How Can I Tell Whether a RAG Assistant Is Actually Using Its Sources?

We’re testing a retrieval-augmented assistant that answers questions from internal documents. It often gives plausible responses, but that alone doesn’t tell us whether it retrieved the right material or simply filled in gaps from its general training.

What would make a useful evaluation set for this? I’m trying to measure retrieval quality separately from answer quality, including cases where the documents don’t contain enough information and the assistant should say so.