Context precision
Context precision asks one question about what your retriever pulled back: of the chunks you handed the model, how many were actually relevant to the question. High precision means little junk in the pile. A system can look fine on other retrieval metrics while still burying the one good chunk in noise, and precision is what exposes that.
Why it matters
Every irrelevant chunk is a chance for the model to grab the wrong detail and answer confidently off it. Low precision also wastes your context window and token budget on text that does nothing. When answers feel vaguely off or quote the wrong policy, bad precision is often why. The right answer was in there, buried under five chunks that were not.
How it works
You mark each retrieved chunk as relevant or not for the question, then look at the fraction that were relevant, usually weighted so chunks ranked near the top count for more. A human can do the labeling, but more often an LLM judge reads the question and each chunk. It pairs with context recall, which checks the opposite failure: whether you retrieved everything you needed, not just whether what you got was clean.
A support bot gets asked about the refund window and retrieves eight chunks: the refund policy, plus seven about shipping, return labels, and gift cards. One of eight was relevant, so context precision is low. Now the model has to find the real answer inside a wall of off-topic text, and that is exactly when it grabs the shipping timeline by mistake.