What is Context Recall?
Context Recall is a RAG evaluation metric that measures whether the retrieved context contains the evidence needed to answer the user's question.
How It Works
Context recall asks whether the retrieval system found the necessary evidence at all. A RAG answer cannot be reliably grounded if the required source passage never reaches the model. Low context recall can come from poor chunking, weak embeddings, missing metadata, strict filters, inadequate top-k, or queries that need rewriting. It should be evaluated together with context precision because blindly increasing recall can flood the model with irrelevant context.
Key Characteristics
- Measures evidence coverage rather than evidence cleanliness
- Identifies whether required facts appear in the retrieved context
- Complements context precision, which penalizes irrelevant retrieved chunks
- Sensitive to top-k, query rewriting, chunking, filters, and retrieval model choice
- A critical diagnostic when RAG answers are incomplete or unsupported
Common Use Cases
- Checking whether gold evidence appears in top-k retrieval results
- Diagnosing RAG failures where the model never saw the right source
- Comparing dense, sparse, and hybrid retrieval recall
- Tuning top-k and metadata filters without losing required evidence
- Building retrieval regression tests for high-value questions
Example
Loading code...Frequently Asked Questions
What does low context recall indicate?
It means the retrieval stage failed to include the evidence needed for a grounded answer, so the generator may guess or answer incompletely.
Can increasing top-k improve context recall?
Often yes, but it can also reduce context precision by adding irrelevant chunks. Reranking and better retrieval are usually needed too.
How is context recall different from answer recall?
Context recall evaluates retrieved evidence coverage, while answer recall evaluates whether the final answer includes expected information.
What improves context recall?
Better chunking, hybrid retrieval, query rewriting, metadata repair, embedding selection, and carefully tuned filters can all help.