RAG evaluation requires separately considering recall (whether retrieval finds the correct content) and faithfulness (whether the generated answer is truly based on the retrieved material).