Failure point mapping connects evaluation dimensions to concrete RAG failure modes, turning metric scores into targeted debugging signals. section 14.0
Map each evaluation dimension to the system area and failure category it diagnoses:
- Context relevance tests retrieval, detecting missing content, missed top-ranked results, or information absent from context (FP1, FP2, FP4). section 14.0
- Faithfulness tests generation, detecting factual inconsistency or hallucination (FP3). section 14.0
- Answer relevance tests generation, detecting information not extracted, incorrect specificity, or incomplete responses (FP5, FP6, FP7). section 14.0
From score to intervention
A low context-relevance score directs attention to retrieval; a low faithfulness score points to generation; a low answer-relevance score identifies extraction, specificity, or completeness problems. These signals support metrics-driven iteration instead of subjective demo testing. section 14.0
