Context Precision — AI Glossary

Reference details

Related terms

• Context Recall
• Faithfulness
• RAGAS

How it works

1. Each retrieved chunk is graded individually for whether it bears on the question, rather than the retrieval being scored as a set.
2. The metric is the fraction of retrieved chunks judged relevant, so it is a property of what the retriever returned and not of the answer.
3. Grading is usually done by a model, which means the metric inherits that grader's error rate and needs its agreement measured before the number is trusted.
4. Position can be weighted, since an irrelevant chunk early in the context distracts more than one at the end.
5. It pairs with context recall: precision asks how much of what came back was useful, recall asks how much of what was needed came back, and a system can fail either way.

Example: Why precision and recall fail differently

Retrieve 10 chunks of which 3 are relevant, and the corpus held 4 relevant chunks in total. Precision is 3/10 = 0.30 — seven chunks of distraction in the prompt. Recall is 3/4 = 0.75 — one needed chunk missed. Lowering k to 3 might give precision 3/3 = 1.00 with recall unchanged at 0.75. The answer may improve because the generator is no longer wading through noise, and nothing about the retriever changed. The two metrics move independently, which is why one alone cannot diagnose a RAG failure.

Commonly confused with

Easily read as a measure of the answer, when it is a measure of the retrieval. A system can score perfect context precision and still answer wrongly, because the generator ignored the context it was given — that is groundedness, a different metric. It is also not context recall: precision is the share of returned chunks that were useful, recall the share of useful chunks that were returned. Tuning k trades them against each other directly, so reporting one without the other hides which half of retrieval is broken.

When to use it

Reach for it when:

• Diagnosing whether a RAG failure originates in retrieval rather than generation.
• Tuning k or a reranker, where the question is how much noise reaches the prompt.
• Alongside context recall, since the pair localises the failure and either alone does not.

Reach for something else when:

• Judging answer quality, which needs groundedness and answer relevance instead.
• Without measuring the grader's own agreement, which bounds how small a change the metric can detect.
• As a single number for RAG health, since it says nothing about what retrieval missed.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Context Precision — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path