hard20 pts · first answer countsragevaluationprompt-debugging

Where is the RAG bug?

A RAG bot answers 30% of eval questions wrongly. Before changing the answer prompt, what is the most useful first step?

A.Switch to a bigger generation model straight awayB.Measure retrieval separately: for each failure, check whether the needed passage was in the retrieved context at allC.Rewrite the system prompt in a friendlier toneD.Double the chunk size

← Follow-up questions fail in RAG Strict JSON hurt accuracy →