RAG

RAG Evaluation

Measure retrieval and generation quality in RAG systems: context metrics, faithfulness, and answer relevance.

1. What does 'context recall' measure in a RAG pipeline?

2. Why is faithfulness evaluated separately from answer relevance?

3. What is a common way to compute faithfulness with an LLM judge?

4. What failure does a high context-recall but low context-precision score indicate?

5. Why should a RAG evaluation set include questions the knowledge base cannot answer?

6. What is the risk of evaluating RAG only with end-to-end answer correctness?

7. What does answer relevance specifically penalize?