Retrieval Recall
Retrieval recall measures what fraction of the actually relevant documents or chunks for a given query a RAG system's retriever successfully finds — a low-recall retriever misses relevant information entirely, which no amount of good generation downstream can fix, since the model can only work with what it was given.
Recall failures are quietly more damaging than precision failures in most RAG systems, because a model faced with irrelevant extra context can often still ignore it and answer correctly, but a model that was never given the one chunk containing the actual answer has no path to a correct response — it will either say it doesn't know (best case) or hallucinate a plausible-sounding answer from its own general knowledge (worst case, and hard to distinguish from a grounded answer without checking).
Recall is also harder to eyeball-test than precision, since a missing chunk produces no obvious error signal — the output just looks like an ordinary wrong or vague answer, with nothing pointing back at retrieval as the cause unless someone specifically checks what was retrieved.
Measuring it requires a labeled test set where the ground-truth relevant document(s) for each query are known in advance, then checking what fraction of those were actually present in what got retrieved — usually reported as recall@k.
