Knowledge and RAG
Knowledge overview
The four kinds of knowledge an agent can have, their limits and credit costs, and what happens when you add one.
Concept2 min readUpdated 23 October 2026
Knowledge is the content an agent answers from. It lives on the agent's Knowledge tab, which has four sections: Documents, Websites, Snippets and Q&A. Each agent has its own knowledge; nothing is shared between agents.
Sources at a glance
| Source | What it's for | Limits | Credits |
|---|---|---|---|
| Documents | Policies, manuals, guides, contracts | PDF up to 50 MB in the app; the ingest API also takes plain-text files | 5 per page, once |
| Websites | A help centre, docs site or product pages | Up to 200 pages per crawl, depth 1 to 10 | 5 per indexed page, per crawl |
| Snippets | Short facts, rules or definitions | 3 to 300 characters | 1 each |
| Q&A | Questions that need an exact answer | Question 3 to 200, answer 3 to 500 characters | 1 per pair |
What happens when you add a source
- Text is taken out of the PDF or the crawled page (main content only, as markdown).
- Credits are charged: per page once the page count is known, before anything is embedded.
- The text is split into chunks of about 400 words that overlap by 40 words. Snippets are kept whole; a Q&A pair becomes two chunks.
- Each chunk is embedded with OpenAI
text-embedding-3-small(1536 dimensions) and stored in Postgres with pgvector under an HNSW index, beside a full-text index of the same text. - The agent's cached answers and search results are cleared, so the next question sees the new content.
How the agent uses it
There is no switch: as soon as an agent has indexed content, every question is searched against it. The best passages go to the model with a grounding instruction, and the stream tells your app which ones were used. See How retrieval works and Sources and citations.
Good to know
- Failed work is refunded: a page that fails to index gives its credits back.
- Deleting a source removes its chunks immediately but doesn't refund credits.
- Searching knowledge is included in the reply's 1 credit.
- Scanned, image-only PDFs have no text to index and fail with "No indexable text found in this document".
Still stuck?
Email us with what you tried and any request id. A person replies.