BGE
A one-stop retrieval toolkit for search and retrieval-augmented generation, providing a family of open embedding, reranking and multimodal-search models known as BGE.
BGE, developed under the FlagEmbedding project, is a one-stop retrieval toolkit for search and retrieval-augmented generation, built around a family of embedding and reranking models for retrieval and retrieval-augmented large language models.
Recent releases include BGE-VL, a multimodal embedding model supporting visual search in any direction, including text-to-image, image-to-text and combined image-and-text search, released with a synthetic training dataset called MegaPairs; an embedding model that uses in-context examples to produce richer query representations; a multilingual embedding model built on Gemma-2-9B reporting new results on multilingual retrieval benchmarks; and a lightweight reranker built on the same base model that uses token compression to reduce resource use while keeping accuracy.
The project also links to related community work, including a unified image-generation model called OmniGen and a memory-inspired retrieval-augmented-generation approach called MemoRAG.