Skip to main content
Wrong answers usually come from the agent receiving the wrong passages, not from the model. Retrieval test runs the exact question through the exact KB with the same configuration the agent uses, then shows you each returned chunk with its score and an automatic diagnosis.

Screen layout

Configuration parameters

Reading the results

Each returned chunk shows its rank, passage title, score (e.g. 0.51), document name, passage content and chunk ID. The first line of the results says how many results came from how many candidates. The Diagnosis block: Click Add to golden question set to save this question to the test set in Quality evaluation.

Debugging wrong results

1

No chunks returned

Lower the score threshold, check that the documents are Processed, and check whether the metadata filter is too strict.
2

Correct chunk ranked low

Switch to Hybrid + rerank, or increase the BM25 weight if the question contains document numbers or codes.
3

Chunk cut off midway

Turn on Expand to parent chunk in the KB’s Settings tab, or edit the chunk manually in the chunk viewer.
4

Outdated documents returned

Turn on Only return documents in effect and fill in the effective date field for the documents.