Question 600
The evaluation suite gives high groundedness scores but poor usefulness on legitimate questions. What additional metric should be included?
Explanation
Correct Answer
Quoting sources without addressing the user's need can be well grounded but still operationally inadequate.
Incorrect Answers
The alternative, Measure response groundedness against the provided retrieved context, belongs to a different technical condition: A RAG answer makes several claims not present in retrieved documents. There the source-specific mechanism or proof is: Groundedness checks whether claims have evidence in source content, not merely whether prose sounds natural. The condition asked about here is: The evaluation suite gives high groundedness scores but poor usefulness on legitimate questions. A suitable correction is Assess response relevance or task satisfaction alongside groundedness; the targeted justification or result is: Quoting sources without addressing the user's need can be well grounded but still operationally inadequate.
Assess response completeness against the required ground truth or checklist. That operation would be relevant if: A support assistant answers one of three required subquestions and omits the others. Its corresponding reason or verification is: Completeness identifies missing necessary elements even when the included facts are correct. However the problem in this question is: The evaluation suite gives high groundedness scores but poor usefulness on legitimate questions. The requirement instead calls for Assess response relevance or task satisfaction alongside groundedness; the question-specific check is: Quoting sources without addressing the user's need can be well grounded but still operationally inadequate.
Measure retrieval relevance separately from final response quality. A separate source case supports that approach when: A retrieval system returns irrelevant passages, but the language model correctly summarizes whatever it receives. The supporting rationale or observable result is: Retrieval evaluation isolates search failures that output-language metrics alone cannot diagnose. This question describes a different issue: The evaluation suite gives high groundedness scores but poor usefulness on legitimate questions. Address it through Assess response relevance or task satisfaction alongside groundedness, with the required justification or evidence: Quoting sources without addressing the user's need can be well grounded but still operationally inadequate.