added llm_tool_calling.py

This commit is contained in:
2026-06-06 17:43:32 -04:00
parent f20bee82ec
commit 5611daab97
2 changed files with 593 additions and 542 deletions
+28 -542
View File
@@ -4,30 +4,35 @@ from __future__ import annotations
import json
import re
import time
from dataclasses import asdict, dataclass, is_dataclass, replace
from os import PathLike
from typing import TYPE_CHECKING
import httpx
from sqlalchemy import or_, select
from sqlalchemy.orm import Session
from python.common import utcnow
from python.orm.richie import Audiobook, AudiobookAuthor, AudiobookSeries
from python.tools.audiobook.llm_tool_calling import (
CatalogToolRegistry,
MetadataResolutionError,
normalize_title_slug,
optional_int,
parse_tool_calls,
required_int,
required_string,
run_tool_calls,
validate_catalog_slug,
validate_title_slug,
)
if TYPE_CHECKING:
from pathlib import Path
from sqlalchemy.engine import Engine
CATALOG_SLUG_PATTERN = re.compile(r"^[a-z0-9]+(?:_[a-z0-9]+)*$")
from python.orm.richie import AudiobookAuthor
FENCED_JSON_PATTERN = re.compile(r"^```(?:json)?\s*(?P<json>.*?)\s*```$", re.IGNORECASE | re.DOTALL)
TITLE_SLUG_PATTERN = re.compile(r"^[a-z0-9]+(?:-[a-z0-9]+)*$")
class MetadataResolutionError(ValueError):
"""Metadata resolution failed validation."""
@dataclass(frozen=True)
@@ -91,14 +96,6 @@ class ResolvedBookFields:
series_index: int
@dataclass(frozen=True)
class EnsuredBook:
"""Book row plus whether it was created."""
book: Audiobook
action: str
@dataclass(frozen=True)
class AgentStepResult:
"""Outcome from one model response."""
@@ -118,7 +115,7 @@ def standard_book_metadata(
) -> StandardBookMetadata:
"""Resolve canonical audiobook metadata with the configured Ollama Cloud model."""
with Session(engine) as session:
registry = CatalogToolRegistry(session, log_path, config)
registry = CatalogToolRegistry(session, log_path, config, write_agent_log)
agent = AudiobookMetadataAgent(
registry=registry, log_path=log_path, ollama_api_key=ollama_api_key, config=config
)
@@ -126,388 +123,15 @@ def standard_book_metadata(
if metadata.needs_review:
session.rollback()
else:
registry.prune_unused_created_rows(metadata)
registry.prune_unused_created_rows(
author_id=metadata.author_id,
book_id=metadata.book_id,
series_id=metadata.series_id,
)
session.commit()
return metadata
class CatalogToolRegistry:
"""Controlled catalog tools exposed to the metadata model."""
def __init__(self, session: Session, log_path: Path, config: AgentConfig) -> None:
"""Create a registry bound to one database session and audit log."""
self._session = session
self._log_path = log_path
self._config = config
self.seen_author_ids: set[int] = set()
self.seen_series_ids: set[int] = set()
self.seen_book_ids: set[int] = set()
self.created_author_ids: set[int] = set()
self.created_series_ids: set[int] = set()
self.created_book_ids: set[int] = set()
def tool_schemas(self) -> list[dict[str, object]]:
"""Return Ollama tool schemas."""
