fixed prompt_bench.py

config.py

@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from datetime import date
 from pathlib import Path
 import tomllib
 
@@ -69,50 +68,10 @@ class BenchmarkConfig:
         return cls(**raw)
 
 
-@dataclass
-class BertTopicTrainConfig:
-    """BERTopic training configuration loaded from TOML."""
-
-    sample_rate: float
-    min_text_length: int
-    n_topics: int
-    model_save_path: str
-    model_version: str | None = None
-
-    @classmethod
-    def from_toml(cls, config_path: Path) -> BertTopicTrainConfig:
-        """Load BERTopic training config from a TOML file."""
-        raw = tomllib.loads(config_path.read_text())["bertopic"]["train"]
-
-        today = date.today().isoformat()
-        if raw.get("model_version") is None:
-            raw["model_version"] = (
-                f"{today}-{raw['sample_rate']}-{raw['min_text_length']}-{raw['n_topics']}"
-            )
-        return cls(**raw)
-
-
-@dataclass
-class BertTopicInferConfig:
-    """BERTopic inference configuration loaded from TOML."""
-
-    min_text_length: int
-    poc_batch_size: int
-    model_version: str
-    model_save_path: str
-
-    @classmethod
-    def from_toml(cls, config_path: Path) -> BertTopicInferConfig:
-        """Load BERTopic inference config from a TOML file."""
-        raw = tomllib.loads(config_path.read_text())["bertopic"]["infer"]
-        return cls(**raw)
-
-
 def get_config_dir() -> Path:
     """Get the path to the config file."""
     return Path(__file__).resolve().parent.parent.parent / "config"
 
-
 def default_config_path() -> Path:
     """Get the path to the config file."""
     return get_config_dir() / "config.toml"
@@ -128,19 +87,3 @@ def get_benchmark_config(config_path: Path | None = None) -> BenchmarkConfig:
     if config_path is None:
         config_path = default_config_path()
     return BenchmarkConfig.from_toml(config_path)
-
-
-def get_bertopic_train_config(
-    config_path: Path | None = None,
-) -> BertTopicTrainConfig:
-    if config_path is None:
-        config_path = default_config_path()
-    return BertTopicTrainConfig.from_toml(config_path)
-
-
-def get_bertopic_infer_config(
-    config_path: Path | None = None,
-) -> BertTopicInferConfig:
-    if config_path is None:
-        config_path = default_config_path()
-    return BertTopicInferConfig.from_toml(config_path)

pipelines/bert_topic/infer.py

@@ -1,116 +0,0 @@
-"""Nornsight — BERTopic POC Inference Script.
-
-Loads the trained model and labels a small batch of posts,
-writing results to main.post_topic for inspection.
-
-POC: processes a single batch of 1k posts to validate the pipeline end-to-end.
-"""
-
-from __future__ import annotations
-
-import logging
-import time
-from collections import Counter
-from pathlib import Path
-
-from bertopic import BERTopic
-from sqlalchemy import Engine, func, insert, select
-from sqlalchemy.orm import Session
-
-from pipelines.config import BertTopicInferConfig, get_bertopic_infer_config
-from pipelines.orm.common import get_postgres_engine
-from pipelines.orm.data_science_dev.posts import PostTopic, Posts
-from pipelines.orm.data_science_dev.posts.lang_filters import ENGLISH_LANGS
-from pipelines.pipelines.common import configure_logger
-
-logger = logging.getLogger(__name__)
-
-
-def main() -> None:
-    """Run BERTopic inference against a sample of posts."""
-    configure_logger()
-
-    config = get_bertopic_infer_config()
-    run_inference(config)
-    logger.info(
-        "POC inference complete. Check main.post_topic in DBeaver to inspect results."
-    )
-
-
-def run_inference(config: BertTopicInferConfig) -> None:
-    model_save_path = Path(config.model_save_path)
-
-    logger.info(f"Loading BERTopic model from {model_save_path}")
-    topic_model = BERTopic.load(str(model_save_path))
-
-    topic_info = topic_model.get_topic_info()
-    label_map: dict[int, str] = dict(zip(topic_info["Topic"], topic_info["Name"]))
-    logger.info(f"Model loaded with {len(label_map)} topics")
-
-    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
-
-    post_ids, texts = get_post_ids_and_test(engine, config)
-
-    logger.info(f"Fetched {len(texts)} posts")
-
-    logger.info("Running BERTopic transform")
-    start = time.perf_counter()
-    topics, _probabilities = topic_model.transform(texts)
-    elapsed = time.perf_counter() - start
-    logger.info(f"Transform complete in {elapsed:.1f}s")
-
-    # Write results to main.post_topic
-    records = [
-        {
-            "post_id": pid,
-            "topic_id": int(topic_id),
-            "topic_label": label_map.get(int(topic_id), "unknown"),
-            "model_version": config.model_version,
-        }
-        for pid, topic_id in zip(post_ids, topics)
-    ]
-    with Session(engine) as session:
-        session.execute(insert(PostTopic), records)
-        session.commit()
-
-    count_topics(records)
-    logger.info(f"Wrote {len(records)} topic labels to main.post_topic")
-
-
-def get_post_ids_and_test(
-    engine: Engine,
-    config: BertTopicInferConfig,
-) -> None | tuple[list[int], list[str]]:
-    with Session(engine) as session:
-        logger.info(f"Fetching {config.poc_batch_size} posts for inference")
-        # Pull a fresh batch for inference — distinct from training sample
-        # using a fixed seed offset so we're not re-labeling training posts
-        stmt = select(Posts).where(
-            Posts.text.is_not(None),
-            Posts.langs.in_(ENGLISH_LANGS),
-            func.length(Posts.text) > config.min_text_length,
-        )
-        if config.poc_batch_size > 0:
-            stmt = stmt.limit(config.poc_batch_size)
-
-        posts = session.scalars(stmt).all()
-        if not posts:
-            logger.warning("No posts were selected for inference")
-            return [], []
-
-        post_ids = [post.post_id for post in posts]
-        texts = [post.text.strip() for post in posts]
-
-    return post_ids, texts
-
-
-def count_topics(records: list[dict]) -> None:
-    topic_counts = Counter(record.get("topic_label", "unknown") for record in records)
-
-    logger.info("Topic distribution in this batch:")
-    for label, count in topic_counts.most_common(10):
-        logger.info("  %s: %d", label, count)
-
-
-if __name__ == "__main__":
-    main()

pipelines/bert_topic/train.py

@@ -1,119 +0,0 @@
-"""Nornsight — BERTopic POC Training Script.
-
-Pulls a small stratified sample (~11.5k posts) from main.posts,
-trains BERTopic with MiniBatchKMeans on Jeeves, and saves the model locally.
-
-POC sample rate: random() < 0.00005 (~0.005% of 230M = ~11.5k posts)
-Full training rate will be: random() < 0.005 (~1.08M posts)
-"""
-
-from __future__ import annotations
-
-import logging
-import time
-from pathlib import Path
-
-from bertopic import BERTopic
-from sklearn.cluster import MiniBatchKMeans
-from sqlalchemy import func, select
-from sqlalchemy.orm import Session
-
-from pipelines.config import BertTopicTrainConfig, get_bertopic_train_config
-from pipelines.orm.common import get_postgres_engine
-from pipelines.orm.data_science_dev.posts import Posts
-from pipelines.orm.data_science_dev.posts.lang_filters import ENGLISH_LANGS
-from pipelines.pipelines.common import configure_logger
-
-logger = logging.getLogger(__name__)
-
-
-def main() -> None:
-    """Train and persist the BERTopic model."""
-    configure_logger()
-
-    config = get_bertopic_train_config()
-    docs = load_sample(config)
-    if not docs:
-        logger.warning("No training documents were selected")
-        return
-
-    train(docs, config)
-    logger.info(f"Done. Model saved as version {config.model_version}")
-    logger.info("Next: run infer.py to label a sample of posts in the database")
-
-
-def load_sample(config: BertTopicTrainConfig) -> list[str]:
-    logger.info("Connecting to PostgreSQL via SQLAlchemy")
-    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
-
-    logger.info(f"Pulling sample from main.posts (sample_rate={config.sample_rate})")
-    start = time.perf_counter()
-
-    with Session(engine) as session:
-        texts = session.scalars(
-            select(Posts.text).where(
-                Posts.text.is_not(None),
-                Posts.langs.in_(ENGLISH_LANGS),
-                func.length(Posts.text) > config.min_text_length,
-                func.random() < config.sample_rate,
-            )
-        ).all()
-
-    elapsed = time.perf_counter() - start
-    logger.info(f"Fetched {len(texts)} rows in {elapsed:.1f}s")
-
-    # Basic cleaning — strip whitespace and deduplicate
-    docs = list({text.strip() for text in texts})
-    logger.info(f"After cleaning and dedup: {len(docs)} posts")
-
-    return docs
-
-
-def train(docs: list[str], config: BertTopicTrainConfig) -> None:
-    logger.info(
-        f"Initialising BERTopic with MiniBatchKMeans (n_topics={config.n_topics})"
-    )
-
-    cluster_model = MiniBatchKMeans(
-        n_clusters=config.n_topics,
-        random_state=42,
-        batch_size=1024,
-        n_init=3,
-        verbose=1,
-    )
-
-    topic_model = BERTopic(
-        hdbscan_model=cluster_model,
-        language="english",
-        calculate_probabilities=False,  # saves memory
-        verbose=True,
-    )
-
-    logger.info(f"Starting fit_transform on {len(docs)} posts (CPU)")
-    start = time.perf_counter()
-
-    topic_model.fit_transform(docs)
-
-    elapsed = time.perf_counter() - start
-    logger.info(f"Training complete in {elapsed:.1f}s ({elapsed / 60:.1f} min)")
-
-    # Log topic summary for quick inspection
-    topic_info = topic_model.get_topic_info()
-    logger.info(f"Topics found: {len(topic_info)}")
-    logger.info(f"\n{topic_info.to_string()}")
-
-    model_save_path = Path(config.model_save_path)
-    model_save_path.mkdir(parents=True, exist_ok=True)
-    logger.info(f"Saving model to {model_save_path}")
-
-    topic_model.save(
-        str(model_save_path),
-        serialization="safetensors",
-        save_ctfidf=True,
-        save_embedding_model=True,
-    )
-    logger.info("Model saved")
-
-
-if __name__ == "__main__":
-    main()

pipelines/containers/__init__.py

@@ -0,0 +1 @@
+"""Prompt benchmarking system for evaluating LLMs via vLLM."""

