
+
+    Args:
+        text (str): Raw text to normalize.
+
+    Returns:
+        str: Normalized tokens joined by single spaces.
+    """
+    return " ".join(token.text for token in tokenize_with_offsets(text))
+
+
+def tokenize(text: str) -> list[str]:
+    """Normalize and split text into phrase-detection tokens.
+
+    Args:
+        text (str): Raw text to tokenize.
+
+    Returns:
+        list[str]: Normalized token strings.
+    """
+    return [token.text for token in tokenize_with_offsets(text)]
+
+
+def tokenize_with_offsets(text: str) -> list[NormalizedToken]:
+    """Normalize text into tokens while preserving original character offsets.
+
+    Args:
+        text (str): Raw text to tokenize.
+
+    Returns:
+        list[NormalizedToken]: Normalized tokens with their source character spans.
+    """
+    tokens: list[NormalizedToken] = []
+    current: list[str] = []
+    start_char: int | None = None
+
+    for index, char in enumerate(text):
+        normalized = normalize_char(char)
+        if normalized == " ":
+            if current and start_char is not None:
+                tokens.append(NormalizedToken(text="".join(current), start_char=start_char, end_char=index))
+                current = []
+                start_char = None
+            continue
+        if start_char is None:
+            start_char = index
+        current.append(normalized)
+
+    if current and start_char is not None:
+        tokens.append(NormalizedToken(text="".join(current), start_char=start_char, end_char=len(text)))
+    return tokens
+
+
+def normalize_char(char: str) -> str:
+    """Normalize one character into a token character or a separator.
+
+    Args:
+        char (str): Single source character to normalize.
+
+    Returns:
+        str: The normalized token character, or a space acting as a separator.
+    """
+    if char in {"\u2019", "\u2018"}:
+        return "'"
+    if char in {"-", "\u2013", "\u2014"}:
+        return " "
+
+    lowered = char.lower()
+    if lowered in "abcdefghijklmnopqrstuvwxyz0123456789'":
+        return lowered
+    return " "

python/ebook_search/rerank.py

@@ -0,0 +1,140 @@
+"""vLLM-backed optional reranking."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, replace
+from typing import TYPE_CHECKING
+
+from python.ebook_search.llm_interface import request_rerank
+
+if TYPE_CHECKING:
+    import httpx
+
+    from python.ebook_search.config import RerankConfig
+    from python.ebook_search.search import SearchResult
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class RerankResult:
+    """A relevance score for one candidate chunk."""
+
+    chunk_id: int
+    score: float
+
+
+async def rerank_chunks(
+    client: httpx.AsyncClient,
+    query: str,
+    candidates: list[SearchResult],
+    config: RerankConfig,
+) -> list[SearchResult]:
+    """Rerank candidates with a vLLM rerank endpoint."""
+    if not candidates:
+        return []
+
+    logger.info(
+        "ebook_rerank_request_start base_url=%s model=%s candidates=%s",
+        config.base_url,
+        config.model,
+        len(candidates),
+    )
+    scores = await score_candidates(client, query, candidates, config)
+    results = sorted(
+        (
+            replace(
+                result,
+                score=final_rerank_score(result, scores[result.chunk_id].score, candidates, config),
+                rerank_score=scores[result.chunk_id].score,
+            )
+            for result in candidates
+        ),
+        key=lambda result: result.score,
+        reverse=True,
+    )
+    logger.info(
+        "ebook_rerank_request_complete base_url=%s model=%s candidates=%s",
+        config.base_url,
+        config.model,
+        len(results),
+    )
+    return results
+
+
+async def score_candidates(
+    client: httpx.AsyncClient,
+    query: str,
+    candidates: list[SearchResult],
+    config: RerankConfig,
+) -> dict[int, RerankResult]:
+    """Score candidate chunks with the configured rerank API."""
+    body = await request_rerank(client, query, [candidate.text for candidate in candidates], config)
+    if body is None:
+        return zero_rerank_scores(candidates)
+
+    scores = parse_vllm_scores(body, candidates)
+    for result in scores.values():
+        logger.debug("ebook_rerank_candidate_scored chunk_id=%s score=%s", result.chunk_id, result.score)
+    return scores
+
+
+def parse_vllm_scores(body: object, candidates: list[SearchResult]) -> dict[int, RerankResult]:
+    """Parse vLLM rerank scores into chunk-id keyed results."""
+    if not isinstance(body, dict):
+        logger.debug("ebook_rerank_response_not_object", extra={"response": body})
+        return zero_rerank_scores(candidates)
+
+    results = body.get("results") or body.get("data")
+    if not isinstance(results, list):
+        logger.debug("ebook_rerank_response_missing_results", extra={"response": body})
+        return zero_rerank_scores(candidates)
+
+    scores = zero_rerank_scores(candidates)
+    for item in results:
+        if not isinstance(item, dict):
+            continue
+        index = item.get("index")
+        score = item.get("relevance_score", item.get("score"))
+        if not isinstance(index, int) or index < 0 or index >= len(candidates):
+            continue
+        if not isinstance(score, int | float):
+            continue
+        chunk_id = candidates[index].chunk_id
+        scores[chunk_id] = RerankResult(chunk_id=chunk_id, score=clamp_score(float(score)))
+    return scores
+
+
+def zero_rerank_scores(candidates: list[SearchResult]) -> dict[int, RerankResult]:
+    """Return zero relevance scores for all candidate chunks."""
+    return {candidate.chunk_id: RerankResult(chunk_id=candidate.chunk_id, score=0.0) for candidate in candidates}
+
+
+def clamp_score(score: float) -> float:
+    """Clamp a rerank score into the supported 0.0 to 1.0 range."""
+    return min(max(score, 0.0), 1.0)
+
+
+def final_rerank_score(
+    result: SearchResult,
+    rerank_score: float,
+    candidates: list[SearchResult],
+    config: RerankConfig,
+) -> float:
+    """Combine rerank relevance with normalized hybrid retrieval evidence."""
+    return (config.score_weight * rerank_score) + (config.hybrid_weight * normalized_hybrid_score(result, candidates))
+
+
+def normalized_hybrid_score(result: SearchResult, candidates: list[SearchResult]) -> float:
+    """Normalize a candidate hybrid score against the rerank candidate set."""
+    hybrid_scores = [
+        candidate.fused_score if candidate.fused_score is not None else candidate.score for candidate in candidates
+    ]
+    low = min(hybrid_scores)
+    high = max(hybrid_scores)
+    if high == low:
+        return 1.0
+
+    score = result.fused_score if result.fused_score is not None else result.score
+    return (score - low) / (high - low)