What does context recall measure in RAG eval and what ground truth does it require?

**Context recall** is a retrieval-quality metric from the RAGAS framework for evaluating RAG pipelines. It answers one question: did the retriever surface all the evidence needed to support a complete answer? To compute it, you start from a reference answer, the gold-standard response written by a human or trusted source. You break that reference into individual claims. Then you check, claim by claim, whether each one can be attributed to the chunks the retriever returned. The fraction that can be attributed is the context recall score. A high score means the retriever found the evidence. A low score means it missed important chunks, which forces the generator to either hallucinate or give an incomplete answer. So recall isolates a retrieval problem from a generation problem. The key practical point in the question is cost. Context recall needs that reference answer. Writing or curating gold answers is slow and expensive. Its sibling metric, context precision, only checks whether the retrieved chunks are relevant, so it does not need a gold answer and is cheaper to run. The wrong options confuse recall with verbatim overlap, with latency, or with how the model uses chunks internally. None of those is what recall measures.

**Context recall** measures retrieval completeness in a RAG eval. It asks what fraction of the claims in a reference answer can be attributed to the retrieved context. The computation is claim-level. Take the ground-truth answer, decompose it into atomic statements, and check each statement against the retrieved chunks. The recall score is the share of reference claims that the chunks could support. High recall means the retriever surfaced the needed evidence. Low recall means it missed chunks, so the generator must hallucinate or answer incompletely. The ground-truth requirement is the crux of the question. Recall is defined relative to a reference answer, so you cannot compute it without one. That gold answer must be written or curated, which is the expensive part. Contrast that with context precision. Precision asks whether the retrieved chunks are relevant and well-ranked. It judges chunk relevance against the question alone. So precision needs no reference answer and is cheaper to run at scale. The two are complementary. Precision catches noisy retrieval. Recall catches missing retrieval. You accept recall's higher labeling cost when missing evidence is the risk you care about.