feat(rag): add active generation schema and store (#438 PR1)

[sqhyz55]

Introduce nullable generation_id columns, active_generations table, and ActiveGenerationStore for crash-safe re-chunking. Keep legacy chunk/embedding merge keys until ingestion publishes generations in a follow-up PR.

[gemini-code-assist]

Code Review

This pull request introduces the ActiveGenerationStore and its LanceDB implementation to track published searchable generations, updating the schemas for chunks and embeddings tables to include generation_id and config_hash while providing migration utilities for existing data. Feedback recommends optimizing performance by caching table existence checks and improving the atomicity of the publication process to reduce database round-trips.

[gemini-code-assist]

The _ensure_table method is called on every database operation (publish, get, list). Since it performs a directory listing via list_table_names to check for table existence, this can become a performance bottleneck during high-volume ingestion. Caching the result of this check after the first successful execution would significantly reduce overhead.

Suggested change

	def _ensure_table(self) -> None:
	from ..LanceDB.schema_manager import ensure_active_generations_table
	ensure_active_generations_table(self._get_sync_connection())
	def _ensure_table(self) -> None:
	if getattr(self, "_table_ensured", False):
	return
	from ..LanceDB.schema_manager import ensure_active_generations_table
	ensure_active_generations_table(self._get_sync_connection())
	self._table_ensured = True

[gemini-code-assist]

The current implementation performs an extra round-trip to the database to fetch the existing record just to preserve the created_at timestamp. This can be achieved more efficiently and atomically using merge_insert with a specific when_matched_update clause that excludes created_at. This avoids the overhead of a separate search operation.

        try:
            now = datetime.now(timezone.utc).replace(tzinfo=None)
            record = {
                "collection": collection,
                "doc_id": doc_id,
                "parse_hash": parse_hash,
                "user_id": user_id,
                "model_tag": normalized_tag,
                "generation_id": generation_id,
                "config_hash": config_hash,
                "created_at": now,
                "updated_at": now,
                "published_at": now,
                "operator": operator or "unknown",
            }

            (
                table.merge_insert(
                    on=[
                        "collection",
                        "doc_id",
                        "parse_hash",
                        "user_id",
                        "model_tag",
                    ]
                )
                .when_matched_update(
                    updates={
                        "generation_id": "source.generation_id",
                        "config_hash": "source.config_hash",
                        "updated_at": "source.updated_at",
                        "published_at": "source.published_at",
                        "operator": "source.operator",
                    }
                )
                .when_not_matched_insert_all()
                .execute([record])
            )
        finally:
            _safe_close_table(table)

[sqhyz55]

当前依赖的 LanceDB 0.30.2 的 merge_insert 只有：

when_matched_update_all()
when_not_matched_insert_all()
when_not_matched_by_source_delete()
没有 when_matched_update(updates={...})