Missing knowledge answers
The agent ignores your documents, answers from memory, or finds the wrong passage. How to diagnose each.
What you will achieve
You will make the agent answer from your documents, and show you when it cannot.
Check in this order
- Is the document indexed? Open the knowledge base: the document should say Indexed with a passage count above zero. Failed or 0 passages means no text was read (a scanned PDF, for example).
- Is it attached? On the agent's Knowledge tab, switch the knowledge base on and Save.
- Does a search find it? Use the base's test retrieval with the caller's wording. If nothing is found, the agent cannot find it either.
- Is the cutoff right? Match cutoff too high hides real matches; too low brings in unrelated passages.
- Does the prompt allow it? Tell the agent in Instructions to answer from the knowledge base and say so when it does not know.
Symptoms
| Symptom | Likely cause | Fix |
|---|---|---|
| The agent answers from general knowledge, not your files | Not attached, or the passage is below the cutoff | Attach; lower the cutoff |
| Exact product codes are not found | Meaning-based search misses literal strings | Turn on Hybrid search |
| An outdated answer | A web page or file changed | Refresh the document or set Automatic refresh |
| Slower replies after attaching | Many bases or passages | Attach fewer; lower Passages per answer |
| It invents an answer | Nothing relevant was found and the prompt does not forbid guessing | Add "If the answer is not in the knowledge, say you do not know" |
| A website did not load | It needs JavaScript, blocks crawlers or is private | Paste its text instead |
| It gives the wrong office location | The address passage was not found for that question | Set Office address under Business details on the agent; it is always sent, whatever the search finds |
| It answers in the language of a document | The agent is on Deepgram Nova 3 Multilingual, which follows the caller | Choose a single-language hearing option; the agent then always replies in its Language |
| Two documents disagree | Both were found | The agent says it is not certain; remove or update the outdated document |
Note. Search needs an embedding model chosen by the platform; the choice cannot be changed per workspace.
Testing
Ask the exact question a caller would, in Talk, and compare the answer with the passage the test retrieval shows.
Expected result
The answer matches your document and the agent says it does not know when something is missing.
Related guides
Something here does not match what you see in the dashboard? Let us know.