What does context recall measure in RAG eval and what ground truth does it require?
Context recall measures whether retrieved chunks cover every claim in a reference answer. It needs ground truth, which makes it costlier than context precision.
Imagine a student writing an essay using only photocopied pages a librarian handed them. Context recall asks one question: did the librarian hand over every page the student needed to write a complete, correct essay? To check this, you first need the model answer, the essay the teacher already wrote. You take each fact in that model answer and ask whether it could have come from one of the handed-over pages. If most facts trace back to the pages, recall is high, the librarian did a good job. If half the facts are missing from the pages, recall is low, the student would have to guess or leave things out. The catch: you can only run this check if a teacher already wrote the model essay. That model answer is the expensive ingredient.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Context recall is one of the four core retrieval and generation metrics in the RAGAS framework, alongside context precision, faithfulness, and answer relevance. It targets a specific and easy to miss failure mode in a RAG pipeline: the retriever quietly omitting evidence that the answer actually needs.
The metric is defined relative to a reference answer. That single design choice, the ground-truth dependency, is what the question is testing and what separates context recall from every other commonly-used RAG retrieval metric. Understanding why recall needs a gold answer, while precision does not, is the difference between picking the right metric for a problem and measuring the wrong thing confidently.
This deep dive defines context recall precisely, walks the claim-level computation RAGAS uses, contrasts it with context precision on both inputs and cost, and shows where each metric belongs in a production eval stack.
What context recall actually measures
Context recall measures retrieval completeness: of all the information a correct, complete answer needs, how much did the retriever actually surface? It is the recall side of a classic precision-recall split, applied to the retrieval step of a RAG pipeline rather than to generation. The word retrieval is load-bearing. Recall scores the retriever, not the generator, not the embedding model in isolation, and not the prompt. It is a verdict on whether the chunks that landed in the context window were enough.
The intuition maps cleanly onto the information-retrieval definition of recall, which is true positives over true positives plus false negatives. A false negative here is a piece of needed evidence that the retriever failed to return. If the retriever misses a third of the evidence, recall is roughly two-thirds, and the generator is now working with an incomplete picture. Notice that recall says nothing about ranking or noise. You can retrieve fifty junk chunks alongside the three you needed and still score perfect recall, which is exactly why recall must be read alongside precision rather than on its own.
The consequence of low recall is concrete. When the needed evidence is absent from the context window, the generator has only two options: leave the answer incomplete, or fill the gap by hallucinating. Either way the product degrades, and crucially the degradation traces to retrieval, not to the language model. Swapping in a stronger generator will not fix a recall problem, because the missing fact was never in front of the model to begin with. Context recall is the metric that localizes the blame to the retriever, which is what makes it actionable: a low recall score points your engineering effort at chunking, indexing, and retrieval depth rather than at prompt tuning.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- RAGAS computes context recall by decomposing a reference answer into claims and entailment-checking each against retrieved context with a judge like Claude Opus 4.7 or GPT-5.5.
- TruLens exposes a context-relevance metric in its RAG triad that plays the precision role, paired with groundedness for the faithfulness side.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you build a golden set to support context recall without it going stale?
Stratify by query type, version each reference answer against a corpus snapshot, and recompute recall whenever chunking, the index, or the embedding model changes. Treat the golden set as code under review, not a one-time artifact.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing context recall with context precision, or with faithfulness. Recall scores the retriever's coverage against a reference answer; precision scores chunk relevance without one.
60 second bullets to scan on the way to the call.
Definition of context recall as reference-claim coverage by retrieved context
Why context recall requires a ground-truth or reference answer
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.