pipelines/containers/bert_topic.py

@@ -0,0 +1,235 @@
+"""Docker container lifecycle management for BERTopic jobs on Jeeves."""
+
+from __future__ import annotations
+
+import logging
+import os
+import subprocess
+from pathlib import Path
+from typing import Annotated, Literal
+
+import typer
+
+logger = logging.getLogger(__name__)
+
+JOBMode = Literal["train", "infer"]
+IMAGE_NAME = "bert-topic:latest"
+REPO_DIR = Path(__file__).resolve().parents[3]
+DEFAULT_CACHE_ROOT = Path("/zfs/storage/main/ds_thing/models/bert_topic")
+DEFAULT_POSTGRES_SOCKET_DIR = Path("/run/postgresql")
+DB_ENV_VARS = (
+    "DATA_SCIENCE_DEV_DB",
+    "DATA_SCIENCE_DEV_HOST",
+    "DATA_SCIENCE_DEV_PORT",
+    "DATA_SCIENCE_DEV_USER",
+    "DATA_SCIENCE_DEV_PASSWORD",
+)
+
+app = typer.Typer(help="BERTopic container management.")
+
+
+def _container_name(mode: JOBMode) -> str:
+    """Return the Docker container name for the selected BERTopic job."""
+    return f"bert-topic-{mode}"
+
+
+def _module_name(mode: JOBMode) -> str:
+    """Return the Python module to run inside the container."""
+    return f"pipelines.bert_topic.{mode}"
+
+
+def _env_args(*, use_postgres_socket: bool) -> list[str]:
+    """Pass through database environment variables from the host shell."""
+    required = [
+        "DATA_SCIENCE_DEV_DB",
+        "DATA_SCIENCE_DEV_PORT",
+        "DATA_SCIENCE_DEV_USER",
+    ]
+    if not use_postgres_socket:
+        required.append("DATA_SCIENCE_DEV_HOST")
+    missing = [name for name in required if not os.getenv(name)]
+    if missing:
+        message = "Missing required database environment variables: " + ", ".join(
+            missing
+        )
+        raise RuntimeError(message)
+    args: list[str] = []
+    if use_postgres_socket:
+        args.extend(["-e", f"DATA_SCIENCE_DEV_HOST={DEFAULT_POSTGRES_SOCKET_DIR}"])
+    for name in DB_ENV_VARS:
+        if use_postgres_socket and name == "DATA_SCIENCE_DEV_HOST":
+            continue
+        if os.getenv(name):
+            args.extend(["-e", name])
+    return args
+
+
+def build_image() -> None:
+    """Build the BERTopic Docker image."""
+    dockerfile = REPO_DIR / "pipelines/containers/docker_files/Dockerfile.bert_topic"
+    logger.info("Building BERTopic image: %s", IMAGE_NAME)
+    result = subprocess.run(
+        [
+            "docker",
+            "build",
+            "--network",
+            "host",
+            "-f",
+            str(dockerfile),
+            "-t",
+            IMAGE_NAME,
+            str(REPO_DIR),
+        ],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        message = (
+            "Failed to build BERTopic image. "
+            f"docker build stderr:\n{result.stderr.strip()}"
+        )
+        raise RuntimeError(message)
+    logger.info("Image built: %s", IMAGE_NAME)
+
+
+def stop_job(*, mode: JOBMode) -> None:
+    """Stop and remove the BERTopic container for the selected mode."""
+    container_name = _container_name(mode)
+    logger.info("Stopping BERTopic container: %s", container_name)
+    subprocess.run(["docker", "stop", container_name], capture_output=True, check=False)
+    subprocess.run(
+        ["docker", "rm", "-f", container_name], capture_output=True, check=False
+    )
+
+
+def start_job(
+    *,
+    mode: JOBMode,
+    cache_root: Path = DEFAULT_CACHE_ROOT,
+    postgres_socket_dir: Path = DEFAULT_POSTGRES_SOCKET_DIR,
+    detach: bool = False,
+) -> None:
+    """Run BERTopic training or inference in Docker on Jeeves."""
+    cache_root = cache_root.resolve()
+    cache_root.mkdir(parents=True, exist_ok=True)
+    postgres_socket_dir = postgres_socket_dir.resolve()
+    stop_job(mode=mode)
+    use_postgres_socket = postgres_socket_dir.exists()
+
+    command = [
+        "docker",
+        "run",
+        "--name",
+        _container_name(mode),
+        "--ipc=host",
+        "-v",
+        f"{cache_root}:/cache",
+        *_env_args(use_postgres_socket=use_postgres_socket),
+        IMAGE_NAME,
+        _module_name(mode),
+    ]
+    if use_postgres_socket:
+        command[7:7] = ["-v", f"{postgres_socket_dir}:{DEFAULT_POSTGRES_SOCKET_DIR}"]
+    if detach:
+        command.insert(2, "-d")
+
+    logger.info("Starting BERTopic %s container", mode)
+    logger.info("  Cache root: %s", cache_root)
+    if use_postgres_socket:
+        logger.info("  Postgres socket: %s", postgres_socket_dir)
+    result = subprocess.run(command, text=True, capture_output=detach, check=False)
+    if result.returncode != 0:
+        detail = (
+            result.stderr.strip() if result.stderr else f"exit code {result.returncode}"
+        )
+        raise RuntimeError(f"BERTopic container failed to start: {detail}")
+    if detach:
+        logger.info("Container started: %s", result.stdout.strip()[:12])
+    else:
+        logger.info("BERTopic %s run complete", mode)
+
+
+def logs_job(*, mode: JOBMode) -> str | None:
+    """Return recent logs from the BERTopic container, or None if absent."""
+    result = subprocess.run(
+        ["docker", "logs", "--tail", "100", _container_name(mode)],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        return None
+    return result.stdout + result.stderr
+
+
+@app.command()
+def build(
+    log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
+) -> None:
+    """Build the BERTopic Docker image."""
+    logging.basicConfig(
+        level=log_level,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+    build_image()
+
+
+@app.command("run")
+def run_job_command(
+    mode: Annotated[JOBMode, typer.Option(help="Which BERTopic job to run")] = "train",
+    cache_root: Annotated[
+        Path, typer.Option(help="Host path mounted to /cache for model and HF cache")
+    ] = DEFAULT_CACHE_ROOT,
+    postgres_socket_dir: Annotated[
+        Path, typer.Option(help="Host Postgres socket directory to mount into the container")
+    ] = DEFAULT_POSTGRES_SOCKET_DIR,
+    detach: Annotated[
+        bool, typer.Option(help="Start the container in the background")
+    ] = False,
+    log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
+) -> None:
+    """Run BERTopic training or inference inside Docker."""
+    logging.basicConfig(
+        level=log_level,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+    start_job(
+        mode=mode,
+        cache_root=cache_root,
+        postgres_socket_dir=postgres_socket_dir,
+        detach=detach,
+    )
+
+
+@app.command("stop")
+def stop_job_command(
+    mode: Annotated[
+        JOBMode, typer.Option(help="Which BERTopic container to stop")
+    ] = "train",
+) -> None:
+    """Stop and remove the BERTopic container."""
+    stop_job(mode=mode)
+
+
+@app.command("logs")
+def logs_job_command(
+    mode: Annotated[
+        JOBMode, typer.Option(help="Which BERTopic container logs to show")
+    ] = "train",
+) -> None:
+    """Show recent logs from the BERTopic container."""
+    output = logs_job(mode=mode)
+    if output is None:
+        typer.echo(f"No BERTopic container found for mode={mode}.")
+        raise typer.Exit(code=1)
+    typer.echo(output)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    app()
+
+
+if __name__ == "__main__":
+    cli()

pipelines/containers/docker_files/Dockerfile.bert_topic

@@ -0,0 +1,38 @@
+FROM python:3.12-bookworm
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV PYTHONUNBUFFERED=1
+ENV PIP_NO_CACHE_DIR=1
+
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    gcc \
+    g++ \
+    git \
+    libgomp1 \
+    libpq-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY pipelines ./pipelines
+
+RUN python -m pip install --upgrade pip setuptools wheel && \
+    python -m pip install \
+    torch \
+    --index-url https://download.pytorch.org/whl/cpu && \
+    python -m pip install \
+    typer \
+    sqlalchemy \
+    bertopic \
+    sentence-transformers \
+    scikit-learn \
+    pandas \
+    numpy \
+    "psycopg[binary]"
+
+ENV HF_HOME=/cache/huggingface
+ENV TRANSFORMERS_CACHE=/cache/huggingface
+
+ENTRYPOINT ["python", "-m"]
+CMD ["pipelines.bert_topic.train"]

pipelines/containers/docker_files/Dockerfile.finetune

@@ -0,0 +1,11 @@
+FROM ghcr.io/unslothai/unsloth:latest
+
+RUN pip install --no-cache-dir typer
+
+WORKDIR /workspace
+COPY python/prompt_bench/finetune.py python/prompt_bench/finetune.py
+COPY config/prompts/summarization_prompts.toml config/prompts/summarization_prompts.toml
+COPY python/prompt_bench/__init__.py python/prompt_bench/__init__.py
+COPY python/__init__.py python/__init__.py
+
+ENTRYPOINT ["python", "-m", "pipelines.prompt_bench.finetune"]

pipelines/containers/finetune.py

@@ -9,7 +9,7 @@ from typing import Annotated
 
 import typer
 
-from pipelines.tools.containers.lib import check_gpu_free
+from pipelines.pipelines.containers.lib import check_gpu_free
 
 logger = logging.getLogger(__name__)
 
