← All models · BGE-reranker-v2-m3
use case · wild
The hardest case for search: distractors in the query's own language, packed with its exact words but answering the wrong question — while the real answer is written in a different language, sharing no words at all. Keyword search ranks the same-language decoys first; a bi-encoder often can't bridge the language gap. Because v2-m3 reads query and passage together on a multilingual backbone, it promotes the foreign-language answer despite zero lexical overlap. The contrast below ranks by the cross-encoder on the left and by shared-word overlap on the right.
The lexical column can only score same-language word matches, so the real answer (in another language) sinks to the bottom there with 0% overlap — exactly where keyword search would bury it. The cross-encoder reads for meaning across the language boundary and pulls it to the top. This is the case only a multilingual reranker gets right.
← Practical · Back to BGE-reranker-v2-m3 · Next: Multi-model →