Knowledge base (pgvector)

Read docs/file-processing.md for the pipeline and docs/howto/add-sync-connector.md / docs/howto/configure-sync-sources.md for sources. Code: app/services/rag/ — ingestion, vectorstore, embeddings, connectors.

The model reaches it through the knowledge capability, whose tool is search_documents. (search_knowledge_base is an internal function in capabilities/knowledge/_search.py, not the tool the model sees.) A collection is searched, never browsed: the model chooses what to look for and can never widen where it looks — collections are resolved from the spec before the run.

CLI

uv run agenticos cmd rag-ingest ./docs/ --collection docs --recursive
uv run agenticos cmd rag-search "your question" --collection docs
uv run agenticos cmd rag-collections
uv run agenticos cmd rag-stats
uv run agenticos cmd rag-drop <collection> --yes
uv run agenticos cmd rag-sources
uv run agenticos cmd rag-source-add
uv run agenticos cmd rag-source-sync --all
uv run agenticos cmd rag-source-remove <id>

Ingestion is parse → chunk → embed → upsert. It runs as an ingest-document job in a worker's child process, under a memory watchdog, never inline in a request - see the background-task skill. An upload queues the job in its own transaction; a document whose parse kills the child is retried once, then dead with why.

The four traps

1. The database must be pgvector/pgvector:pg16. The store issues CREATE EXTENSION IF NOT EXISTS vector on first write; stock Postgres answers extension "vector" is not available — a 500 before any row is committed. If ingestion 500s on a fresh environment, check the image first. Every compose file and both CI jobs pin it; they used to pin postgres:16-alpine, which is why no ingestion path had ever been exercised locally or in CI.

2. A format list can lie, and the upload is accepted anyway. GET /rag/supported-formats and the upload validator answer from PARSER_FORMATS; DocumentProcessor.process_file is what actually routes. When the two disagree the upload succeeds — file stored, document row committed, task dispatched — and dies in a worker, so the document sits in the listing with no explanation. tests/test_supported_formats.py pins each parser's set against what the pipeline can route, in both directions. Widening a format set is what that test exists for.

3. Narrowing an IngestionConfig rule breaks existing rows. It lives in a JSONB column, so a tighter rule does not only reject new input — it makes stored rows unreadable, and a Pydantic model refusing one field of one row takes the whole listing endpoint down with a 500. Adding a field is safe (missing keys take defaults); narrowing needs a data migration in the same change. the OCR language codes are the worked example. See the alembic-migration skill.

4. A document that parses to nothing. Silently indexing an empty result is indistinguishable afterwards from a document that ingested fine and never matches. Markdown reconstruction returns an empty fenced block rather than whitespace for an unreadable scan, so .strip() is not the check.

Credentials

Embedding keys, connector credentials and a LlamaParse key are organization secrets in the vault. CHANNEL_ENCRYPTION_KEY is gone — see the vault-secrets skill. app/services/embedding_resolution.py decides which key a collection embeds with, and it is in the gated platform layer.

Adding a connector

Implement it in app/services/rag/connectors/ following the Google Drive / S3 connectors, register it so the sync service discovers it, and expose its config fields. docs/howto/add-sync-connector.md is the walkthrough; docs/patterns.md has the registration shape.

Debugging bad results

1. Is the collection populated? rag-stats.
2. Is it bound to this agent's spec (collection_ids)? An unbound collection is invisible, and knowledge bound with no collections contributes nothing at all.
3. Does the org have knowledge:read?
4. Same embedding model throughout? Do not mix embeddings within a collection — re-ingest if it changed.
5. Did the document actually parse? See trap 4.

Coverage

app/services/rag/* is template-inherited and outside the coverage gate — which is exactly why the invariants above are pinned by explicit named tests rather than left to a percentage. Three format lists disagreed there for months without anything failing.