What is Recall and Precision?
Also called Precision and Recall.
Recall and precision are the two complementary measures of retrieval quality. Recall is the share of all relevant items that a system actually returned, while precision is the share of returned items that were relevant. Improving one often degrades the other, so retrieval systems are tuned toward a target balance rather than a single number.
Consider a corpus containing ten passages that answer a question. If the system returns twenty passages and eight of them are among the ten, recall is eighty percent and precision is forty percent. Returning fewer results generally raises precision and lowers recall, while returning more results does the opposite.
In retrieval for generation the two stages have different priorities. The first stage should favor recall, because a passage never retrieved can never be used no matter how good the later stages are. The reranking and selection stages should favor precision, because irrelevant passages in the prompt dilute attention and invite mistakes.
Both are usually measured at a cutoff, written as recall at ten or precision at five, since ranked lists are consumed from the top. Related summary measures include mean reciprocal rank, which rewards placing the first relevant result early, and normalized discounted cumulative gain, which accounts for graded relevance across the list.
Measurement requires labeled data: a set of representative questions with the passages known to answer them. Building even a modest labeled set of a few dozen questions turns retrieval tuning from guesswork into measurement, and it is the step most often skipped. Without it, changes to chunking or models cannot be judged.
Key points
- Recall: how much relevant material was returned.
- Precision: how much of what returned was relevant.
- Retrieval favors recall, reranking favors precision.
- Measured at a cutoff, such as recall at ten.
- Requires a labeled question set to evaluate at all.
In practice
A team changes chunk size and measures the effect on fifty labeled questions. Recall at twenty rises from 0.72 to 0.86, meaning the correct passage now reaches the shortlist far more often, while precision at five falls slightly. Because a reranker follows, the tradeoff is accepted and end to end answer accuracy improves.