What four metrics does the RAGAS paper propose and what does each measure?

**RAGAS** is a framework for evaluating retrieval-augmented generation pipelines without armies of human annotators. A RAG pipeline has two halves: a retriever that pulls context chunks, and a generator that writes an answer from them. RAGAS scores both with four metrics. Two metrics judge the generator. Faithfulness asks whether every claim in the answer is actually supported by the retrieved context, so it catches hallucinations. Answer relevance asks whether the answer addresses the question rather than wandering off topic. Two metrics judge the retriever. Context precision asks what fraction of the retrieved chunks are actually relevant, so it catches noisy retrieval. Context recall asks whether the retrieval pulled all the information needed to answer fully, so it catches missing context. The headline feature is that three of the four are reference-free: faithfulness, answer relevance, and context precision need no human-written gold answer. Only context recall needs a reference answer, because you cannot know what was missed without knowing what should have been there. That is why RAGAS scales cheaply for continuous evaluation.

**RAGAS** proposes four metrics that decompose RAG quality into retrieval and generation. Generation side. Faithfulness measures whether answer claims are entailed by the retrieved context. It decomposes the answer into atomic statements and checks each against the context, catching hallucinations. Answer relevance measures whether the answer addresses the question. It generates candidate questions from the answer and scores their similarity to the real one. Retrieval side. Context precision measures the fraction of retrieved chunks that are relevant, rewarding tight retrieval. Context recall measures whether the retrieved context covers every claim in the reference answer, catching missing evidence. The headline property is reference-free scoring. Faithfulness, answer relevance, and context precision need only the question, context, and answer. None needs a human gold answer. Only context recall requires a reference, because measuring what was missed needs a ground truth. This makes three of four metrics cheap to run continuously. In practice you watch them as a panel. Low faithfulness with high context precision points at the generator. Low context recall points at the retriever.