ModelRefs / Faithfulness — AI Glossary

Faithfulness — AI Glossary

A RAG evaluation metric measuring whether generated answers are fully supported by the retrieved context without hallucination.

Overview

Faithfulness (RAGAS metric) scores whether each claim in the generated answer can be attributed to the retrieved context. A faithful response avoids making claims beyond what the retrieval supports. Automated evaluation uses an LLM judge to decompose statements and verify attribution. Distinct from answer relevance (correctness to the query).

Reference details

Topic	evaluation
Last reviewed	2026-06-24

Example: Faithful and useless are compatible

Ask what the refund window is. The retriever returns the shipping policy. A model that answers “the provided context does not state the refund window” is perfectly faithful — every claim is supported — and completely unhelpful. Another that answers “30 days” from parametric memory may be right and is unfaithful, because nothing retrieved supports it. Faithfulness scores the second one worse, which is correct for what it measures and wrong as a summary of quality. It has to be read beside answer relevance and beside retrieval quality; alone, it rewards abstention and cannot distinguish a good retriever from a cautious generator.

Commonly confused with

Faithfulness asks whether the answer follows from the retrieved context; answer relevance asks whether it addresses the question; context precision and recall describe what retrieval supplied. Only the last pair can tell you the retriever was at fault. A generation metric will happily blame the model for a retrieval failure, which is the most common misreading in RAG evaluation.

When to use it

Reach for it when:

RAG systems where unsupported claims are the primary risk
Alongside retrieval metrics, so a low score can be attributed to the right stage
Enforcing citation requirements, where every claim must trace to a source

Reach for something else when:

As a single quality number — it is maximised by refusing to answer
Without checking what the judge counts as a claim; decomposition drives the score
Where the answer legitimately requires general knowledge the corpus does not contain

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Faithfulness — AI Glossary.

Frequently asked questions

What is Faithfulness?