schemas = [
{
"type": "function",
"function": {
"name": "search_authors",
"description": "Search canonical audiobook authors by slug or noisy source text.",
"parameters": {
"type": "object",
"properties": {"query": {"type": "string"}},
"required": ["query"],
},
},
},
{
"type": "function",
"function": {
"name": "search_series",
"description": "Search canonical audiobook series by slug or noisy source text.",
"parameters": {
"type": "object",
"properties": {
"query": {"type": "string"},
"author_id": {"type": ["integer", "null"]},
},
"required": ["query"],
},
},
},
{
"type": "function",
"function": {
"name": "search_books",
"description": "Search canonical audiobook titles with optional author and series filters.",
"parameters": {
"type": "object",
"properties": {
"query": {"type": "string"},
"author_id": {"type": ["integer", "null"]},
"series_id": {"type": ["integer", "null"]},
},
"required": ["query"],
},
},
},
{
"type": "function",
"function": {
"name": "ensure_author",
"description": "Normalize an author name to a catalog slug, then return or create that author.",
"parameters": {
"type": "object",
"properties": {"name": {"type": "string"}},
"required": ["name"],
},
},
},
{
"type": "function",
"function": {
"name": "ensure_series",
"description": "Normalize a series name to a catalog slug, then return or create it for an author.",
"parameters": {
"type": "object",
"properties": {
"name": {"type": "string"},
"author_id": {"type": "integer"},
},
"required": ["name", "author_id"],
},
},
},
{
"type": "function",
"function": {
"name": "ensure_book",
"description": "Normalize a title to a book slug, then return or create it for an author/series.",
"parameters": {
"type": "object",
"properties": {
"title": {"type": "string"},
"author_id": {"type": "integer"},
"series_id": {"type": ["integer", "null"]},
"series_index": {"type": "integer"},
},
"required": ["title", "author_id", "series_id", "series_index"],
},
},
},
]
enabled_tool_names = set(self._config.tool_names)
return [schema for schema in schemas if schema["function"]["name"] in enabled_tool_names]
def run(self, name: str, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Run a single read-only catalog tool."""
handlers = {
"search_authors": self.run_search_authors,
"search_series": self.run_search_series,
"search_books": self.run_search_books,
"ensure_author": self.run_ensure_author,
"ensure_series": self.run_ensure_series,
"ensure_book": self.run_ensure_book,
}
handler = handlers.get(name)
if handler is None:
write_agent_log(self._log_path, "tool_error", tool=name, arguments=arguments, error="unknown_tool")
msg = f"Unknown audiobook metadata tool: {name}"
raise MetadataResolutionError(msg)
if name not in self._config.tool_names:
write_agent_log(self._log_path, "tool_error", tool=name, arguments=arguments, error="tool_not_enabled")
msg = f"Audiobook metadata tool is not enabled: {name}"
raise MetadataResolutionError(msg)
started = time.perf_counter()
write_agent_log(self._log_path, "tool_call", tool=name, arguments=arguments)
result = handler(arguments)
duration_ms = round((time.perf_counter() - started) * 1000, 3)
write_agent_log(
self._log_path,
"tool_result",
tool=name,
duration_ms=duration_ms,
result_count=len(result),
preview=result[:3],
)
return result
def get_author(self, author_id: int) -> AudiobookAuthor | None:
"""Return an author by id."""
return self._session.get(AudiobookAuthor, author_id)
def get_book(self, book_id: int) -> Audiobook | None:
"""Return a book by id."""
return self._session.get(Audiobook, book_id)
def get_series(self, series_id: int) -> AudiobookSeries | None:
"""Return a series by id."""
return self._session.get(AudiobookSeries, series_id)
def prune_unused_created_rows(self, metadata: StandardBookMetadata) -> None:
"""Remove catalog rows created during this run but not used by final metadata."""
used_book_ids = {metadata.book_id} if metadata.book_id is not None else set()
for book_id in self.created_book_ids - used_book_ids:
if book := self.get_book(book_id):
self._session.delete(book)
self._session.flush()
used_series_ids = {metadata.series_id} if metadata.series_id is not None else set()
for series_id in self.created_series_ids - used_series_ids:
series = self.get_series(series_id)
if series and not series.books:
self._session.delete(series)
self._session.flush()
for author_id in self.created_author_ids - {metadata.author_id}:
author = self.get_author(author_id)
if author and not author.books and not author.series:
self._session.delete(author)
def run_search_authors(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Search authors from tool arguments and remember returned ids."""
query = required_string(arguments, "query")
statement = select(AudiobookAuthor).order_by(AudiobookAuthor.name).limit(self._config.max_tool_results)
if terms := query_terms(query):
statement = statement.where(or_(*(AudiobookAuthor.name.ilike(f"%{term}%") for term in terms)))
authors = self._session.scalars(statement).all()
self.seen_author_ids.update(author.id for author in authors)
return [{"id": author.id, "name": author.name} for author in authors]
def run_search_series(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Search series from tool arguments and remember returned ids."""
