RAG evaluation
Also known as: RAGAS · RAG metrics
RAG evaluation separates the two halves of RAG so you can measure each:
- Retrieval quality: did we find the right chunks? (precision/recall of retrieval)
- Answer quality: given the chunks, is the answer faithful, complete, and grounded? (faithfulness, answer relevance)
Frameworks like RAGAS formalize these.
Diagnose, don't guess: bad RAG is usually *either* bad retrieval *or* bad synthesis. Evaluate both halves separately so you fix the right one.