RAG Knowledge Base ({{ cookiecutter.vector_store }})

The RAG stack lives in backend/app/services/rag/ (ingestion, vectorstore, embeddings, connectors). Retrieval is exposed to the agent as the search_knowledge_base tool, and to operators via the CLI and the dashboard.

CLI (run from backend/)

uv run {{ cookiecutter.project_slug }} cmd rag-ingest ./docs/ --collection docs --recursive   # ingest files/folder
uv run {{ cookiecutter.project_slug }} cmd rag-search "your question" --collection docs        # semantic search
uv run {{ cookiecutter.project_slug }} cmd rag-collections                                     # list collections
uv run {{ cookiecutter.project_slug }} cmd rag-stats                                           # chunk/vector counts
uv run {{ cookiecutter.project_slug }} cmd rag-drop <collection> --yes                         # delete a collection

Ingestion = parse → chunk → embed → upsert into {{ cookiecutter.vector_store }}. Re-ingesting the same source updates it (use --no-replace / --sync-mode to control dedupe).

Sync sources (connectors)

Connectors keep a collection in sync with an external source (Google Drive, S3/MinIO) on a schedule, and can be managed per-organization from the dashboard (/orgs/[id]/integrations) or via CLI:

uv run {{ cookiecutter.project_slug }} cmd rag-sources                  # list
uv run {{ cookiecutter.project_slug }} cmd rag-source-add               # add (interactive)
uv run {{ cookiecutter.project_slug }} cmd rag-source-sync --all        # trigger a sync

Connector credentials are encrypted at rest with CHANNEL_ENCRYPTION_KEY (Fernet).

Adding a new connector type

Implement a connector in backend/app/services/rag/connectors/ following the existing Google Drive / S3 connectors, register it in the connector registry, and expose its config fields. See docs/howto/add-sync-connector.md and docs/howto/configure-sync-sources.md.

Tuning retrieval

• Chunk size/overlap and parser (PyMuPDF / LlamaParse) are configured via env — see docs/configuration.md and docs/rag.md.
• Reranking (Cohere or local CrossEncoder) improves result ordering when enabled.
• If search returns poor results: confirm the collection is populated (rag-stats), check the active collection in the chat's KB selector, and verify the embedding provider/key.

Rules

• Embedding provider and dimensions are fixed per project ({{ cookiecutter.embedding_provider }}) — don't mix embeddings across a collection; re-ingest if you change them.
• Heavy ingestion runs as a background job, not inline in a request.
• See docs/rag.md for the full pipeline reference.