query = required_string(arguments, "query")
author_id = optional_int(arguments.get("author_id"), "author_id")
statement = select(AudiobookSeries).order_by(AudiobookSeries.name).limit(self._config.max_tool_results)
if terms := query_terms(query):
statement = statement.where(or_(*(AudiobookSeries.name.ilike(f"%{term}%") for term in terms)))
if author_id is not None:
statement = statement.where(AudiobookSeries.author_id == author_id)
series_rows = self._session.scalars(statement).all()
self.seen_series_ids.update(series.id for series in series_rows)
self.seen_author_ids.update(series.author_id for series in series_rows)
return [
{
"id": series.id,
"name": series.name,
"author_id": series.author_id,
"author": series.author.name,
}
for series in series_rows
]
def run_search_books(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Search books from tool arguments and remember returned ids."""
query = required_string(arguments, "query")
author_id = optional_int(arguments.get("author_id"), "author_id")
series_id = optional_int(arguments.get("series_id"), "series_id")
statement = select(Audiobook).order_by(Audiobook.title).limit(self._config.max_tool_results)
if terms := query_terms(query):
statement = statement.where(or_(*(Audiobook.title.ilike(f"%{term}%") for term in terms)))
if author_id is not None:
statement = statement.where(Audiobook.author_id == author_id)
if series_id is not None:
statement = statement.where(Audiobook.series_id == series_id)
books = self._session.scalars(statement).all()
self.seen_book_ids.update(book.id for book in books)
self.seen_author_ids.update(book.author_id for book in books)
self.seen_series_ids.update(book.series_id for book in books if book.series_id is not None)
return [
{
"id": book.id,
"title": book.title,
"author_id": book.author_id,
"author": book.author.name,
"series_id": book.series_id,
"series": book.series.name if book.series else self._config.standalone_series,
"series_index": book.series_index,
}
for book in books
]
def run_ensure_author(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Ensure an author from tool arguments and return a tool result."""
name = normalize_catalog_slug(required_string(arguments, "name"))
validate_catalog_slug(name, "author")
author = self._session.scalar(select(AudiobookAuthor).where(AudiobookAuthor.name == name))
action = "existing"
if author is None:
author = AudiobookAuthor(name=name)
self._session.add(author)
self._session.flush()
self.created_author_ids.add(author.id)
action = "created"
self.seen_author_ids.add(author.id)
return [{"id": author.id, "name": author.name, "action": action}]
def run_ensure_series(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Ensure a series from tool arguments and return a tool result."""
name = normalize_catalog_slug(required_string(arguments, "name"))
author_id = required_int(arguments, "author_id")
validate_catalog_slug(name, "series")
author = self.required_author(author_id)
series = self._session.scalar(
select(AudiobookSeries).where(
AudiobookSeries.name == name,
AudiobookSeries.author_id == author.id,
),
)
action = "existing"
if series is None:
series = AudiobookSeries(name=name, author=author)
self._session.add(series)
self._session.flush()
self.created_series_ids.add(series.id)
action = "created"
self.seen_author_ids.add(author.id)
self.seen_series_ids.add(series.id)
return [self.series_result(series, action)]
def run_ensure_book(self, arguments: dict[str, object]) -> list[dict[str, object]]:
"""Ensure a book from tool arguments and return a tool result."""
title = required_string(arguments, "title")
author_id = required_int(arguments, "author_id")
series_id = optional_int(arguments.get("series_id"), "series_id")
series_index = required_int(arguments, "series_index")
ensured = self.ensure_book(title, author_id, series_id, series_index)
return [self.book_result(ensured.book, ensured.action)]
def ensure_book(
self,
title: str,
author_id: int,
series_id: int | None,
series_index: int,
) -> EnsuredBook:
"""Return an existing book row, or create it after validating ownership."""
