Recall vs Precision
Also called Recall vs Precision
The trade-off between catching everything relevant and only returning what’s actually relevant.
Think of it like
A fishing net: a wide net catches every fish but also boots and seaweed; a narrow one keeps the catch clean but misses fish.
Example
Set retrieval to return 20 chunks and you rarely miss the answer (high recall) but drag in noise; return 3 and every chunk is on-point (high precision) but you sometimes miss the key one.
How it actually works
Recall is the share of all relevant items you actually retrieved; precision is the share of what you retrieved that was relevant. Turning one dial usually moves the other the wrong way. In RAG the sweet spot depends on the model: too few chunks and it lacks the answer, too many and it drowns in distractors. Re-ranking is the usual way to get precision back after casting a wide, high-recall net.
For product teams
The core quality lever in search — decide whether missing an answer or showing junk hurts your users more.
For engineers
Recall = TP/(TP+FN), precision = TP/(TP+FP); tune top-k to trade them, then restore precision with a reranker.
Related
- Reranking — The usual tool to raise precision after wide recall.
- Retrieval Latency — How many results you fetch sets the balance.
- RAG — A quality dial inside RAG.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome