Which metric best measures whether a RAG answer is grounded in the retrieved context?

**RAG** systems retrieve documents, stuff them into the prompt, and ask the model to answer based on those documents. The most important thing to evaluate is whether the model actually USED the retrieved documents, or whether it ignored them and made something up. **Faithfulness** is the metric designed for exactly this. The procedure: decompose the model's answer into atomic claims, then for each claim check whether the retrieved context entails it. If every claim is supported, faithfulness is 1.0; if half the claims have no source, it is 0.5. This is what catches hallucinations in RAG. The other options measure different things and miss the question: - BLEU measures n-gram overlap between the answer and a reference answer. It cares about word-level similarity to a gold answer, not about whether the claims came from the retrieved context. - ROUGE also measures n-gram overlap, typically against a reference summary. Same problem: surface similarity, not grounding. And ROUGE against the QUERY (as the option states) is even less meaningful. - Perplexity measures how confidently the model produced the output. A confident hallucination has low perplexity, so this misses the point entirely. Faithfulness is the answer.

**Faithfulness** is the right metric because it asks the specific question 'are the claims in the answer supported by the retrieved context?' The standard implementation, from the RAGAS framework: 1. Decompose the candidate answer into atomic claims. 2. For each claim, ask a judge whether the retrieved context entails it. 3. Faithfulness = (claims supported) / (total claims). A score of 1.0 means every claim has a source; 0.5 means half the answer is hallucinated; 0.0 means the answer ignored retrieval entirely. The other options fail because they measure the wrong axis: - BLEU and ROUGE measure surface overlap against a reference. They reward style matching, not grounding. A fluent fabrication that uses the right vocabulary scores high; a correct answer in different words scores low. - Perplexity measures the model's confidence in its output. A confident hallucination has LOW perplexity, which is the opposite of what you want. Faithfulness is one of four canonical RAG metrics (the others being answer relevance, context precision, and context recall). All four together give a full picture of retrieval and generation quality.