← Back to concept explainers

Concept Fable

Retrieval recall

Conclusion first:Retrieval Recall measures the proportion of all relevant results that the system successfully retrieves. In RAG, it answers the question: was the key evidence recalled? Low recall causes the model to answer based on incomplete context, which is especially dangerous in compliance, contract, after-sales and healthcare scenarios. Recall can be improved through better chunking, metadata filters, hybrid search, query expansion, multi-query retrieval and more suitable embedding models, but reranking is still needed to control noise.

Retrieval Recall · RAG evaluation

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Whether the evidence that should have been found was brought back

The county office is hearing a land deed case, and the clerk is responsible for finding evidence in the archives. The archive holds old deeds, tax registers, neighbors' testimony, and past rulings. Whether the case can be judged clearly depends largely on whether the right materials are brought back.

On the first day, the clerk brought back only the three neatest contracts. They all supported the plaintiff, and the county magistrate almost ruled. But the defendant, an old farmer, said that thirty years earlier there had been an agreement about the water channel, and that if that document was not examined, the boundaries would be wrong.

The clerk went back to the archive to search again and indeed found the water channel register in a damp corner. The county magistrate realized that the materials brought back looked clean but had missed key evidence, and the entire judgment had been built on a gap.

For the first fix, the secretary decided to grab more rather than less. He brought two large boxes of paper, stuffed with irrelevant tax receipts, old cases from neighboring villages, and duplicate copies. The court spent the whole day flipping through paper, and the key materials ended up buried at the bottom.

The old archivist taught him to first list the "types of evidence that must be found": original deeds, subsequent changes, tax records, testimony from neighboring land, and past disputes. For each type, at least one credible source had to be found; if not found, the gap had to be marked.

When the legal advisor handled cases again, he did not seek to empty the archive, nor was he satisfied with the few most pleasing pages. What concerned him was whether important evidence that should appear had been recalled, especially the part that would change the judgment.

The county magistrate later often asked one question: has all the evidence that should come arrived? If the answer was no, then no matter how elegant the argument, it was only standing on empty ground. Before finding accurately, one must first find enough to support a judgment.

Later, every time the secretary submitted a report, he attached a gap sheet: which evidence had been found, where no records could be located, and what still needed review. When the county magistrate saw the gaps, he actually dared to judge more, because he knew what materials the judgment stood on. Recall does not seek to bring the whole world over, but to bring back as much as possible of what should affect the conclusion. The old archive clerk later carved this sentence beside the archive door. Before entering, the new secretary first thought clearly about what materials the case needed, then began searching. How much evidence is brought back is not important; what matters is whether evidence that would affect the judgment has been left in the dark. Later the county office had far fewer overconfident verdicts, because gaps would be seen first. This sentence was later written into the newcomers' handbook, reminding everyone to look first at the cost on site before talking about saving effort. And precisely because of this reminder, when similar difficult problems arose later, people would not rush to apply old methods.

Reveal

This story is about: retrieval recall

Retrieval Recall measures the proportion of all relevant results that the system successfully retrieves. In RAG, it answers the question: was the key evidence recalled? Low recall causes the model to answer based on incomplete context, which is especially dangerous in compliance, contract, after-sales and healthcare scenarios. Recall can be improved through better chunking, metadata filters, hybrid search, query expansion, multi-query retrieval and more suitable embedding models, but reranking is still needed to control noise.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Missing supplementary agreement: key evidence was not recalled
  • The returned materials are all accurate: high precision but possibly low recall
  • Required evidence checklist: golden labels needed to evaluate recall
  • Master agreement, DPA, and approval records: enterprise issues often require multiple pieces of evidence to support together
  • Parent-child chunks and metadata: an index design that improves recall
  • Hybrid retrieval: combine keywords and vectors to reduce missed recalls
  • Key evidence never absent: the core value of RAG recall

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

precisionrerankingtop-kground truth

Related distinctions

This concept is easily confused with adjacent concepts; it is recommended to continue to the comparison page.

京ICP备2026014272号