Internal RAG evaluation (Ragas-style, LLM-as-judge): faithfulness, answer relevancy, context relevancy. Not user-facing.