Explain context recall in RAG evaluation and why it is more expensive than context precision

In a RAG system, a retriever pulls chunks of text from a knowledge base, then the model writes an answer using those chunks. **Context recall** measures one specific thing: did the retriever pull back everything the answer actually needed? To grade recall you need a reference answer, the gold standard, written by a human or curated from a trusted source. You break that gold answer into individual claims. Then, for each claim, you check whether at least one retrieved chunk supports it. Recall is the share of gold claims that were covered. Context precision asks the opposite-facing question: of the chunks you retrieved, how many were actually relevant? You can judge relevance just by reading the question and the chunk. No gold answer is required. That difference is why recall costs more. Precision needs only the question and the chunks, both of which you already have. Recall needs a hand-written reference answer for every query in your eval set, and producing those answers is slow human work. The practical upshot: teams compute precision cheaply on lots of queries, but recall only on a smaller curated set where the gold answers exist.

**Context recall** measures whether retrieval pulled back all the information a correct answer needs. It is a coverage metric over the reference answer. Computation: take a gold reference answer, split it into atomic claims, then check each claim against the retrieved chunks. The score is a ratio: $$\text{ContextRecall} = \frac{\text{reference claims attributable to retrieved chunks}}{\text{total reference claims}}$$ The required ground truth is a reference answer per query, usually human-written or curated. That is the expensive part. You must build a labelled gold set before you can measure recall at all. Context precision is cheaper because it is reference-free. It only asks whether each retrieved chunk is relevant to the question. An LLM judge scores that from the query and the chunk alone. So the cost gap is structural, not incidental. Precision reuses inputs you already have. Recall demands new annotation that scales with eval-set size. That budget is why teams run precision broadly but recall on a small curated slice. Low recall also caps the pipeline: if a needed fact was never retrieved, no judge or generator can recover it downstream.