A RAG answer is faithful when every claim it makes is supported by the retrieved context. Unfaithful answers are hallucinations dressed up with citations.
Why Answers Drift From Sources
- The model fills gaps with its general knowledge.
- Retrieved chunks are partially relevant, inviting inference.
- Instructions don't make grounding requirements clear.
- The answer combines sources in ways none of them states.
Prompting for Faithfulness
- Answer only from the provided documents.
- If the documents don't contain the answer, say so.
- Quote or cite supporting passages.
- Don't combine information into conclusions the sources don't state.
Checking Faithfulness
- Claim decomposition: split the answer into individual claims and check each against the context.
- Model-based graders that judge support per claim.
- Natural language inference models that classify whether a passage entails a statement.
Measuring It
Report the share of answers with no unsupported claims, and the share of claims that are supported. Validate automated graders with human review.
Faithful Isn't Always Correct
If the retrieved source is outdated or wrong, a faithful answer repeats the error. Faithfulness must be paired with good, current sources.
When Some Inference Is Wanted
For analytical questions, allow reasoning but require it to be labelled as inference and grounded in cited facts.