How a search ranking system uses a two-stage funnel to meet latency limits

Scoring every item in a large catalog with an expensive ranking model is computationally infeasible within a request's latency budget. The practical design instead narrows millions of candidates to a few hundred through fast retrieval, then applies the costly model only to those survivors. Candidate generation typically combines lexical matching with semantic embeddings searched via approximate nearest-neighbour methods. The ranker relies on rich features from a feature store and learns from user clicks and watches, which are biased and must be handled carefully. The article presents this retrieval-then-rank pattern, split between offline model building and online serving, as the backbone of modern search and recommendation systems.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in