Recall@k

Recall at k asks a simple question: when we take the top k results, did the right material make it in? You build a labeled set of real queries, each paired with the passages or documents that answer it, run retrieval, and count how often the relevant items appear in the first k results. Some teams compute it per relevant item, others as the share of queries with at least one hit; be consistent.

For RAG, it's the most important retrieval number, because the model can only use what you send it. If the answer isn't in the top k passages you include in the prompt, no prompt or model can fix that. Choose k to match your pipeline: recall at the size of your candidate list for first-stage retrieval, and recall at the number of passages you actually send after reranking.

Recall ignores order within the top k, so pair it with a rank-sensitive metric such as mean reciprocal rank. It also says nothing about the generated answer, which needs its own metrics.

Track it on every change to chunking, embeddings, index settings, or retrieval logic. Small regressions are easy to miss by eye and easy to catch with a number.