Faithfulness, also called grounding or groundedness, is the property that every claim in a model's output is supported by the source material it was given: retrieved documents, a transcript, a database result. A faithful answer can still be wrong if the source is wrong; an unfaithful answer adds, alters or contradicts what the sources say.
Faithfulness versus factuality
Factuality asks whether a claim is true about the world. Faithfulness asks whether it follows from the provided context. The distinction matters in retrieval-augmented systems: if a policy document says the refund window is 30 days and the model says 60 because it remembers a different company's policy, the answer is unfaithful even though it might be "true" somewhere. Unfaithful output is one form of hallucination.
How it is measured
The common recipe is claim-level: split the answer into atomic claims, then check each claim against the sources with an entailment model or a calibrated LLM judge. The score is
Reports usually add the share of answers with at least one unsupported claim, since a single invented figure can make a whole answer unusable.
Worked example
Illustrative numbers: an answer about a loan agreement contains 8 claims. Six are stated in the retrieved contract, one is a reasonable paraphrase that the checker accepts, and one (an interest rate) appears nowhere in the sources. Faithfulness is , and the answer fails an "all claims supported" gate. Across 200 answers, if 14 have any unsupported claim, the answer-level failure rate is 7%.
Pitfalls
- Claim splitting is itself a model step and can drop or merge claims.
- Checkers struggle with numbers, negation and claims that need combining two passages.
- Citations are not proof: a cited passage may not support the sentence it is attached to. Check citation precision separately.
Why it matters
In finance, healthcare and legal workflows, the point of retrieval is that answers trace back to an authoritative source. Faithfulness is the metric that tells you whether they do. It belongs in every eval suite for a retrieval system and in online evaluation, because new documents change how often the model drifts from them.