@@ -27,7 +27,7 @@ def build_image() -> None:
             "docker",
             "build",
             "-f",
-            str(REPO_DIR / "python/prompt_bench/Dockerfile.finetune"),
+            str(REPO_DIR / "pipelines/containers/docker_files/Dockerfile.finetune"),
             "-t",
             FINETUNE_IMAGE,
             ".",

pipelines/jobs/calculate_legislator_scores.py

@@ -0,0 +1,574 @@
+"""Calculate legislator topic scores from bill topics and roll-call votes."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from typing import Annotated, Sequence
+
+import typer
+from sqlalchemy import (
+    ColumnElement,
+    Integer,
+    Select,
+    and_,
+    case,
+    cast,
+    delete,
+    extract,
+    func,
+    or_,
+    select,
+    tuple_,
+)
+from sqlalchemy.orm import Session
+
+from pipelines.congress_vote_context import create_score_run, finalize_score_run
+from pipelines.orm.common import get_postgres_engine
+from pipelines.orm.data_science_dev.congress import (
+    BillTopic,
+    BillTopicPosition,
+    LegislatorScore,
+    SubjectType,
+    Vote,
+    VoteClassification,
+    VoteEffect,
+    VoteMeasureLink,
+    VoteMeasureRole,
+    VotePositionMeaning,
+    VoteRelationship,
+    VoteRecord,
+)
+from pipelines.pipelines.jobs.extract_bill_topics import normalize_topic_label
+from pipelines.web.scoring import (
+    OPPOSE_POSITIONS,
+    SUPPORT_POSITIONS,
+    normalized_position_expression,
+)
+
+logger = logging.getLogger(__name__)
+
+DELETE_BATCH_SIZE = 5_000
+
+
+@dataclass(frozen=True)
+class ScoreDiagnostics:
+    """Counts for the input stages required to calculate legislator scores."""
+
+    bill_topic_rows: int
+    linked_vote_rows: int
+    vote_record_rows: int
+    topic_vote_links: int
+    scorable_vote_records: int
+
+
+@dataclass(frozen=True)
+class LegislatorScoreInput:
+    """One aggregated score ready to store in legislator_score."""
+
+    legislator_id: int
+    year: int
+    topic: str
+    score: float
+    supportive: int
+    opposed: int
+
+
+def create_legislator_score_query(
+    *,
+    congress: int | None = None,
+    bill_ids: Sequence[int] | None = None,
+    topics: Sequence[str] | None = None,
+) -> Select:
+    """Build the aggregate score query from extracted bill topics and vote records."""
+    normalized_vote = normalized_position_expression(VoteRecord.position)
+    supportive_vote = _supportive_vote_expression(normalized_vote)
+    opposed_vote = _opposed_vote_expression(normalized_vote)
+    supportive_count = func.sum(supportive_vote)
+    opposed_count = func.sum(opposed_vote)
+    total_count = supportive_count + opposed_count
+    vote_year = cast(extract("year", Vote.vote_date), Integer)
+    score = (100.0 * supportive_count / func.nullif(total_count, 0)).label("score")
+
+    stmt = (
+        select(
+            VoteRecord.legislator_id.label("legislator_id"),
+            vote_year.label("year"),
+            BillTopic.topic.label("topic"),
+            score,
+            supportive_count.label("supportive"),
+            opposed_count.label("opposed"),
+            total_count.label("total"),
+        )
+        .select_from(VoteRecord)
+        .join(Vote, Vote.id == VoteRecord.vote_id)
+        .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+        .join(VotePositionMeaning, VotePositionMeaning.vote_id == Vote.id)
+        .join(
+            VoteMeasureLink,
+            and_(
+                VoteMeasureLink.vote_id == Vote.id,
+                VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+            ),
+        )
+        .join(BillTopic, BillTopic.bill_id == VoteMeasureLink.measure_id)
+        .where(
+            *_eligible_vote_filters(),
+            _is_scorable_position(normalized_vote),
+        )
+        .group_by(VoteRecord.legislator_id, vote_year, BillTopic.topic)
+        .having(total_count > 0)
+        .order_by(VoteRecord.legislator_id, vote_year, BillTopic.topic)
+    )
+    if congress is not None:
+        stmt = stmt.where(Vote.congress == congress)
+    if bill_ids:
+        stmt = stmt.where(VoteMeasureLink.measure_id.in_(list(bill_ids)))
+
+    normalized_topics = _normalize_topics(topics)
+    if normalized_topics:
+        stmt = stmt.where(BillTopic.topic.in_(normalized_topics))
+
+    return stmt
+
+
+def collect_legislator_scores(
+    session: Session,
+    *,
+    congress: int | None = None,
+    bill_ids: Sequence[int] | None = None,
+    topics: Sequence[str] | None = None,
+) -> list[LegislatorScoreInput]:
+    """Run the aggregate query and return score rows."""
+    rows = session.execute(
+        create_legislator_score_query(
+            congress=congress,
+            bill_ids=bill_ids,
+            topics=topics,
+        )
+    )
+    return [
+        LegislatorScoreInput(
+            legislator_id=int(row.legislator_id),
+            year=int(row.year),
+            topic=str(row.topic),
+            score=float(row.score),
+            supportive=int(row.supportive),
+            opposed=int(row.opposed),
+        )
+        for row in rows
+        if row.score is not None
+    ]
+
+
+def collect_score_diagnostics(
+    session: Session,
+    *,
+    congress: int | None = None,
+    bill_ids: Sequence[int] | None = None,
+    topics: Sequence[str] | None = None,
+) -> ScoreDiagnostics:
+    """Count score pipeline inputs for explaining empty score runs."""
+    normalized_topics = _normalize_topics(topics)
+    vote_filters = _vote_scope_filters(congress=congress, bill_ids=bill_ids)
+    topic_filters = _topic_scope_filters(bill_ids=bill_ids, topics=normalized_topics)
+    normalized_vote = normalized_position_expression(VoteRecord.position)
+    eligible_vote_filters = _eligible_vote_filters()
+
+    bill_topic_rows = session.scalar(
+        select(func.count(BillTopic.id)).where(*topic_filters)
+    )
+    linked_vote_rows = session.scalar(
+        select(func.count(func.distinct(Vote.id)))
+        .select_from(Vote)
+        .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+        .join(
+            VoteMeasureLink,
+            and_(
+                VoteMeasureLink.vote_id == Vote.id,
+                VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+            ),
+        )
+        .where(*vote_filters, *eligible_vote_filters)
+    )
+    vote_record_rows = session.scalar(
+        select(func.count())
+        .select_from(VoteRecord)
+        .join(Vote, Vote.id == VoteRecord.vote_id)
+        .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+        .where(*vote_filters, *eligible_vote_filters)
+    )
+    topic_vote_links = session.scalar(
+        select(func.count())
+        .select_from(VoteRecord)
+        .join(Vote, Vote.id == VoteRecord.vote_id)
+        .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+        .join(VotePositionMeaning, VotePositionMeaning.vote_id == Vote.id)
+        .join(
+            VoteMeasureLink,
+            and_(
+                VoteMeasureLink.vote_id == Vote.id,
+                VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+            ),
+        )
+        .join(BillTopic, BillTopic.bill_id == VoteMeasureLink.measure_id)
+        .where(*vote_filters, *topic_filters, *eligible_vote_filters)
+    )
+    scorable_vote_records = session.scalar(
+        select(func.count())
+        .select_from(VoteRecord)
+        .join(Vote, Vote.id == VoteRecord.vote_id)
+        .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+        .join(VotePositionMeaning, VotePositionMeaning.vote_id == Vote.id)
+        .join(
+            VoteMeasureLink,
+            and_(
+                VoteMeasureLink.vote_id == Vote.id,
+                VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+            ),
+        )
+        .join(BillTopic, BillTopic.bill_id == VoteMeasureLink.measure_id)
+        .where(
+            *vote_filters,
+            *topic_filters,
+            *eligible_vote_filters,
+            _is_scorable_position(normalized_vote),
+        )
+    )
+
+    return ScoreDiagnostics(
+        bill_topic_rows=bill_topic_rows or 0,
+        linked_vote_rows=linked_vote_rows or 0,
+        vote_record_rows=vote_record_rows or 0,
+        topic_vote_links=topic_vote_links or 0,
+        scorable_vote_records=scorable_vote_records or 0,
+    )
+
+
+def store_legislator_scores(
+    session: Session,
+    rows: Sequence[LegislatorScoreInput],
+    *,
+    score_run_id: int | None,
+    replace_all: bool = False,
+) -> int:
+    """Replace matching score rows and insert the newly calculated scores."""
+    if replace_all:
+        session.execute(delete(LegislatorScore))
+    elif rows:
+        keys = [
+            (row.legislator_id, row.year, row.topic)
+            for row in rows
+        ]
+        for key_batch in _batched(keys, DELETE_BATCH_SIZE):
+            session.execute(
+                delete(LegislatorScore).where(
+                    tuple_(
+                        LegislatorScore.legislator_id,
+                        LegislatorScore.year,
+                        LegislatorScore.topic,
+                    ).in_(key_batch)
+                )
+            )
+
+    session.add_all(
+        [
+            LegislatorScore(
+                legislator_id=row.legislator_id,
+                year=row.year,
+                topic=row.topic,
+                score=row.score,
+                score_run_id=score_run_id,
+            )
+            for row in rows
+        ]
+    )
+    return len(rows)
+
+
+def _supportive_vote_expression(
+    normalized_vote: ColumnElement[str | None],
