Context recall.
The fraction of ground-truth information that the retriever actually found. Range 0 to 1, higher is better.
How it’s computed
The judge decomposes the ground-truth answer into atomic claims and checks which of those claims are attributable to the retrieved chunks.
context_recall = count(ground_truth_claims in retrieved)
/ count(all_ground_truth_claims)Unlike most RAGAS metrics, context recall is not reference-free — it requires a ground-truth answer in the dataset so the judge has a canonical list of claims to check against.
Worked example
The ground-truth answer to “What does Yoke’s RAG workbench do?” contains 5 atomic claims (ingestion, dataset generation, grid-search, scoring, improvement report). The retrieved chunks cover 4 of them. context_recall = 4 / 5 = 0.8.
How Yoke Agent uses it
Context recall is a RAGAS core metric, used whenever you have ground-truth answers in the dataset. It drives decisions about chunk size, overlap, and retrieval strategy — low recall tells you the retriever is missing relevant material, independent of how it ranks what it did find.
Frequently asked
Does this require a ground-truth answer?
Yes. This is the one RAGAS metric that is not reference-free. If you don’t have ground truth, you can’t compute context recall.
What usually improves a low score?
Chunking first (smaller chunks or adding overlap), then retriever breadth (top-k up, or hybrid dense+BM25). Advanced strategies like HyDE and Multi-Query help on multi-hop questions.
Is there a trade-off with context precision?
Yes. Retrieving more chunks helps recall but usually hurts precision. The sweet spot depends on your corpus.