RAG Evaluation Checklist (Triad)

• [ ] Context relevance (retrieved chunks topically relevant)
• [ ] Groundedness (answer supported by retrieved text)
• [ ] Answer relevance (addresses the query)
• [ ] Latency: p50, p95 (ms); Token counts (prompt/response)
• [ ] Citations contain stable doc IDs + spans
• [ ] Regression set: 10–20 fixed questions with expected sources