← All models · BGE-reranker-v2-m3
use case · multi-model
The two-stage pipeline behind real search and RAG, running entirely in your browser with two models. Stage 1: a multilingual E5 bi-encoder embeds every passage and the query, and ranks by cosine similarity — fast, but only surface-level. Stage 2: the v2-m3 cross-encoder reads each shortlisted (query, passage) pair together and reorders for precision. Both stages are multilingual, so an English query surfaces and correctly reranks answers written in other languages. Compare the two orders below.
The left column is what embeddings alone would return; the right column is after the cross-encoder rereads the shortlist. When the top result changes, that's stage 2 earning its keep — the reranker is slower per pair, which is exactly why you only run it on the shortlist the retriever already narrowed down.
← Wild · Back to BGE-reranker · Related: Multilingual embeddings → · v1 reranker →