Precision Recall: Definition and Examples

Precision and recall are fundamental metrics in artificial intelligence and information retrieval. They allow evaluating the performance of a system that must identify relevant items among a dataset. These two measures are inseparable because they capture two different facets of result quality.

Precision answers the question: 'Among all the items the model identified as positive, how many are actually positive?' For example, if a spam filter classifies 100 emails as spam and 90 of them are indeed spam, the precision is 90%. A high-precision system produces few false positives. Recall answers the question: 'Among all the actually positive items, how many were correctly identified?' If the inbox contains 120 total spams and the filter detects 90, recall is 75%.

There is generally a trade-off between these two metrics, known as the precision-recall trade-off. Increasing precision tends to decrease recall, and vice versa. A very conservative system will have high precision but low recall (it only flags cases it is sure about), while a permissive system will have high recall but lower precision (it captures everything, including false positives). The F1-score, harmonic mean of precision and recall, helps find a balance between the two.

In the context of prompt engineering, understanding these metrics helps formulate more effective instructions. When asking an LLM to extract information or classify content, you can guide its responses toward more precision ('return only results you are certain of') or more recall ('list all possible items, even uncertain ones'). This understanding is essential for calibrating expectations and refining obtained results.