Evaluating grounded answers in RAG systems

A retrieval system can sound confident and still be wrong. The hard part of building with RAG is not generating an answer, it is knowing whether that answer is actually supported by the sources it retrieved. This is an evaluation problem, and it is worth treating as seriously as the retrieval itself.

What to measure

We find it useful to separate retrieval quality from answer quality. A system can retrieve the right documents and then ignore them, or retrieve poorly and still guess well. Measuring them separately tells you where to spend effort.

• Context relevance: did retrieval return the passages that matter?
• Faithfulness: is every claim in the answer supported by those passages?
• Answer relevance: does the answer address what was actually asked?

Building the evaluation set

Cherry-picked demos hide the failures that matter. We build an evaluation set from real questions, including the awkward and out-of-scope ones, and keep it versioned so results are comparable across changes. When a change improves one metric and quietly harms another, a fixed set makes that visible.

Let the model help, but verify

Model-graded evaluation scales well and is a reasonable first signal, but it is not ground truth. We spot-check graded results by hand, especially on faithfulness, because a model asked to grade its own kind of output can be too forgiving.

Watching it in production

Offline evaluation tells you whether a change is safe to ship. Production monitoring tells you whether quality is holding up under real traffic. We log retrieved sources alongside answers so that when something looks wrong, we can see whether retrieval or generation was at fault.

Retrieval quality, not the model, was usually the lever that mattered most.

Practical recommendations

1. 01Measure retrieval and answer quality separately.
2. 02Build a versioned evaluation set from real, messy questions.
3. 03Use model-graded scores as a signal, then verify faithfulness by hand.
4. 04Log retrieved sources with every answer so failures are debuggable.
5. 05Attach citations so users can check the answer themselves.