A versatile embedding model from BAAI that simultaneously supports dense retrieval, sparse retrieval, and multi-vector retrieval, with multilingual support for 100+ languages and multi-granularity processing from short sentences to 8192-token documents.
BGE-M3 stands for Multi-Functionality, Multi-Linguality, and Multi-Granularity. It is a groundbreaking embedding model that can simultaneously perform three common retrieval functionalities in a single model.
Uses the normalized hidden state of the [CLS] token as the dense embedding for semantic similarity search.
Generates sparse vectors (vocabulary-sized with mostly zeros) calculating weights only for tokens present in the text, similar to BM25 but learned.
Uses multiple vectors to represent text, enabling fine-grained similarity matching at the token level.
The optimal setup is hybrid retrieval + re-ranking:
On the MLDR test set (13-language long document retrieval):
Free and open-source. Available through various commercial API providers with usage-based pricing.
Loading more......