Reranking : trier ce que la recherche a trouvé
The second pass of a document search. The first one, retrieval, moves fast across a large collection: it compares vectors computed in advance and brings back, say, the fifty passages closest to the question. Reranking takes that short list and rereads each passage with the question in view, using a model that scores every question and passage pair, then reorders. Only the top few go on to the model that writes the answer. The two steps do different jobs. Retrieval is fast because it summarised each document as a vector before the question even existed, and that summary loses information. The reranker is more accurate because it actually reads, and far slower for the same reason: Pinecone estimates that running 40 million documents through a reranker would take more than fifty hours, where vector search answers in under 100 milliseconds. Hence the order: find wide and fast, then sort short and well.