title = normalize_title_slug(title)
validate_title_slug(title)
author = self.required_author(author_id)
series = None
if series_id is None:
if series_index != 0:
msg = "standalone books must use series_index 0"
raise MetadataResolutionError(msg)
else:
series = self.required_series(series_id)
if series.author_id != author.id:
msg = f"series_id {series_id} does not belong to author_id {author_id}"
raise MetadataResolutionError(msg)
if series_index <= 0:
msg = "series books must use a positive series_index"
raise MetadataResolutionError(msg)
statement = select(Audiobook).where(
Audiobook.title == title,
Audiobook.author_id == author.id,
)
if series is None:
statement = statement.where(Audiobook.series_id.is_(None))
else:
statement = statement.where(Audiobook.series_id == series.id)
book = self._session.scalar(statement)
if book is None:
book = Audiobook(title=title, author=author, series=series, series_index=series_index)
self._session.add(book)
self._session.flush()
self.created_book_ids.add(book.id)
action = "created"
else:
action = "existing"
self.seen_book_ids.add(book.id)
self.seen_author_ids.add(author.id)
if book.series_id is not None:
self.seen_series_ids.add(book.series_id)
return EnsuredBook(book=book, action=action)
def required_author(self, author_id: int) -> AudiobookAuthor:
"""Return an author or fail metadata resolution."""
author = self.get_author(author_id)
if author is None:
msg = f"author_id {author_id} does not exist"
raise MetadataResolutionError(msg)
return author
def required_series(self, series_id: int) -> AudiobookSeries:
"""Return a series or fail metadata resolution."""
series = self.get_series(series_id)
if series is None:
msg = f"series_id {series_id} does not exist"
raise MetadataResolutionError(msg)
return series
def series_result(self, series: AudiobookSeries, action: str) -> dict[str, object]:
"""Build a normalized series tool result."""
return {
"id": series.id,
"name": series.name,
"author_id": series.author_id,
"author": series.author.name,
"action": action,
}
def book_result(self, book: Audiobook, action: str) -> dict[str, object]:
"""Build a normalized book tool result."""
return {
"id": book.id,
"title": book.title,
"author_id": book.author_id,
"author": book.author.name,
"series_id": book.series_id,
"series": book.series.name if book.series else self._config.standalone_series,
"series_index": book.series_index,
"action": action,
}
class AudiobookMetadataAgent:
"""Ollama-backed metadata resolver with a fixed local tool registry."""
@@ -571,45 +195,15 @@ class AudiobookMetadataAgent:
should_continue=False,
)
if tool_calls:
return self.handle_tool_calls(messages, message, tool_calls, invalid_final_count)
return self.handle_final_message(messages, message, invalid_final_count)
def handle_tool_calls(
self,
messages: list[dict[str, object]],
message: dict[str, object],
tool_calls: list[tuple[str, dict[str, object]]],
invalid_final_count: int,
) -> AgentStepResult:
"""Run tool calls from one model response and append tool results."""
messages.append(message)
for tool_name, arguments in tool_calls:
try:
tool_result = self._registry.run(tool_name, arguments)
except MetadataResolutionError as error:
if is_fatal_tool_error(error):
return AgentStepResult(
metadata=review_metadata(str(error), self._config),
invalid_final_count=invalid_final_count,
should_continue=False,
)
write_agent_log(self._log_path, "tool_error", tool=tool_name, arguments=arguments, error=str(error))
messages.append(
{
"role": "tool",
"tool_name": tool_name,
"content": json.dumps({"error": str(error)}, sort_keys=True),
},
fatal_error = run_tool_calls(messages, message, tool_calls, self._registry, self._log_path, write_agent_log)
if fatal_error is not None:
return AgentStepResult(
metadata=review_metadata(fatal_error, self._config),
invalid_final_count=invalid_final_count,
should_continue=False,
)
continue
messages.append(
{
"role": "tool",
"tool_name": tool_name,
"content": json.dumps(tool_result, sort_keys=True),
},
)
return AgentStepResult(metadata=None, invalid_final_count=invalid_final_count, should_continue=True)
return AgentStepResult(metadata=None, invalid_final_count=invalid_final_count, should_continue=True)
return self.handle_final_message(messages, message, invalid_final_count)
def handle_final_message(
self,
@@ -905,43 +499,6 @@ def user_prompt(aax_file_name: str, metadata: dict[str, str]) -> str:
)
def parse_tool_calls(message: dict[str, object]) -> list[tuple[str, dict[str, object]]]:
"""Parse Ollama tool calls from a response message."""
