Faithfulness

What it means

Faithfulness, sometimes called groundedness, checks whether an answer stays true to the context it was given rather than inventing information. The common way to compute it is to break the response into individual factual claims, then verify each claim against the retrieved documents. The score is the fraction of claims that are supported. A faithfulness of one means every statement traces back to the source; a low score means the model added material the context did not justify. Note that faithfulness is about support, not about correctness in the wider world - an answer can be faithful to a wrong document.

Why it matters

For retrieval-augmented generation, faithfulness is the direct measure of hallucination. It isolates the generation step from retrieval - it asks whether the model used the context honestly, regardless of whether the context itself was any good. That separation is valuable because it tells you where a bad answer came from. Unlike accuracy, faithfulness is reference-free - it needs the retrieved context but not a hand-written ground-truth answer, which makes it cheap enough to run on live production traffic and not just a curated eval dataset.

In practice

Faithfulness is usually paired with answer relevancy and context precision to give a rounded view of a RAG pipeline - relevancy asks whether the answer addressed the question, context metrics ask whether retrieval fetched the right documents, and faithfulness asks whether the answer stuck to them. The scoring itself is typically done by an LLM judge that performs the claim extraction and verification. Tools like DeepEval provide a ready faithfulness scorer, and platforms like Langfuse let you attach the resulting scores to traces. See the RAG evaluation guides for the full metric set and how to combine them.