How do you evaluate a RAG system?

Evaluate RAG in three layers: (1) Retrieval quality: Recall@k, MRR, nDCG@k against a held-out query set with ground-truth chunk IDs; (2) Generation quality: faithfulness (whether the answer is grounded in retrieved context, RAGAS "faithfulness"), answer-relevance (whether the answer addresses the query, RAGAS "answer_relevancy"), context-relevance (RAGAS "context_precision" / "context_recall"); (3) End-to-end: human eval via pairwise comparison, A/B testing, or LLM-as-judge. Frameworks: RAGAS, DeepEval, TruLens, Phoenix (Arize). Senior nuance: RAGAS-style metrics are useful but biased (LLM judge biases toward verbose, "confident" answers); pair them with explicit ground-truth checks and human spot-audits. Also: retrieval changes suddenly when docs change, so eval must be re-run continuously.