Context Recall — AI Glossary

Reference details

Related terms

• Context Precision
• Faithfulness
• RAGAS

How it works

1. A ground-truth answer is required, which is the structural constraint on the metric: it cannot be computed on live traffic where no ideal answer exists.
2. That reference answer is decomposed into individual claims, so the unit of measurement is a claim rather than a document or a chunk.
3. Each claim is checked against the retrieved context by a judge, which decides whether the context contains enough to support it.
4. The score is the proportion of reference claims that are supported, so a value of 0.4 means three fifths of what the ideal answer asserts was never retrieved.
5. The system's own generated answer plays no part in the computation, which is what makes the metric a measurement of retrieval alone rather than of the pipeline.
6. That isolation is the diagnostic value: because generation cannot influence the number, a low score localises the fault upstream without any further experiment.

Example: The metric that tells you where to look, and one reason it falls

Answer quality is poor. Context recall is 0.4 — the retriever is not returning the material the answer needs, so no prompt change will fix it. Context recall is 0.95 and quality is still poor — retrieval is fine and the generation step or the prompt is at fault. One number redirects the whole investigation. When it is low, chunk boundaries are worth checking before the retriever itself. A span straddling a boundary is split, and neither half states the whole rule, so both can retrieve while neither supports the claim. With 512-token chunks and 50 tokens of overlap the stride is 462, and a 120-token clause lands across a boundary with probability 120 / 462 ≈ 26.0%. So roughly a quarter of medium-length clauses are split before retrieval is even attempted, which caps recall at a level no amount of reranking recovers.

Commonly confused with

Context recall is not context precision. Recall asks whether the needed material was retrieved; precision asks whether what was retrieved is relevant. You can have perfect recall alongside terrible precision by returning everything, which is why the two are always read together. Recall is also not faithfulness, and the difference is which text gets decomposed: recall splits the reference answer and asks what retrieval covered, while faithfulness splits the generated answer and asks what the context supported — so one needs ground truth and the other can run on live traffic.

When to use it

Reach for it when:

• Diagnosing whether a RAG failure is retrieval or generation
• Tuning chunk size, overlap or top-k with a measurable target
• You have ground-truth answers to evaluate against
• Before blaming the retriever, since chunk boundaries split a quarter of medium clauses above

Reach for something else when:

• Without ground truth — the metric requires an ideal answer to compare with
• As an overall quality score; it measures one stage, not the product
• Confused with faithfulness, which decomposes the generated answer rather than the reference
• On live traffic, where no reference answer exists for it to measure against

Continue your research

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

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