Inaccurate KB answers
When your agent answers from a knowledge base incorrectly, review the indexed content, fix it, and re-index.
Symptoms
- The agent gives wrong, outdated, or incomplete answers from a knowledge base.
- It misses details that are clearly in your original document.
Why this happens
Your agent doesn't re-read the original file on every call — it retrieves from the indexed content, the text that was extracted from your upload when the knowledge base was processed. If extraction missed a section, garbled a table, or the source has since changed, the answers will be off. The fix is to review that indexed text and re-index it.
Fix it
Open the knowledge base
Go to Knowledge Base in the sidebar and click the knowledge base's name to open it.
Review the indexed content
The viewer shows the indexed content — exactly what the agent retrieves on a call. Compare it against your source; for a PDF, use View original to open the uploaded file side by side. Look for missing sections, garbled tables, or extra clutter (navigation, repeated headers).
Make your changes
Edit the text directly in the viewer — correct mistakes, add missing details, and remove noise. An Unsaved changes note appears as you edit.
Save & Re-index
Click Save & Re-index. Your edits are re-chunked and re-embedded — you'll see Re-indexing started, and the status returns to Ready in a few seconds.
Re-test the agent
Ask the agent the same question again to confirm the answer is now correct.
For a website knowledge base that's gone stale, use Recrawl (in the ⋯ menu) to pull fresh content from the URL. Note: recrawling replaces your manual edits with the newly scraped text.
Only a knowledge base in the Ready state is used on calls. If it's stuck on Processing, Indexing, or shows Failed, wait for it to finish — or re-upload — before testing.
For how knowledge bases are extracted, chunked, and retrieved, see Knowledge Base.