BGE Reranker Base

BGE Reranker Base (BAAI/bge-reranker-base) is a cross-encoder reranking model from the Beijing Academy of Artificial Intelligence's BGE family. Where bi-encoder embedding models encode queries and documents independently, this model processes each query-document pair jointly through the full transformer, producing a fine-grained relevance score used to reorder retrieval results.

It is a standard second-stage reranker in RAG pipelines: fast vector search returns candidates, and the cross-encoder re-sorts the top results for precision. The model supports Chinese and English, takes up to 512 tokens, and is released under the MIT license. The release date below is era-level (2023), matching the BGE family release window.

It ships as a default option in many open-source reranking stacks (including FastEmbed) and is frequently cited as the strong, simple baseline that newer rerankers are measured against.