Knowledge Bases
Upload files, crawl websites, manage sources, and reprocess documents.
Knowledge powers RAG answers. Documents are chunked, embedded, and stored in Postgres + pgvector. Ingestion runs in a background worker inside the web app process.
Supported inputs
| Type | Notes |
|---|---|
| Files | PDF, text, Markdown, DOCX, CSV, HTML, JSON |
| Text / FAQ | Paste or FAQ pairs from the UI |
| Website crawl | URL / sitemap-style source with sync jobs |
Upload flow
- Open the assistant → Knowledge
- Upload a file or add text/FAQ
- Status moves from processing → ready (chunk count shown when done)
- Ask in the Playground once status is ready
Failed jobs show an error on the document (for example embedding dimension mismatches — see Ollama).
Website sources
Add a website source, configure crawl limits/exclusions, then sync. Sync upserts URL documents and enqueues ingest jobs. Re-sync after content changes; scheduled cron refresh is deferred.
Reprocess
Use reprocess after:
- Changing chunking mode (
standard↔parent_child) - Changing embedding model (same dimensions only)
- Fixing a failed ingest
Reprocess rebuilds chunks and embeddings for that document.
Tips
- Prefer a few clear documents over huge noisy dumps
- FAQs work well for short policy answers
- After
pnpm seed:demo, wait for the refund policy document to reach ready before testing
Related
- Assistants
- RAG and debugger
- Providers — embedding dimensions