It may rewrite the query, search several indexes, filter candidates, fuse scores, rerank passages, and fetch source text. Evaluating each stage separately helps locate recall, latency, and relevance failures.
A retrieval pipeline is the sequence of processing steps that turns a query into ranked evidence.
It may rewrite the query, search several indexes, filter candidates, fuse scores, rerank passages, and fetch source text. Evaluating each stage separately helps locate recall, latency, and relevance failures.