Of all the actual positive cases in the data, the fraction the model correctly identified — a measure of how much the model misses.
Recall answers the question: of all the actual positive cases, how many did the model catch? Formally, it is the number of true positives divided by all actual positives (true positives plus false negatives). A recall of 0.85 means the model found 85% of the genuine positives and missed 15%.
Recall is the metric to prioritise when missing a positive case is costly. A well-formedness classifier used to flag queries for expansion should have high recall — it's better to over-flag slightly ambiguous queries for review than to miss genuinely ambiguous ones and leave them unhandled. A grounding classifier with low recall would fail to trigger live search for queries that genuinely needed it, producing outdated or hallucinated answers.
Recall and precision pull in opposite directions. Lowering the decision threshold catches more actual positives (higher recall) but also catches more false positives (lower precision). The right balance depends on what the model is used for. When neither error type dominates, the F1 score — the harmonic mean of the two — provides a single summary number.
For models with severe class imbalance, recall for the minority class is often the number that matters most — it's easy to score high overall accuracy by always predicting the majority class, but recall will reveal that the model never finds the rare cases it actually needs to catch.
