Answer Relevance

Also known as: response relevance

Answer relevance measures how directly and completely a model's response addresses the user's question. A response can be factually accurate and faithful to its source context yet still score low on relevance if it answers a related but different question, buries the key point, or omits important aspects of what was asked. Answer relevance is typically assessed by an LLM judge or by comparing the response embedding to the question embedding.