The agent doesn't use my knowledge

Knowledge turns on by itself as soon as an agent has indexed content: there's no switch. For each question SentientOne runs a vector search and a keyword search, fuses and reranks the results, and adds the best passages to the prompt only if the top one is a good enough match. When an answer ignores your content, one of those steps came up short. The trace tells you which.

Check these in order

1. 1

   The source is Ready

   On the agent's Knowledge tab, each document or page should show Ready. A failed PDF usually has no text layer: scanned, image-only PDFs fail with "No indexable text found". Export a PDF with real text, or paste the content as snippets.

2. 2

   It's on the right agent

   Knowledge belongs to one agent. Check you added it to the agent you're asking, not another one.

3. 3

   Look at what was retrieved

   In the Playground, each reply shows the sources it used. In the API stream, look for the sources event. No sources means no passage scored high enough to be used. Open the trace for the rag.* scores. See Debug a bad answer.

4. 4

   The match is strong enough

   Passages are only added when the best one scores at least 0.30 after reranking (0.50 when the agent also has MCP tools, so it doesn't answer from a weak match when it could look something up). Questions phrased very differently from your content can fall below that. Add a Q&A pair with the question as your users ask it.

5. 5

   The answer isn't cached

   The first question of a new conversation can be answered from a 15-minute cache. Adding knowledge clears it, but if you changed something else, ask in an existing conversation or wait 15 minutes.

6. 6

   The prompt doesn't fight it

   A system prompt that tells the agent to answer from general knowledge, or to keep answers very short, can override the grounding rule that's added automatically. See Write a system prompt.

More questions