An enterprise search assistant gives wrong answers for about 5% of questions. The team cannot tell whether the search layer returned the wrong passages, the prompt template dropped them, or the model misread them. Which capability would isolate the failing stage?
Show answer and explanation
Answer: D
Per-request tracing captures each stage of the pipeline for a given run: the retrieved passages, the assembled prompt, and the model response, along with timing. Reviewing those traces shows which stage introduced the error, so the team fixes the right component. Cost reports, stylesheets, and disabled logging do not reveal where a RAG answer went wrong.
Why the other options are wrong
- A. A cost report shows spending by resource group. It does not reveal which retrieved passage or prompt led to a wrong answer.
- B. Interface styling affects appearance only. It has no bearing on whether retrieval or generation produced the error.
- C. Disabling logs removes the evidence needed for diagnosis. Failures become harder to trace, not easier.