BAAI/bge-reranker-base · Hugging Face

Abstraction: BAAI BGE cross-encoder reranker for RAG retrieval pipelines

Key points:

• bge-reranker-base is a cross-encoder that takes a (query, document) pair as input and directly outputs a similarity score — more accurate than bi-encoder embeddings but slower
• Recommended use: retrieve top-100 with a fast embedding model, then re-rank with bge-reranker to get final top-k
• BGE-M3 (Jan 2024) supports 100+ languages, up to 8192 token input, and unifies dense, lexical, and multi-vector (ColBERT) retrieval in one model
• bge-large-en ranked #1 on MTEB leaderboard; bge-large-zh ranked #1 on C-MTEB (31 Chinese datasets, 6 tasks)
• Similarity scores are not bounded (trained with cross-entropy loss); for filtering use relative ordering, not absolute thresholds
• FlagEmbedding library (MIT license) integrates with sentence-transformers, LangChain, and HuggingFace Transformers

Connections: Baai · Hugging Face · Retrieval Augmented Generation · Text Embeddings · Information Retrieval

Source: https://huggingface.co/BAAI/bge-reranker-base