+) -> ColumnElement[int]:
+    supports_text = _position_matches_effect(normalized_vote, VoteEffect.SUPPORTS_TEXT)
+    opposes_text = _position_matches_effect(normalized_vote, VoteEffect.OPPOSES_TEXT)
+    return case(
+        (
+            and_(
+                BillTopic.support_position == BillTopicPosition.FOR,
+                supports_text,
+            ),
+            1,
+        ),
+        (
+            and_(
+                BillTopic.support_position == BillTopicPosition.AGAINST,
+                opposes_text,
+            ),
+            1,
+        ),
+        else_=0,
+    )
+
+
+def _opposed_vote_expression(
+    normalized_vote: ColumnElement[str | None],
+) -> ColumnElement[int]:
+    supports_text = _position_matches_effect(normalized_vote, VoteEffect.SUPPORTS_TEXT)
+    opposes_text = _position_matches_effect(normalized_vote, VoteEffect.OPPOSES_TEXT)
+    return case(
+        (
+            and_(
+                BillTopic.support_position == BillTopicPosition.FOR,
+                opposes_text,
+            ),
+            1,
+        ),
+        (
+            and_(
+                BillTopic.support_position == BillTopicPosition.AGAINST,
+                supports_text,
+            ),
+            1,
+        ),
+        else_=0,
+    )
+
+
+def _position_matches_effect(
+    normalized_vote: ColumnElement[str | None],
+    effect: VoteEffect,
+) -> ColumnElement[bool]:
+    return or_(
+        and_(
+            normalized_vote.in_(sorted(SUPPORT_POSITIONS)),
+            VotePositionMeaning.yea_effect == effect,
+        ),
+        and_(
+            normalized_vote.in_(sorted(OPPOSE_POSITIONS)),
+            VotePositionMeaning.nay_effect == effect,
+        ),
+        and_(
+            normalized_vote == "present",
+            VotePositionMeaning.present_effect == effect,
+        ),
+    )
+
+
+def _is_scorable_position(normalized_vote: ColumnElement[str | None]) -> ColumnElement[bool]:
+    return or_(
+        _position_matches_effect(normalized_vote, VoteEffect.SUPPORTS_TEXT),
+        _position_matches_effect(normalized_vote, VoteEffect.OPPOSES_TEXT),
+    )
+
+
+def _normalize_topics(topics: Sequence[str] | None) -> list[str]:
+    normalized: list[str] = []
+    seen: set[str] = set()
+    for topic in topics or []:
+        value = normalize_topic_label(topic)
+        if value and value not in seen:
+            normalized.append(value)
+            seen.add(value)
+    return normalized
+
+
+def _batched[T](items: Sequence[T], batch_size: int) -> list[Sequence[T]]:
+    return [
+        items[index : index + batch_size]
+        for index in range(0, len(items), batch_size)
+    ]
+
+
+def _vote_scope_filters(
+    *,
+    congress: int | None,
+    bill_ids: Sequence[int] | None,
+) -> list[ColumnElement[bool]]:
+    filters: list[ColumnElement[bool]] = []
+    if congress is not None:
+        filters.append(Vote.congress == congress)
+    if bill_ids:
+        filters.append(VoteMeasureLink.measure_id.in_(list(bill_ids)))
+    return filters
+
+
+def _topic_scope_filters(
+    *,
+    bill_ids: Sequence[int] | None,
+    topics: Sequence[str],
+) -> list[ColumnElement[bool]]:
+    filters: list[ColumnElement[bool]] = []
+    if bill_ids:
+        filters.append(BillTopic.bill_id.in_(list(bill_ids)))
+    if topics:
+        filters.append(BillTopic.topic.in_(list(topics)))
+    return filters
+
+
+def _has_score_scope(
+    *,
+    congress: int | None,
+    bill_ids: Sequence[int] | None,
+    topics: Sequence[str] | None,
+) -> bool:
+    return congress is not None or bool(bill_ids) or bool(topics)
+
+
+def _eligible_vote_filters() -> list[ColumnElement[bool]]:
+    return [
+        VoteClassification.subject_type == SubjectType.MEASURE,
+        VoteClassification.vote_relationship == VoteRelationship.DIRECT_TEXT_VOTE,
+        VoteClassification.is_direct_vote_on_legislative_text.is_(True),
+        VoteClassification.is_substantive_policy_vote.is_(True),
+        VoteClassification.is_special_rule.is_(False),
+    ]
+
+
+def main(
+    congress: Annotated[
+        int | None,
+        typer.Option(help="Only score votes from one Congress."),
+    ] = None,
+    bill_ids: Annotated[
+        list[int] | None,
+        typer.Option(
+            "--bill-id",
+            help="Only score votes linked to one internal bill.id. Repeatable.",
+        ),
+    ] = None,
+    topics: Annotated[
+        list[str] | None,
+        typer.Option("--topic", help="Only score one normalized topic. Repeatable."),
+    ] = None,
+    replace_all: Annotated[
+        bool,
+        typer.Option(
+            help="Delete every existing legislator score before inserting. "
+            "Unfiltered runs do this automatically."
+        ),
+    ] = False,
+    dry_run: Annotated[
+        bool,
+        typer.Option(help="Calculate scores without writing to the database."),
+    ] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+    diagnose: Annotated[
+        bool,
+        typer.Option(help="Log input-stage counts even when rows are calculated."),
+    ] = False,
+) -> None:
+    """CLI entrypoint for calculating and storing legislator topic scores."""
+    logging.basicConfig(
+        level=log_level,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+
+    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
+    with Session(engine) as session:
+        rows = collect_legislator_scores(
+            session,
+            congress=congress,
+            bill_ids=bill_ids,
+            topics=topics,
+        )
+        logger.info("Calculated %d legislator topic score rows", len(rows))
+        if diagnose or not rows:
+            diagnostics = collect_score_diagnostics(
+                session,
+                congress=congress,
+                bill_ids=bill_ids,
+                topics=topics,
+            )
+            _log_diagnostics(diagnostics)
+
+        if dry_run:
+            session.rollback()
+            return
+
+        score_run = create_score_run(session)
+        should_replace_all = replace_all or not _has_score_scope(
+            congress=congress,
+            bill_ids=bill_ids,
+            topics=topics,
+        )
+        written = store_legislator_scores(
+            session,
+            rows,
+            score_run_id=score_run.id,
+            replace_all=should_replace_all,
+        )
+        included_vote_count = session.scalar(
+            select(func.count(func.distinct(Vote.id)))
+            .select_from(VoteRecord)
+            .join(Vote, Vote.id == VoteRecord.vote_id)
+            .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+            .join(VotePositionMeaning, VotePositionMeaning.vote_id == Vote.id)
+            .join(
+                VoteMeasureLink,
+                and_(
+                    VoteMeasureLink.vote_id == Vote.id,
+                    VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+                ),
+            )
+            .join(BillTopic, BillTopic.bill_id == VoteMeasureLink.measure_id)
+            .where(
+                *_vote_scope_filters(congress=congress, bill_ids=bill_ids),
+                *_topic_scope_filters(bill_ids=bill_ids, topics=_normalize_topics(topics)),
+                *_eligible_vote_filters(),
+                _is_scorable_position(normalized_position_expression(VoteRecord.position)),
+            )
+        ) or 0
+        total_scoped_votes = session.scalar(
+            select(func.count(func.distinct(Vote.id)))
+            .select_from(Vote)
+            .join(VoteClassification, VoteClassification.vote_id == Vote.id)
+            .join(
+                VoteMeasureLink,
+                and_(
+                    VoteMeasureLink.vote_id == Vote.id,
+                    VoteMeasureLink.role == VoteMeasureRole.VOTED_ON,
+                ),
+            )
+            .where(*_vote_scope_filters(congress=congress, bill_ids=bill_ids))
+        ) or 0
+        finalize_score_run(
+            session,
+            score_run=score_run,
+            included_vote_count=included_vote_count,
+            excluded_vote_count=max(total_scoped_votes - included_vote_count, 0),
+        )
+        session.commit()
+        logger.info("Stored %d legislator topic score rows", written)
+
+
+def _log_diagnostics(diagnostics: ScoreDiagnostics) -> None:
+    logger.info(
+        "Score input diagnostics: bill_topic_rows=%d linked_vote_rows=%d "
+        "vote_record_rows=%d topic_vote_links=%d scorable_vote_records=%d",
+        diagnostics.bill_topic_rows,
+        diagnostics.linked_vote_rows,
+        diagnostics.vote_record_rows,
+        diagnostics.topic_vote_links,
+        diagnostics.scorable_vote_records,
+    )
+    if diagnostics.bill_topic_rows == 0:
+        logger.warning(
+            "No extracted bill topics matched the score scope. Run "
+            "pipelines.tools.extract_bill_topics after bill summarization."
+        )
+    elif diagnostics.linked_vote_rows == 0:
+        logger.warning("No direct substantive text votes matched the score scope.")
+    elif diagnostics.vote_record_rows == 0:
+        logger.warning("No individual vote records matched the score scope.")
+    elif diagnostics.topic_vote_links == 0:
+        logger.warning(
+            "Bill topics exist, but none are attached to bills that have eligible scored votes."
+        )
+    elif diagnostics.scorable_vote_records == 0:
+        logger.warning(
+            "Topic-vote links exist, but no joined vote records had Yea/Aye/Yes/Nay/No positions."
+        )
+
+
+if __name__ == "__main__":
+    typer.run(main)

