Precision and Recall

Precision and recall are the two halves of a classifier's or retriever's quality, and they answer different questions. Precision asks: of the items the system flagged, what share were actually correct? Recall asks: of the items that should have been flagged, what share did the system find? A spam filter with high precision rarely sends good mail to the junk folder but may let spam through; one with high recall catches nearly all spam but occasionally quarantines a real invoice. Accuracy — the share of all predictions that were right — hides this trade-off and becomes actively misleading when one class is rare, since a model that never flags anything can be 99% accurate on a 1% problem while finding nothing. The trade-off is usually a threshold, not a fixed property. Lowering the score at which the system acts raises recall and lowers precision; raising it does the reverse. Which side to favour is a product decision, not a modelling one: it depends on what a false positive costs relative to a false negative in your workflow. Where a single number is needed, the F1 score combines the two as a harmonic mean, but reporting F1 alone hides which side of the trade-off a change moved. In retrieval-augmented systems the same pair applies to the retrieval step, where recall usually matters more up front — a passage that is never retrieved cannot be reranked — and precision is recovered later by reranking.