What is recall in machine learning?
Recall = TP / (TP + FN), the true positives over everything that was actually positive. It is also called the true positive rate (Google Machine Learning Crash Course, last updated 12 January 2026). It is the metric that exposes the failure accuracy hides on imbalanced data: a fraud model that flags nothing scores high accuracy and zero recall. Recall ignores false alarms entirely, so a model that flags everything scores a perfect 100 percent. That is why it is never read alone in classification.
Recall vs precision: what is the difference?
They divide the same true positives by different totals.
| Metric | Formula | Question it answers | Falls when the model |
|---|---|---|---|
| Recall | TP / (TP + FN) | Of the real positives, how many did we catch? | Misses cases |
| Precision | TP / (TP + FP) | Of the items we flagged, how many were real? | Raises false alarms |
| F1 score | 2 × P × R / (P + R) | One balance of both | Is weak on either |
A worked example: 1,000 insurance claims, 50 of them fraudulent. The model flags 60, and 40 of those are fraud. Recall is 40 / 50 = 80 percent; precision is 40 / 60 = 67 percent. Moving the decision threshold trades one for the other: flag more and recall rises while precision falls.
What is recall in RAG and retrieval evaluation?
In retrieval, recall@k is the share of relevant documents that appear in the top k results for a query. It is the first number to check in RAG evaluation, because the model cannot use a passage the retriever never returned, and no prompt fixes a retrieval miss. Measuring it needs ground truth: for each test question, the documents that actually answer it. The standard architecture splits the work: retrieve wide for recall, with hybrid search combining keyword and vector matches, then restore precision with reranking before the context reaches the model. Fix retrieval recall before tuning the generator, since generation quality is capped by what was retrieved.
When should a team optimize for recall?
When a miss costs more than a false alarm. Fraud screening, compliance review, safety guardrails that detect harmful content and first-stage retrieval all favor recall. Pair a high-recall stage with a second stage that restores precision, a reranker, a stricter model or a person under human-in-the-loop review, so the extra flags cost reviewer minutes rather than wrong decisions.