The Faithfulness Metric Explained - How to Catch RAG Hallucination in 2026

Faithfulness is the single most useful metric for a RAG system, because it answers the question that keeps teams up at night - is the model making things up? But “faithfulness” gets used loosely, and few explanations actually show you how the score is computed. This guide does.

We cover what faithfulness measures, the claim-extract-and-verify method behind the number, a worked example, the crucial gap between faithful and correct, and where the tools fit.

What faithfulness actually measures

Faithfulness, also called groundedness, measures whether an answer is supported by the context it was given. Every claim the model makes should trace back to the retrieved source material. Anything it asserts that is not in the context is invented - a hallucination.

The critical nuance, and the thing most people get wrong - faithfulness is about support, not truth. It does not ask whether the answer is correct in the real world. It asks whether the answer honestly used the documents it was handed. An answer can be perfectly faithful to a wrong document. Hold that thought - it drives everything below. For the base concept, see the faithfulness and hallucination glossary entries.

How the score is computed

The standard approach is claim-extract-and-verify, and it is almost always run by an LLM judge because the steps require language understanding.

1. Decompose the answer into atomic claims. The judge breaks the response into individual factual statements, so one sentence with three assertions becomes three claims.
2. Verify each claim against the context. For every claim, the judge checks whether the retrieved context supports it, contradicts it, or says nothing about it.
3. Score. Faithfulness is the number of supported claims divided by the total number of claims.

So the formula is simply:

faithfulness = supported_claims / total_claims

A score of 1.0 means every statement in the answer is grounded in the context. A score of 0.6 means 40% of the claims are unsupported - the model went beyond its evidence.

The property that makes this powerful - it is reference-free. You need the answer and the retrieved context, but you do not need a hand-written “correct” answer. That is why faithfulness can run on live production traffic, where no ground-truth label exists, and not just on a curated eval dataset.

A worked example

Your RAG app retrieves this context for a question about a product’s warranty:

Context - “The Model X carries a 2-year limited warranty covering manufacturing defects.”

The model answers:

“The Model X has a 2-year warranty covering manufacturing defects, and it includes free shipping on all repairs.”

The judge extracts three claims:

Claim	In context?	Verdict
Model X has a 2-year warranty	Yes	Supported
Warranty covers manufacturing defects	Yes	Supported
Free shipping on all repairs	No	Unsupported

Two of three claims are supported, so faithfulness = 0.67. The metric caught the hallucination - “free shipping” appears nowhere in the context. That third claim might even be true in reality, but it is not grounded in what the model was given, and that is precisely what faithfulness flags.

Faithful is not the same as correct

This is the limitation that trips teams up, so it is worth stating flatly. If your retriever pulls a wrong or outdated document and the model faithfully repeats it, the answer scores high on faithfulness while being factually wrong.

Faithfulness deliberately isolates the generation step. It tells you the model used the context honestly. It says nothing about whether the context deserved to be trusted. That separation is a feature, not a bug - it lets you localize failures - but only if you also measure retrieval quality.

The rule - faithfulness checks the generator, retrieval metrics check the retriever. Pair faithfulness with context precision and recall, and you can tell whether a bad answer came from a hallucinating model or a bad retrieval. Our guide on context precision vs recall covers that retrieval side, and the RAG triad shows how the three metrics read together.

Common mistakes

• Treating a high faithfulness score as proof the answer is right. It only proves the answer matched the context. Check retrieval quality separately.
• Running faithfulness alone. Without answer relevancy, a model could answer a different question faithfully and still fail the user. Without retrieval metrics, you cannot tell a hallucination from a bad source. See RAG evaluation metrics explained for the full set.
• Ignoring judge cost. Claim extraction plus per-claim verification means multiple judge operations per answer, so faithfulness is one of the more expensive metrics to run at scale. Budget for it - the same LLM-as-judge cost dynamic we cover in how to measure LLM hallucination.
• Vague claim boundaries. If the judge decomposes claims inconsistently, scores wobble. Keep the judge prompt tight and calibrate against a few human-labeled answers.

Where the tools fit

You do not hand-build the claim-extract-and-verify loop. Three platforms ship faithfulness scorers with different strengths.

For pytest-style RAG tests, DeepEval. It ships a faithfulness scorer among its 50+ research-backed metrics, and you run it with deepeval test run in CI. The Apache-2.0 framework is free. The catch - nearly every metric is LLM-as-judge, so faithfulness on a big suite fires a lot of inference calls and runs take minutes. Cost-model it before you scale.

For the deepest RAG scoring, Arize Phoenix. Reviewers single out its RAG evaluation as the best in the category, it ships 50+ pre-built metrics with no judge prompts to write, and it runs locally in under a minute. Note the license - the Phoenix server is Elastic License 2.0, source-available rather than OSI open source, which only matters if you plan to resell it as a hosted service.

For attaching faithfulness scores to production traces, Langfuse. It runs LLM-as-judge and custom scorers and lets you attach the resulting scores to live traces, so you can watch faithfulness on real traffic rather than only in an offline suite. It is MIT-licensed and free to self-host.

Bottom line - faithfulness is your RAG hallucination detector, and it works by decomposing an answer into claims and checking each against the retrieved context. Remember its one blind spot - it measures support, not truth - and always pair it with retrieval metrics so you can tell a hallucinating model from a bad source. For the tool-by-tool comparison, see best RAG evaluation tools, and for the end-to-end workflow, how to evaluate RAG.