Concept Contrast Fable

Retrieval recall and context recall

Conclusion first:Both metrics are called "recall," and both relate to "completeness." But one measures how much was retrieved from the warehouse, while the other measures whether what actually reached the workers' hands is complete enough. Misreading the numbers means missing real process weaknesses.

Both metrics are called "recall," and both relate to "completeness." But one measures how much was retrieved from the warehouse, while the other measures whether what actually reached the workers' hands is complete enough. Misreading the numbers means missing real process weaknesses.

Fable

Two completeness-rate reports

Old Chen is a quality inspection supervisor at a parts assembly factory. The factory has two inspectors, one surnamed Zhao and one surnamed Qian. Both their jobs are called "checking whether it's complete," but they stand in different positions.

Inspector Zhao stood at the warehouse door. Each time a new order came in, workers went to the warehouse to pick parts, and Inspector Zhao checked: were all the parts that should be picked actually picked? Was anything missed? Was anything picked wrong? At month's end, Inspector Zhao's numbers looked great—a pick completeness rate of ninety-eight percent.

Qian Quality Inspection stood beside the assembly station. His job was to check: did the parts package actually delivered to the assembly station contain everything? Had anything fallen during transport? Had anything been swapped out temporarily? On the month-end report, Qian Quality Inspection's numbers did not look so good—assembly completeness was only eighty-four percent.

The factory manager looked at the two reports and found it strange. For the same batch of orders, the warehouse pick completeness rate was 98 percent, so how did it drop to only 84 percent by the assembly bench? A gap of fourteen points in between—where did it go?

The factory manager called Inspector Zhao and Inspector Qian together to reconcile order by order. As soon as they compared, the problem surfaced: some parts, though taken out of the warehouse, were temporarily pulled aside by the dispatcher in the sorting area and diverted to rush orders—Inspector Zhao had no idea, because his inspection only went as far as the warehouse door. Other parts had the right specs but the wrong version; old-version parts counted as "retrieved" in the warehouse, but couldn't be installed at the assembly station, so Inspector Qian could only mark them as "missing."

The factory manager understood. Inspector Zhao was measuring the stretch from the warehouse to the sorting area, while Inspector Qian was measuring the state of the parts package when it finally reached the assembly station. Both used the same ruler called "completeness rate," but they were not measuring the same process segment.

From then on, the factory manager added a handover registration point between the sorting area and the assembly bench, so every transfer of every kind of part was recorded. He also stopped looking at only one number—warehouse completeness rate and assembly completeness rate were tracked separately, and whoever dropped a part was responsible for it.

At the year-end summary, the factory manager shared one lesson with peers: do not assume the assembly bench is fine just because the warehouse reports look good. In a system, between being obtainable and being usable lies an entire logistics chain.

Retrieval recall

Measures how many relevant documents a retrieval system found from the knowledge base. It focuses on whether the retrieval step was complete, independent of how those documents are used afterward.

Context recall

Measures whether the information finally packed into the model's context window covers all key evidence needed for the answer. It focuses on 'whether what's ultimately used is complete,' and after retrieval it may still undergo truncation, re-ranking, and filtering.

Story Mapping

Warehouse entrance: the retrieval system exit; retrieval recall measures this segment.
Assembly table: the context window the model ultimately sees; context recall measures this segment.
The sorting-area dispatcher pulls out parts: reranking, truncation or filtering after retrieval drops relevant content.
The wrong version cannot be installed: the retrieved document content is relevant but the information does not match, and it is judged missing in context evaluation.
A fourteen-percentage-point gap: the difference between two recall rates is usually a problem with ranking and truncation strategy in a RAG system.

In the project

Each time you evaluate a RAG system, record two recall rates at the same time. If retrieval recall is low, the problem lies in the vector model, indexing strategy, or knowledge base coverage. If retrieval recall is high but context recall is low, the problem lies in reranking, truncation windows, or filtering logic. Putting both numbers side by side in the same table is more useful than looking at either one alone.

Retrieval recall Context recall