pipelines/jobs/extract_bill_topics.py

@@ -0,0 +1,682 @@
+"""Extract bill topics from bill text using a configurable topic catalog."""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Annotated, Any, Sequence
+
+import httpx
+import typer
+from sqlalchemy import ColumnElement, Select, delete, exists, func, select
+from sqlalchemy.orm import Session, selectinload
+
+from pipelines.config import OpenAIConfig, get_config_dir, get_openai_config
+from pipelines.orm.common import get_postgres_engine
+from pipelines.orm.data_science_dev.congress import (
+    Bill,
+    BillText,
+    BillTopic,
+    BillTopicPosition,
+    SubjectType,
+    VoteClassification,
+    VoteRelationship,
+    VoteTextTarget,
+)
+
+logger = logging.getLogger(__name__)
+
+OPENAI_PROJECT_ID = "proj_fQBPEXFgnS87Fk6wZwploFwE"
+OPENAI_CHAT_COMPLETIONS_URL = "https://api.openai.com/v1/chat/completions"
+REQUEST_TIMEOUT_SECONDS = 60
+DEFAULT_TOPICS_PATH = get_config_dir() / "congressional_issues_comprehensive.json"
+
+
+class TopicExtractionError(RuntimeError):
+    """Raised when a topic extraction request or response is invalid."""
+
+
+@dataclass(frozen=True)
+class TopicCatalog:
+    """Loaded topic catalog with categories for prompting and flat candidates."""
+
+    topics_by_category: dict[str, list[str]]
+    candidate_topics: list[str]
+
+
+@dataclass(frozen=True)
+class TopicExtractionDiagnostics:
+    """Counts for the bill summary inputs needed by topic extraction."""
+
+    bill_rows: int
+    bill_text_rows: int
+    summarized_bill_text_rows: int
+    bills_with_summaries: int
+    bill_topic_rows: int
+    selected_bills: int
+
+
+@dataclass(frozen=True)
+class ExtractedBillTopic:
+    """One extracted bill topic and yes-vote stance."""
+
+    topic: str
+    support_position: BillTopicPosition
+    confidence: float | None = None
+    evidence: str | None = None
+
+
+def _select_bill_text_for_topic_extraction(bill: Bill) -> BillText | None:
+    """Pick one summarized bill_text row from the already-loaded relationship."""
+    for bill_text in bill.bill_texts:
+        if bill_text.summary and bill_text.summary.strip():
+            return bill_text
+    return None
+
+
+def normalize_topic_label(value: str) -> str:
+    """Normalize a topic label for storage, comparison, and de-duping."""
+    normalized = value.strip().strip("\"'")
+    normalized = normalized.strip().rstrip(".").strip()
+    return re.sub(r"\s+", " ", normalized).lower()
+
+
+def load_topic_catalog(path: Path | None = None) -> TopicCatalog:
+    """Load, validate, normalize, and flatten the bill topic catalog."""
+    topics_path = path or DEFAULT_TOPICS_PATH
+    try:
+        raw = json.loads(topics_path.read_text())
+    except FileNotFoundError as exc:
+        msg = f"Topic catalog not found: {topics_path}"
+        raise TopicExtractionError(msg) from exc
+    except json.JSONDecodeError as exc:
+        msg = f"Topic catalog is not valid JSON: {topics_path}: {exc}"
+        raise TopicExtractionError(msg) from exc
+
+    if not isinstance(raw, dict):
+        msg = "Topic catalog root must be an object mapping category names to lists"
+        raise TopicExtractionError(msg)
+
+    topics_by_category: dict[str, list[str]] = {}
+    candidate_topics: list[str] = []
+    seen_topics: set[str] = set()
+
+    for category, topics in raw.items():
+        if not isinstance(category, str) or not category.strip():
+            msg = "Topic catalog category names must be non-empty strings"
+            raise TopicExtractionError(msg)
+        if not isinstance(topics, list):
+            msg = f"Topic catalog category {category!r} must contain a list"
+            raise TopicExtractionError(msg)
+
+        normalized_topics: list[str] = []
+        for topic in topics:
+            if not isinstance(topic, str):
+                msg = f"Topic catalog category {category!r} contains a non-string topic"
+                raise TopicExtractionError(msg)
+            normalized_topic = normalize_topic_label(topic)
+            if not normalized_topic:
+                msg = f"Topic catalog category {category!r} contains a blank topic"
+                raise TopicExtractionError(msg)
+            if normalized_topic in seen_topics:
+                continue
+            seen_topics.add(normalized_topic)
+            normalized_topics.append(normalized_topic)
+            candidate_topics.append(normalized_topic)
+
+        topics_by_category[category.strip()] = normalized_topics
+
+    return TopicCatalog(
+        topics_by_category=topics_by_category,
+        candidate_topics=candidate_topics,
+    )
+
+
+def build_topic_extraction_messages(
+    *,
+    bill: Bill,
+    bill_text: str,
+    candidate_topics: Sequence[str],
+) -> list[dict[str, str]]:
+    """Build GPT messages for extracting a bill's scored topics."""
+    normalized_candidates = [normalize_topic_label(topic) for topic in candidate_topics]
+    candidate_list = "\n".join(f"- {topic}" for topic in normalized_candidates)
+    metadata = "\n".join(
+        (
+            f"Congress: {bill.congress}",
+            f"Bill: {bill.bill_type} {bill.number}",
+            f"Title: {bill.title_short or bill.title or bill.official_title or ''}",
+            f"Top subject term: {bill.subjects_top_term or ''}",
+        )
+    )
+
+    system_prompt = (
+        "You extract policy topics from U.S. congressional bills.\n"
+        'For each selected topic, decide whether a Yes/Yea vote on the bill is "for" or "against" that topic.\n'
+        'Use "support_position": "for" when a Yes/Yea vote advances or supports the topic.\n'
+        'Use "support_position": "against" when a Yes/Yea vote restricts, repeals, blocks, or opposes the topic.\n'
+        "Select only topics from the provided candidate topic list.\n"
+        "Omit topics that are not materially addressed by the bill.\n"
+        "Return strict JSON only, with this shape:\n"
+        '{"topics":[{"topic":"candidate topic","support_position":"for","confidence":0.0,"evidence":"short reason"}]}'
+    )
+    user_prompt = "\n\n".join(
+        (
+            "BILL METADATA:",
+            metadata,
+            "CANDIDATE TOPICS:",
+            candidate_list,
+            "BILL TEXT:",
+            bill_text,
+        )
+    )
+    return [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": user_prompt},
+    ]
+
+
+def call_openai_topic_extraction(
+    *,
+    openai_config: OpenAIConfig,
+    messages: list[dict[str, str]],
+) -> str:
+    """Call GPT and return the assistant message content."""
+
+    response = httpx.post(
+        openai_config.openai_chat_completions_url,
+        headers={
+            "Authorization": f"Bearer {openai_config.api_key}",
+            "OpenAI-Project": openai_config.openai_project_id,
+            "Content-Type": "application/json",
+        },
+        json={
+            "model": "gpt-5.4-mini",
+            "messages": messages,
+        },
+        timeout=openai_config.timeout_seconds,
+    )
+    response.raise_for_status()
+    return extract_message_content(response.json())
+
+
+def extract_message_content(data: dict[str, Any]) -> str:
+    """Extract message content from a chat-completions response body."""
+    choices = data.get("choices")
+    if not isinstance(choices, list) or not choices:
+        msg = "Chat completion response did not contain choices"
+        raise TopicExtractionError(msg)
+
+    first = choices[0]
+    if not isinstance(first, dict):
+        msg = "Chat completion choice must be an object"
+        raise TopicExtractionError(msg)
+
+    message = first.get("message")
+    if isinstance(message, dict) and isinstance(message.get("content"), str):
+        return message["content"]
+    if isinstance(first.get("text"), str):
+        return first["text"]
+
+    msg = "Chat completion response did not contain message content"
+    raise TopicExtractionError(msg)
+
+
+def parse_topic_extraction_response(response_text: str) -> list[ExtractedBillTopic]:
+    """Parse, normalize, validate, and de-dupe a topic extraction response."""
+    payload = _load_json_response(response_text)
+    topics = payload.get("topics")
+    if not isinstance(topics, list):
+        msg = "Topic extraction response must contain a topics list"
+        raise TopicExtractionError(msg)
+
+    deduped: dict[tuple[str, BillTopicPosition], ExtractedBillTopic] = {}
+    for item in topics:
+        if not isinstance(item, dict):
+            msg = "Topic extraction response topics must be objects"
+            raise TopicExtractionError(msg)
+
+        raw_topic = _extract_topic_label(item)
+        topic = normalize_topic_label(raw_topic)
+        if not topic:
+            msg = "Topic extraction response topic must not be blank"
+            raise TopicExtractionError(msg)
+
+        raw_position = item.get("support_position")
+        try:
+            support_position = BillTopicPosition(raw_position)
+        except ValueError as exc:
+            msg = f"Invalid support_position: {raw_position!r}"
+            raise TopicExtractionError(msg) from exc
+
+        confidence = _parse_confidence(item.get("confidence"))
+        evidence = item.get("evidence")
+        if evidence is not None and not isinstance(evidence, str):
+            evidence = str(evidence)
+
+        extracted = ExtractedBillTopic(
+            topic=topic,
+            support_position=support_position,
+            confidence=confidence,
+            evidence=evidence,
+        )
+        key = (topic, support_position)
+        existing = deduped.get(key)
+        if existing is None or _confidence_rank(extracted) > _confidence_rank(existing):
+            deduped[key] = extracted
+
+    return list(deduped.values())
+
+
+def extract_topics_for_bill_text(
+    *,
+    openai_config: OpenAIConfig,
+    bill: Bill,
+    text: str,
+    candidate_topics: Sequence[str],
+) -> list[ExtractedBillTopic]:
+    """Extract accepted catalog topics for a bill text string."""
+    normalized_candidates = {normalize_topic_label(topic) for topic in candidate_topics}
+    messages = build_topic_extraction_messages(
+        bill=bill,
+        bill_text=text,
+        candidate_topics=sorted(normalized_candidates),
+    )
+    response_text = call_openai_topic_extraction(
+        openai_config=openai_config,
+        messages=messages,
+    )
+    extracted_topics = parse_topic_extraction_response(response_text)
+    return [topic for topic in extracted_topics if topic.topic in normalized_candidates]
+
+
+def store_bill_topic_result(
+    *,
+    session: Session,
+    bill: Bill,
+    topics: Sequence[ExtractedBillTopic],
+    replace_existing: bool = True,
+) -> None:
+    """Store extracted topics for one bill."""
+    if replace_existing:
+        session.execute(delete(BillTopic).where(BillTopic.bill_id == bill.id))
+
+    for topic in topics:
+        session.add(
+            BillTopic(
+                bill_id=bill.id,
+                topic=normalize_topic_label(topic.topic),
+                support_position=topic.support_position,
+            )
+        )
+
+
+def create_select_bills_for_topic_extraction(
+    congress: int | None = None,
+    bill_ids: list[int] | None = None,
+    bill_text_ids: list[int] | None = None,
+    with_votes_only: bool = False,
+    force: bool = False,
+    limit: int | None = None,
+) -> Select[tuple[Bill]]:
+    """Select bill rows that have summarized bill_text rows for topic extraction."""
+    has_summary = (BillText.summary.is_not(None), BillText.summary != "")
+    summarized_text_filters: list[ColumnElement[bool]] = [
+        BillText.bill_id == Bill.id,
+        *has_summary,
+    ]
+    if with_votes_only:
+        summarized_text_filters.append(
+            exists(
+                select(VoteTextTarget.vote_id)
+                .join(
+                    VoteClassification,
+                    VoteClassification.vote_id == VoteTextTarget.vote_id,
+                )
+                .where(
+                    VoteTextTarget.voted_text_version_id == BillText.id,
+                    VoteClassification.subject_type == SubjectType.MEASURE,
