Grounding and "I don't know"
Why the agent refuses to guess, how to trace an answer back to its sources, and what to do when it falls back too often.
Every answer is grounded: it is written from passages retrieved out of your knowledge base, not from the model's general training. When retrieval finds nothing relevant, the agent says so instead of producing something plausible.
The fallback is a feature
A support agent that invents a refund window is worse than one that says it does not know. The fallback is deterministic — no model call is made — so it cannot drift into a confident guess, and it always reads the same:
I couldn't find that in the knowledge base. Please contact our support team
for help.Set your own wording in the chatbot's fallback message, and point it somewhere useful — a contact form, a phone number, or a nudge to ask for a human. See Chatbot settings.
Tracing an answer
Open the Playground and turn on Debug. For each answer you get the retrieved chunks and their scores, the latency, the token usage, and the assembled prompt. That tells you which of two very different problems you have:
| What Debug shows | Diagnosis | Fix |
|---|---|---|
| No chunks retrieved | Nothing in the knowledge base matches | Add the content, or add an FAQ pair |
| The right chunk retrieved, poor answer | A wording or instruction problem | Adjust the system prompt or tone |
| The wrong chunk retrieved with a high score | Competing or contradictory content | Delete the stale source; keep one answer per fact |
| The right chunk retrieved but scored below the threshold | Threshold too high | Lower the similarity threshold slightly |
Sources are attached to each message as metadata rather than printed as [1] citations, which is why you read them in Debug rather than in the answer text.
When it falls back too often
In rough order of how often it is the cause:
- 1The content genuinely is not there. Check the
knowledge gaps list — it is exactly the set of questions that got no sources.
- 2The content is there but unreadable. A scanned PDF or a JavaScript-only
page extracts to nothing. Inspect the chunks — see Reindexing and freshness.
- 3The source is still processing, or disabled. Check its status.
- 4The similarity threshold is too high. Move it in small steps and re-test.
Resist fixing a fallback by loosening the threshold far: you trade "I don't know" for confident answers built from loosely related passages, which is the failure mode grounding exists to prevent.
Where to go next
Last updated