← Back to concept explainers

Concept Fable

Context precision

Conclusion first:Context Precision measures how much of the context retrieved or provided to the model is truly relevant to the question in the earlier positions. It focuses not only on whether the answer was retrieved, but also on whether relevant material is ranked near the front and whether there is too much irrelevant noise. For RAG systems, high Context Precision makes it easier for the model to answer based on correct evidence; low Context Precision makes it vulnerable to noise.

Context Precision · RAG evaluation

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

Fable

The librarian hid the answer in the seventh book

The library in Fengling Town was very large. When townspeople asked questions, the librarian would bring over a stack of books and let them find the answers themselves. The librarian had read a lot and knew which books might be relevant. But the more books were brought over, the more readers got lost at the table.

One day an apprentice at the medicine shop asked, "Which old prescription should I look at for a night cough?" The administrator brought over ten medical books, nine of which mentioned coughs, and only the seventh recorded the subdivision of nighttime cough and cold-dampness. The apprentice flipped through them all morning and almost copied down the prescription for daytime cough.

The curator first asked the administrator to find more books, afraid of missing the answer. The table went from ten books to twenty, and the apprentice became more confused. Finding everything seemed safe, but the more irrelevant material there was, the easier it was for the real answer to be drowned out.

The second repair was to sort the books by thickness, looking at thin books first. Speed increased, but it still could not guarantee that most of the content on the desk was truly relevant. The librarian realized that the problem was not whether the answer had been carried over, but how much of the material carried over could directly help.

So the library changed its rules. After each retrieval, the librarian had to mark which pages directly supported the answer, which were merely background, and which, though they shared the same words, should not be put on the table. If most of a stack of books was useless, then even if the answer was hidden among them, it could not count as a good search.

Later, readers received less material, but it was more accurate. The curator understood that retrieval is not about piling everything possibly relevant onto the table, but about giving the reader context with as few impurities as possible; the cleaner the desk, the less easily the answer is led astray by noise.

The administrator later also retained the ability to "find everything." When encountering difficult cases, he would first expand the scope, then remove irrelevant pages, and place only the passages that could support an inference on the answer desk. Broad finding and precise placement were split into two steps, and readers were no longer intimidated by a sea of book spines.

The curator spot-checked the desk once a week: if the answer page appeared but was surrounded by nine-tenths miscellaneous books, it still had to be redone. Because the material finally handed to the reader affects judgment. Precise context is not about showing off how many books one has, but about letting useful evidence occupy the main position. Later, when the administrator trained apprentices, he did not reward the one who carried the most books, only the one who could remove irrelevant books. The reader asks for the answer, not a mountain of books. The closer the context is to the question, the steadier the later judgment. If there are too many impurities in the material, even a very clever reader later will be led astray.

Reveal

This story is about: context precision

Context Precision measures how much of the context retrieved or provided to the model is truly relevant to the question in the earlier positions. It focuses not only on whether the answer was retrieved, but also on whether relevant material is ranked near the front and whether there is too much irrelevant noise. For RAG systems, high Context Precision makes it easier for the model to answer based on correct evidence; low Context Precision makes it vulnerable to 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

  • Librarian: retrieval system or RAG pipeline
  • Bringing over a stack of books: retrieval results or context fragments provided to the model
  • The answer is in the fifth book: the relevant evidence was retrieved, but the ranking wasn't good enough
  • Bridge history, legends, ballads: seemingly relevant noise context that does little to help the current problem
  • Put the wood-wax manual first: truly relevant evidence comes first
  • Whether the answer is in the top few: Context Precision focuses on the relevance of the earlier part of retrieval results
  • Too much noise in relevance is also a form of slowness: context quality directly affects model answer quality and efficiency

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.

context recallfaithfulnessRAGAS
京ICP备2026014272号