How should RAG evaluation be designed for production?

Maintain labeled queries, measure recall and precision at k, context relevance, faithfulness, answer correctness, latency, and cost, and review failures by category.