+                    VoteClassification.vote_relationship
+                    == VoteRelationship.DIRECT_TEXT_VOTE,
+                    VoteClassification.is_direct_vote_on_legislative_text.is_(True),
+                    VoteClassification.is_substantive_policy_vote.is_(True),
+                    VoteClassification.is_special_rule.is_(False),
+                )
+            )
+        )
+    summarized_text_exists = exists(select(BillText.id).where(*summarized_text_filters))
+    stmt = (
+        select(Bill)
+        .where(summarized_text_exists)
+        .options(selectinload(Bill.bill_texts.and_(*summarized_text_filters[1:])))
+        .order_by(Bill.id)
+    )
+    if congress is not None:
+        stmt = stmt.where(Bill.congress == congress)
+    if bill_ids:
+        stmt = stmt.where(Bill.id.in_(bill_ids))
+    if bill_text_ids:
+        selected_text_exists = exists(
+            select(BillText.id).where(
+                BillText.bill_id == Bill.id,
+                BillText.id.in_(bill_text_ids),
+                *summarized_text_filters[1:],
+            )
+        )
+        stmt = stmt.where(selected_text_exists)
+    if not force:
+        stmt = stmt.where(
+            ~exists(select(BillTopic.id).where(BillTopic.bill_id == Bill.id))
+        )
+    if limit is not None:
+        stmt = stmt.limit(limit)
+    return stmt
+
+
+def collect_topic_extraction_diagnostics(
+    session: Session,
+    *,
+    congress: int | None = None,
+    bill_ids: list[int] | None = None,
+    bill_text_ids: list[int] | None = None,
+    with_votes_only: bool = False,
+    force: bool = False,
+    limit: int | None = None,
+) -> TopicExtractionDiagnostics:
+    """Count topic extraction inputs for explaining empty selections."""
+    bill_filters = []
+    bill_text_filters: list[ColumnElement[bool]] = []
+    if congress is not None:
+        bill_filters.append(Bill.congress == congress)
+    if bill_ids:
+        bill_filters.append(Bill.id.in_(bill_ids))
+        bill_text_filters.append(BillText.bill_id.in_(bill_ids))
+    if bill_text_ids:
+        bill_text_filters.append(BillText.id.in_(bill_text_ids))
+    if with_votes_only:
+        bill_text_filters.append(
+            exists(
+                select(VoteTextTarget.vote_id)
+                .join(
+                    VoteClassification,
+                    VoteClassification.vote_id == VoteTextTarget.vote_id,
+                )
+                .where(
+                    VoteTextTarget.voted_text_version_id == BillText.id,
+                    VoteClassification.subject_type == SubjectType.MEASURE,
+                    VoteClassification.vote_relationship
+                    == VoteRelationship.DIRECT_TEXT_VOTE,
+                    VoteClassification.is_direct_vote_on_legislative_text.is_(True),
+                    VoteClassification.is_substantive_policy_vote.is_(True),
+                    VoteClassification.is_special_rule.is_(False),
+                )
+            )
+        )
+
+    has_summary = (BillText.summary.is_not(None), BillText.summary != "")
+    summary_filters = [*bill_text_filters, *has_summary]
+
+    bills_with_summaries = session.scalar(
+        select(func.count(func.distinct(Bill.id)))
+        .select_from(Bill)
+        .join(BillText, BillText.bill_id == Bill.id)
+        .where(*bill_filters, *summary_filters)
+    )
+    selected_bills = session.scalar(
+        select(func.count()).select_from(
+            create_select_bills_for_topic_extraction(
+                congress=congress,
+                bill_ids=bill_ids,
+                bill_text_ids=bill_text_ids,
+                with_votes_only=with_votes_only,
+                force=force,
+                limit=limit,
+            ).subquery()
+        )
+    )
+
+    return TopicExtractionDiagnostics(
+        bill_rows=session.scalar(select(func.count(Bill.id)).where(*bill_filters)) or 0,
+        bill_text_rows=_count_bill_texts(
+            session,
+            bill_filters=bill_filters,
+            bill_text_filters=bill_text_filters,
+        ),
+        summarized_bill_text_rows=_count_bill_texts(
+            session,
+            bill_filters=bill_filters,
+            bill_text_filters=summary_filters,
+        ),
+        bills_with_summaries=bills_with_summaries or 0,
+        bill_topic_rows=session.scalar(select(func.count(BillTopic.id))) or 0,
+        selected_bills=selected_bills or 0,
+    )
+
+
+def _load_json_response(response_text: str) -> dict[str, Any]:
+    text = response_text.strip()
+    fenced = re.fullmatch(r"```(?:json)?\s*(.*?)\s*```", text, flags=re.DOTALL)
+    if fenced:
+        text = fenced.group(1).strip()
+
+    try:
+        payload = json.loads(text)
+    except json.JSONDecodeError as exc:
+        msg = f"Topic extraction response is not valid JSON: {exc}"
+        raise TopicExtractionError(msg) from exc
+    if not isinstance(payload, dict):
+        msg = "Topic extraction response must be a JSON object"
+        raise TopicExtractionError(msg)
+    return payload
+
+
+def _parse_confidence(raw: Any) -> float | None:
+    if raw is None:
+        return None
+    try:
+        return float(raw)
+    except (TypeError, ValueError) as exc:
+        msg = f"Invalid confidence: {raw!r}"
+        raise TopicExtractionError(msg) from exc
+
+
+def _confidence_rank(topic: ExtractedBillTopic) -> tuple[int, float]:
+    if topic.confidence is None:
+        return (0, 0.0)
+    return (1, topic.confidence)
+
+
+def _extract_topic_label(item: dict[str, Any]) -> str:
+    raw_topic = item.get("topic")
+    if isinstance(raw_topic, str):
+        return raw_topic
+    if isinstance(raw_topic, dict):
+        for key in ("topic", "label", "name", "title"):
+            value = raw_topic.get(key)
+            if isinstance(value, str):
+                return value
+
+    msg = "Topic extraction response topic must be a string"
+    raise TopicExtractionError(msg)
+
+
+def _count_bill_texts(
+    session: Session,
+    *,
+    bill_filters: Sequence[ColumnElement[bool]],
+    bill_text_filters: Sequence[ColumnElement[bool]],
+) -> int:
+    stmt = select(func.count(BillText.id))
+    if bill_filters:
+        stmt = stmt.join(Bill, Bill.id == BillText.bill_id).where(*bill_filters)
+    return session.scalar(stmt.where(*bill_text_filters)) or 0
+
+
+def main(
+    topics_path: Annotated[
+        Path, typer.Option(help="Path to congressional issue topic JSON.")
+    ] = DEFAULT_TOPICS_PATH,
+    congress: Annotated[
+        int | None, typer.Option(help="Only process one Congress.")
+    ] = None,
+    bill_ids: Annotated[
+        list[int] | None,
+        typer.Option(
+            "--bill-id",
+            help="Only process one internal bill.id. Repeat for multiple bills.",
+        ),
+    ] = None,
+    bill_text_ids: Annotated[
+        list[int] | None,
+        typer.Option(
+            "--bill-text-id",
+            help="Only process one internal bill_text.id. Repeat for multiple rows.",
+        ),
+    ] = None,
+    with_votes_only: Annotated[
+        bool,
+        typer.Option(
+            "--with-votes-only",
+            help="Only process summarized bill_text rows linked to at least one vote.",
+        ),
+    ] = True,
+    limit: Annotated[int | None, typer.Option(help="Maximum rows to process.")] = None,
+    force: Annotated[
+        bool,
+        typer.Option(help="Regenerate topics for bills that already have topics."),
+    ] = False,
+    dry_run: Annotated[
+        bool,
+        typer.Option(help="Select bills and print diagnostics without calling OpenAI."),
+    ] = False,
+    diagnose: Annotated[
+        bool,
+        typer.Option(help="Log input-stage counts before processing."),
+    ] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+) -> None:
+    """CLI entrypoint for generating and storing bill topics."""
+    logging.basicConfig(
+        level=log_level,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+
+    topic_catalog = load_topic_catalog(topics_path)
+    logger.info(
+        "Loaded %d candidate topics from %s",
+        len(topic_catalog.candidate_topics),
+        topics_path,
+    )
+
+    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
+    with Session(engine) as session:
+        if diagnose or dry_run:
+            diagnostics = collect_topic_extraction_diagnostics(
+                session,
+                congress=congress,
+                bill_ids=bill_ids,
+                bill_text_ids=bill_text_ids,
+                with_votes_only=with_votes_only,
+                force=force,
+                limit=limit,
+            )
+            _log_topic_extraction_diagnostics(diagnostics)
+            if dry_run:
+                return
+
+        openai_config = get_openai_config()
+
+        stmt = create_select_bills_for_topic_extraction(
+            congress=congress,
+            bill_ids=bill_ids,
+            bill_text_ids=bill_text_ids,
+            with_votes_only=with_votes_only,
+            force=force,
+            limit=limit,
+        )
+        bills = session.scalars(stmt).all()
+        logger.info("Selected %d bills for topic extraction", len(bills))
+
+        written = 0
+        failed = 0
+        for index, bill in enumerate(bills, 1):
+            bill_text = _select_bill_text_for_topic_extraction(bill)
+            if bill_text is None:
+                logger.warning("Skipping bill id=%s: no usable summary", bill.id)
+                continue
+            summary = bill_text.summary.strip()
+
+            try:
+                extracted_topics = extract_topics_for_bill_text(
+                    openai_config=openai_config,
+                    bill=bill,
+                    text=summary,
+                    candidate_topics=topic_catalog.candidate_topics,
+                )
+            except (httpx.HTTPError, TopicExtractionError):
+                failed += 1
+                logger.exception(
+                    "Skipping bill id=%s after topic extraction failure", bill.id
+                )
+                continue
+
+            store_bill_topic_result(
+                session=session,
+                bill=bill,
+                topics=extracted_topics,
+                replace_existing=True,
+            )
+            written += 1
+            if index % 100 == 0:
+                session.commit()
+            logger.info(
+                "Stored %d topics for bill id=%s",
+                len(extracted_topics),
+                bill.id,
+            )
+
+        session.commit()
+        logger.info(
+            "Done: stored topic results for %d bills; failed %d bills",
+            written,
+            failed,
+        )
+
+
+def _log_topic_extraction_diagnostics(
+    diagnostics: TopicExtractionDiagnostics,
+) -> None:
+    logger.info(
+        "Topic extraction diagnostics: bill_rows=%d bill_text_rows=%d "
+        "summarized_bill_text_rows=%d bills_with_summaries=%d "
+        "bill_topic_rows=%d selected_bills=%d",
+        diagnostics.bill_rows,
+        diagnostics.bill_text_rows,
+        diagnostics.summarized_bill_text_rows,
+        diagnostics.bills_with_summaries,
+        diagnostics.bill_topic_rows,
+        diagnostics.selected_bills,
+    )
+    if diagnostics.bill_rows == 0:
+        logger.warning("No bills matched the topic extraction scope.")
+    elif diagnostics.bill_text_rows == 0:
+        logger.warning("No bill_text rows matched the topic extraction scope.")
+    elif diagnostics.summarized_bill_text_rows == 0:
+        logger.warning(
+            "No summarized bill_text rows matched the topic extraction scope. "
+            "Run pipelines.tools.summarize_bills first."
+        )
+    elif diagnostics.selected_bills == 0 and diagnostics.bill_topic_rows > 0:
+        logger.warning(
+            "No bills selected because matching bills already have topics. "
+            "Use --force to regenerate them."
+        )
+    elif diagnostics.selected_bills == 0:
+        logger.warning("No bills selected for topic extraction.")
+
+
+if __name__ == "__main__":
+    typer.run(main)

