Context Precision and Recall

This concept gives you the two numbers that tell you whether your retrieval system is actually finding the right evidence for your AI assistant. Without these, you are flying blind — you cannot tell if a bad answer came from bad retrieval or bad generation. Mastering this lets you diagnose and fix retrieval problems, which is the first step toward building reliable AI systems that answer from your documents, not from guesswork. This unlocks further study in RAG components like reranking, hybrid search, and agentic retrieval — all of which are about improving precision or recall.

Imagine you ask an AI assistant a question, and it goes to a library to find the answer before responding. The library returns a stack of books, but you want to know: Did it give me the right books, and did it miss any important ones? Precision and recall are exactly these two questions, measured with numbers. Precision: Of all the books it gave me, how many were actually relevant? Recall: Of all the books that were relevant in the whole library, how many did it actually give me? High precision means no junk — every book it pulled is on-topic. High recall means no misses — every book that had the answer was pulled. Most real systems trade one off against the other: if you pull only one book, precision might be high but recall low; if you pull fifty, recall goes up but precision drops. In AI retrieval, we measure these on the retrieved chunks of text that get inserted into the model's context before it writes an answer.

Here is a richer analogy to carry through. Suppose you are a detective looking for clues about a crime. You have a file cabinet (the corpus). You send your assistant to pull files. Precision: Of the files pulled, how many actually contain relevant clues? Recall: Of all files in the cabinet that contain relevant clues, how many did the assistant bring back? If the assistant brings back 10 files but only 3 have relevant clues, precision is 30%. If there were 20 relevant files in the cabinet but only 3 were brought, recall is 15%. Now here is the crucial asymmetry: if precision is low, you can throw away the irrelevant files — no harm to the investigation, just wasted time. But if recall is low, you have missed clues entirely — you might never solve the case. That is why the library notes say 'Low recall is unfixable downstream — the evidence simply is not there.' In AI, low precision wastes the model's context budget (it gets distracted by irrelevant chunks), but low recall means the model can only guess or hallucinate. The analogy stops when you realise that in retrieval, you do not know the true set of relevant files — you can only estimate it by judging the retrieved set and maybe sampling the corpus.

Context precision is the fraction of retrieved chunks that are relevant to the query; context recall is the fraction of all relevant chunks in the corpus that were successfully retrieved.

This concept belongs to retrieval evaluation, part of the broader field of information retrieval. It connects forward to RAG evaluation frameworks like RAGAS, which decompose quality into context precision, context recall, answer relevancy, and answer correctness. It also underlies every technique you will learn later: reranking improves precision, hybrid search improves recall by combining keyword and semantic matching, agentic retrieval improves recall through iterative search, and chunking strategy directly affects both. Your library notes say that retrieval and answer quality are separately measurable and separately fixable — knowing these two numbers lets you fix the right half of a RAG system.