Why it matters

BGE is top embedding family. Understanding shapes design.

Advertisement

The architecture

Contrastive pretraining.

Fine-tuning on retrieval data.

BGE-M3: dense + sparse + multi-vector.

BGE embedding familyContrastive pretrainInfoNCEFine-tune on retrievalsupervised pairsBGE-M3 multi-modedense + sparse + multi-vecBAAI open-source; BGE-M3 recent multi-mode; top on MTEB
BGE math.
Advertisement

How it works end to end

BGE-base + BGE-large + BGE-M3 variants.

Reranker BGE-reranker.

MTEB leaderboard top.