Use offline tests, online monitoring, user feedback, and safety checks.

Evaluation should measure answer quality, grounding, retrieval quality, latency, cost, safety, and task completion across realistic examples.