pipelines/jobs/summarize_bills.py

@@ -0,0 +1,309 @@
+"""Summarize bill_text rows with GPT-5 and store results in the database."""
+
+from __future__ import annotations
+
+import logging
+import tomllib
+from os import getenv
+from typing import Annotated, Any
+
+import httpx
+import typer
+from sqlalchemy import Select, exists, or_, select
+from sqlalchemy.orm import Session, selectinload
+
+from tiktoken import get_encoding
+
+
+from pipelines.config import get_config_dir
+from pipelines.orm.common import get_postgres_engine
+from pipelines.orm.data_science_dev.congress import (
+    Bill,
+    BillText,
+    SubjectType,
+    VoteClassification,
+    VoteRelationship,
+    VoteTextTarget,
+)
+from pipelines.tools.bill_token_compression import compress_bill_text
+
+logger = logging.getLogger(__name__)
+
+OPENAI_CHAT_COMPLETIONS_URL = "https://api.openai.com/v1/chat/completions"
+OPENAI_PROJECT_ID = "proj_fQBPEXFgnS87Fk6wZwploFwE"
+REQUEST_TIMEOUT_SECONDS = 60
+
+
+def load_summarization_prompts(
+    section: str = "summarization",
+) -> dict[str, str]:
+    summarization_prompts = get_config_dir() / "prompts" / "summarization_prompts.toml"
+
+    return tomllib.loads(summarization_prompts.read_text())[section]
+
+
+class BillSummaryError(RuntimeError):
+    """Raised when a bill summary request or response is invalid."""
+
+
+def call_openai_summary(
+    *,
+    model: str,
+    messages: list[dict[str, str]],
+) -> str:
+    """Call GPT and return the assistant message content."""
+    api_key = getenv("CLOSEDAI_TOKEN")
+    if not api_key:
+        msg = "CLOSEDAI_TOKEN is required"
+        raise BillSummaryError(msg)
+
+    response = httpx.post(
+        OPENAI_CHAT_COMPLETIONS_URL,
+        headers={
+            "Authorization": f"Bearer {api_key}",
+            "OpenAI-Project": OPENAI_PROJECT_ID,
+            "Content-Type": "application/json",
+        },
+        json={
+            "model": model,
+            "messages": messages,
+        },
+        timeout=REQUEST_TIMEOUT_SECONDS,
+    )
+    logger.info(f"{response.text=}")
+    response.raise_for_status()
+    return extract_message_content(response.json())
+
+
+def build_bill_summary_messages(
+    *,
+    bill_text: BillText,
+    summarization_prompts: dict[str, str],
+) -> list[dict[str, str]]:
+    """Build the GPT prompt messages plus compressed text and user prompt."""
+    if not bill_text.text_content:
+        msg = f"bill_text id={bill_text.id} has no text_content"
+        raise BillSummaryError(msg)
+
+    compressed_text = compress_bill_text(bill_text.text_content)
+    if not compressed_text:
+        msg = f"bill_text id={bill_text.id} has no summarizable text_content"
+        raise BillSummaryError(msg)
+
+    user_prompt = summarization_prompts["user_template"].format(
+        text_content=compressed_text
+    )
+
+    user_prompt_tokens = len(get_encoding("o200k_base").encode(user_prompt))
+    logger.info(f"{user_prompt_tokens=}")
+
+    messages = [
+        {"role": "system", "content": summarization_prompts["system_prompt"]},
+        {
+            "role": "user",
+            "content": user_prompt,
+        },
+    ]
+    return messages, user_prompt_tokens
+
+
+def summarize_bill_text(
+    *,
+    model: str,
+    bill_text: BillText,
+    summarization_prompts: dict[str, str],
+) -> str:
+    """Generate and return a summary for one bill_text row."""
+    messages, user_prompt_tokens = build_bill_summary_messages(
+        bill_text=bill_text,
+        summarization_prompts=summarization_prompts,
+    )
+    # This may only be for gpt-5.4 mini I need to read the docs
+    if user_prompt_tokens > 272000:
+        msg = f"Compressed bill_text id={bill_text.id} is too long for summarization ({user_prompt_tokens} tokens)"
+        logger.warning(msg)
+        return None
+
+    summary = call_openai_summary(
+        model=model,
+        messages=messages,
+    ).strip()
+    if not summary:
+        msg = f"Model returned an empty summary for bill_text id={bill_text.id}"
+        raise BillSummaryError(msg)
+    return summary
+
+
+def store_bill_summary_result(
+    *,
+    bill_text: BillText,
+    summary: str,
+    model: str,
+) -> None:
+    """Store a generated summary and the prompt/model metadata that produced it."""
+    bill_text.summary = summary
+    bill_text.summarization_model = model
+    bill_text.summarization_system_prompt_version = "v1.2"
+    bill_text.summarization_user_prompt_version = "v1"
+
+
+def create_select_bill_texts_for_summarization(
+    congress: int | None = None,
+    bill_ids: list[int] | None = None,
+    bill_text_ids: list[int] | None = None,
+    with_votes_only: bool = False,
+    force: bool = False,
+    limit: int | None = None,
+) -> Select:
+    """Select bill_text rows that have source text and need summaries."""
+    stmt = (
+        select(BillText)
+        .join(Bill, Bill.id == BillText.bill_id)
+        .where(BillText.text_content.is_not(None), BillText.text_content != "")
+        .options(selectinload(BillText.bill))
+        .order_by(BillText.id)
+    )
+    if congress is not None:
+        stmt = stmt.where(Bill.congress == congress)
+    if bill_ids:
+        stmt = stmt.where(BillText.bill_id.in_(bill_ids))
+    if bill_text_ids:
+        stmt = stmt.where(BillText.id.in_(bill_text_ids))
+    if with_votes_only:
+        stmt = stmt.where(
+            exists(
+                select(VoteTextTarget.vote_id)
+                .join(
+                    VoteClassification,
+                    VoteClassification.vote_id == VoteTextTarget.vote_id,
+                )
+                .where(
+                    VoteTextTarget.voted_text_version_id == BillText.id,
+                    VoteClassification.subject_type == SubjectType.MEASURE,
+                    VoteClassification.vote_relationship
+                    == VoteRelationship.DIRECT_TEXT_VOTE,
+                    VoteClassification.is_direct_vote_on_legislative_text.is_(True),
+                    VoteClassification.is_substantive_policy_vote.is_(True),
+                    VoteClassification.is_special_rule.is_(False),
+                )
+            )
+        )
+    if not force:
+        stmt = stmt.where(or_(BillText.summary.is_(None), BillText.summary == ""))
+    if limit is not None:
+        stmt = stmt.limit(limit)
+    return stmt
+
+
+def extract_message_content(data: dict[str, Any]) -> str:
+    """Extract message content from a chat-completions response body."""
+    choices = data.get("choices")
+    if not isinstance(choices, list) or not choices:
+        msg = "Chat completion response did not contain choices"
+        raise BillSummaryError(msg)
+
+    first = choices[0]
+    if not isinstance(first, dict):
+        msg = "Chat completion choice must be an object"
