Faithfulness
An eval metric measuring whether every claim in an answer is supported by the retrieved context.
Simple explanation
Faithfulness asks: 'Is everything the model said actually in the sources?' A faithful answer doesn't add, embellish or hallucinate. It's the #1 RAG eval metric.
Engineering explanation
Faithfulness is scored by an LLM-as-judge (different model from the system) that checks each claim in the answer against the retrieved context. Score = (supported claims / total claims). Production RAG targets faithfulness > 0.85. Below 0.7 = hallucination risk. Eval gate blocks deploy if faithfulness drops > 5% from baseline.
Example
Answer: 'Refunds take 5-7 days and include a 10% restocking fee.' Context only mentions 5-7 days. Faithfulness = 0.5 (one claim supported, one hallucinated).
When to use it
Any RAG eval pipeline. The primary generation-quality metric alongside answer_relevance.
Common misconception
Faithfulness is NOT the same as accuracy. Faithfulness = supported by context. Accuracy = factually true. An answer can be faithful (matches wrong context) but inaccurate.