raw_tool_calls = message.get("tool_calls") or []
if not isinstance(raw_tool_calls, list):
msg = "tool_calls must be a list"
raise MetadataResolutionError(msg)
tool_calls = []
for raw_call in raw_tool_calls:
if not isinstance(raw_call, dict):
msg = "tool call must be an object"
raise MetadataResolutionError(msg)
function = raw_call.get("function")
if not isinstance(function, dict):
msg = "tool call is missing function"
raise MetadataResolutionError(msg)
name = function.get("name")
if not isinstance(name, str) or not name:
msg = "tool call is missing function name"
raise MetadataResolutionError(msg)
arguments = parse_tool_arguments(function.get("arguments", {}))
tool_calls.append((name, arguments))
return tool_calls
def parse_tool_arguments(raw_arguments: object) -> dict[str, object]:
"""Parse tool call arguments returned by Ollama."""
if isinstance(raw_arguments, dict):
return {str(key): value for key, value in raw_arguments.items()}
if isinstance(raw_arguments, str):
parsed = json.loads(raw_arguments) if raw_arguments else {}
if isinstance(parsed, dict):
return {str(key): value for key, value in parsed.items()}
msg = "tool arguments must be an object"
raise MetadataResolutionError(msg)
def parse_final_json_content(content: str) -> object:
"""Parse final model content, accepting bare or fenced JSON."""
stripped = content.strip()
@@ -967,41 +524,6 @@ def parse_final_metadata_fields(raw_metadata: object) -> FinalMetadataFields:
)
def validate_title_slug(title: str) -> None:
"""Validate a canonical book title slug."""
if not TITLE_SLUG_PATTERN.fullmatch(title):
msg = f"title slug is invalid: {title}"
raise MetadataResolutionError(msg)
def validate_catalog_slug(value: str, label: str) -> None:
"""Validate a canonical catalog slug."""
if not CATALOG_SLUG_PATTERN.fullmatch(value):
msg = f"{label} slug is invalid: {value}"
raise MetadataResolutionError(msg)
def normalize_catalog_slug(value: str) -> str:
"""Normalize noisy catalog names into lower snake-case slugs."""
return re.sub(r"[^a-z0-9]+", "_", value.strip().casefold()).strip("_")
def normalize_title_slug(value: str) -> str:
"""Normalize noisy book titles into lower kebab-case slugs."""
return re.sub(r"[^a-z0-9]+", "-", value.strip().casefold()).strip("-")
def is_fatal_tool_error(error: MetadataResolutionError) -> bool:
"""Return whether a tool error should stop the agent immediately."""
message = str(error)
return message.startswith(
(
"Unknown audiobook metadata tool",
"Audiobook metadata tool is not enabled",
),
)
def review_metadata(reason: str, config: AgentConfig) -> StandardBookMetadata:
"""Return a metadata result that must be reviewed manually."""
return StandardBookMetadata(
@@ -1018,42 +540,6 @@ def review_metadata(reason: str, config: AgentConfig) -> StandardBookMetadata:
)
def query_terms(query: str) -> tuple[str, ...]:
"""Return text variants useful for matching noisy audiobook metadata."""
normalized = query.strip().casefold()
underscore_slug = normalize_catalog_slug(normalized)
hyphen_slug = normalize_title_slug(normalized)
return tuple(dict.fromkeys(term for term in (normalized, underscore_slug, hyphen_slug) if term))
def required_string(data: dict[str, object], key: str) -> str:
"""Read a required string field."""
value = data.get(key)
if not isinstance(value, str) or not value.strip():
msg = f"{key} must be a non-empty string"
raise MetadataResolutionError(msg)
return value.strip()
def required_int(data: dict[str, object], key: str) -> int:
"""Read a required integer field."""
value = data.get(key)
if isinstance(value, bool) or not isinstance(value, int):
msg = f"{key} must be an integer"
raise MetadataResolutionError(msg)
return value
def optional_int(value: object, key: str) -> int | None:
"""Read an optional integer field."""
if value is None:
return None
if isinstance(value, bool) or not isinstance(value, int):
msg = f"{key} must be an integer or null"
raise MetadataResolutionError(msg)
return value
def required_float(data: dict[str, object], key: str) -> float:
"""Read a required float field."""
value = data.get(key)