+        raise BillSummaryError(msg)
+
+    message = first.get("message")
+    if isinstance(message, dict) and isinstance(message.get("content"), str):
+        return message["content"]
+    if isinstance(first.get("text"), str):
+        return first["text"]
+
+    msg = "Chat completion response did not contain message content"
+    raise BillSummaryError(msg)
+
+
+def main(
+    model: Annotated[str, typer.Option(help="OpenAI model id.")] = "gpt-5.4-mini",
+    congress: Annotated[
+        int | None, typer.Option(help="Only process one Congress.")
+    ] = None,
+    bill_ids: Annotated[
+        list[int] | None,
+        typer.Option(
+            "--bill-id",
+            help="Only process one internal bill.id. Repeat for multiple bills.",
+        ),
+    ] = None,
+    bill_text_ids: Annotated[
+        list[int] | None,
+        typer.Option(
+            "--bill-text-id",
+            help="Only process one internal bill_text.id. Repeat for multiple rows.",
+        ),
+    ] = None,
+    with_votes_only: Annotated[
+        bool,
+        typer.Option(
+            "--with-votes-only",
+            help="Only process bill_text rows linked to at least one vote.",
+        ),
+    ] = False,
+    limit: Annotated[int | None, typer.Option(help="Maximum rows to process.")] = None,
+    force: Annotated[
+        bool,
+        typer.Option(help="Regenerate summaries for rows that already have a summary."),
+    ] = False,
+    dry_run: Annotated[
+        bool, typer.Option(help="Print summaries without writing them to the database.")
+    ] = False,
+    log_level: Annotated[str, typer.Option(help="Log level.")] = "INFO",
+) -> None:
+    """CLI entrypoint for generating and storing bill summaries."""
+    logging.basicConfig(
+        level=log_level,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+    if not getenv("CLOSEDAI_TOKEN"):
+        message = "CLOSEDAI_TOKEN is required"
+        raise typer.BadParameter(message)
+
+    summarization_prompts = load_summarization_prompts()
+    engine = get_postgres_engine(name="DATA_SCIENCE_DEV")
+    with Session(engine) as session:
+        stmt = create_select_bill_texts_for_summarization(
+            congress=congress,
+            bill_ids=bill_ids,
+            bill_text_ids=bill_text_ids,
+            with_votes_only=with_votes_only,
+            force=force,
+            limit=limit,
+        )
+        bill_texts = session.scalars(stmt).all()
+        logger.info("Selected %d bill_text rows for summarization", len(bill_texts))
+
+        written = 0
+        for index, bill_text in enumerate(bill_texts, 1):
+            summary = summarize_bill_text(
+                model=model,
+                bill_text=bill_text,
+                summarization_prompts=summarization_prompts,
+            )
+            if summary is None:
+                logger.warning("Skipping bill_text id=%s", bill_text.id)
+                continue
+            store_bill_summary_result(
+                bill_text=bill_text,
+                summary=summary,
+                model=model,
+            )
+            if index % 100 == 0:
+                session.commit()
+            written += 1
+            session.commit()
+            logger.info("Stored summary for bill_text id=%s", bill_text.id)
+
+        logger.info("Done: stored %d summaries", written)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

pipelines/tools/Dockerfile.finetune

@@ -1,25 +0,0 @@
-# Unsloth fine-tuning container for Qwen 3.5 4B on RTX 3090.
-#
-# Build:
-#   docker build -f python/prompt_bench/Dockerfile.finetune -t bill-finetune .
-#
-# Run:
-#   docker run --rm --device=nvidia.com/gpu=all --ipc=host \
-#     -v $(pwd)/output:/workspace/output \
-#     -v $(pwd)/output/finetune_dataset.jsonl:/workspace/dataset.jsonl:ro \
-#     -v /zfs/models/hf:/models \
-#     bill-finetune \
-#     --dataset /workspace/dataset.jsonl \
-#     --output-dir /workspace/output/qwen-bill-summarizer
-
-FROM ghcr.io/unslothai/unsloth:latest
-
-RUN pip install --no-cache-dir typer
-
-WORKDIR /workspace
-COPY python/prompt_bench/finetune.py python/prompt_bench/finetune.py
-COPY config/prompts/summarization_prompts.toml config/prompts/summarization_prompts.toml
-COPY python/prompt_bench/__init__.py python/prompt_bench/__init__.py
-COPY python/__init__.py python/__init__.py
-
-ENTRYPOINT ["python", "-m", "pipelines.prompt_bench.finetune"]

pipelines/tools/batch_bill_summarizer.py

@@ -23,14 +23,10 @@ import httpx
 import typer
 from tiktoken import Encoding, get_encoding
 
+from pipelines.config import get_config_dir
 from pipelines.tools.bill_token_compression import compress_bill_text
 
-_PROMPTS_PATH = (
-    Path(__file__).resolve().parents[2]
-    / "config"
-    / "prompts"
-    / "summarization_prompts.toml"
-)
+_PROMPTS_PATH = get_config_dir() / "prompts" / "summarization_prompts.toml"
 _PROMPTS = tomllib.loads(_PROMPTS_PATH.read_text())["summarization"]
 SUMMARIZATION_SYSTEM_PROMPT: str = _PROMPTS["system_prompt"]
 SUMMARIZATION_USER_TEMPLATE: str = _PROMPTS["user_template"]

pipelines/tools/compresion_test.py

@@ -24,14 +24,10 @@ from typing import Annotated
 import httpx
 import typer
 
+from pipelines.config import get_config_dir
 from pipelines.tools.bill_token_compression import compress_bill_text
 
-_PROMPTS_PATH = (
-    Path(__file__).resolve().parents[2]
-    / "config"
-    / "prompts"
-    / "summarization_prompts.toml"
-)
+_PROMPTS_PATH = get_config_dir() / "prompts" / "summarization_prompts.toml"
 _PROMPTS = tomllib.loads(_PROMPTS_PATH.read_text())["summarization"]
 SUMMARIZATION_SYSTEM_PROMPT: str = _PROMPTS["system_prompt"]
 SUMMARIZATION_USER_TEMPLATE: str = _PROMPTS["user_template"]

pipelines/tools/finetune.py

@@ -25,6 +25,8 @@ from datasets import Dataset
 from transformers import TrainingArguments
 from trl import SFTTrainer
 
+from pipelines.config import default_config_path
+
 logger = logging.getLogger(__name__)
 
 
@@ -123,7 +125,7 @@ def main(
     config_path: Annotated[
         Path,
         typer.Option("--config", help="TOML config file"),
-    ] = Path(__file__).parent / "config.toml",
+    ] = default_config_path(),
     save_gguf: Annotated[
         bool, typer.Option("--save-gguf/--no-save-gguf", help="Also save GGUF")
     ] = False,

pipelines/tools/prompt_bench.py

@@ -11,8 +11,8 @@ from typing import Annotated
 
 import typer
 
-from pipelines.tools.containers.lib import check_gpu_free
-from pipelines.tools.containers.vllm import start_vllm, stop_vllm
+from pipelines.containers.lib import check_gpu_free
+from pipelines.containers.vllm import start_vllm, stop_vllm
 from pipelines.tools.downloader import is_model_present
 from pipelines.tools.models import BenchmarkConfig
 from pipelines.tools.vllm_client import VLLMClient