bge-reranker-base
BAAI/bge-reranker-base
A solid general-purpose reranker for improving retrieval precision in RAG pipelines. Small enough to be cheap, good enough to matter.
about this model
The BAAI/bge-reranker-base is a cross-encoder reranking model hosted on gigarouter as an OpenAI-compatible API. It takes a query and a set of candidate documents and directly outputs a relevance score, allowing for more precise reordering of retrieval results than embedding-based similarity alone. This model is designed to be used as a second-stage re-ranker after an initial retrieval step (e.g., using a BGE embedding model) to improve top-k accuracy.
Key Strengths
- Cross-encoder architecture provides higher ranking accuracy than embedding-based cosine similarity, at the cost of higher per-pair latency.
- Supports both Chinese and English input.
- Part of the BAAI general embedding (bge) family, built on a base-scale transformer.
- An updated version (March 2024) extends input length, adds multi-lingual support, and delivers major performance gains on standard benchmarks.
Best for
- Re-ranking the top 100 documents returned by a fast dual-encoder retriever to produce final top-3 results.
- Applications requiring higher relevance precision, such as question answering, search, and retrieval-augmented generation (RAG).
Performance
The updated reranker achieves massive improvements in ranking performance on BEIR, C-MTEB/Retrieval, MIRACL, and LlamaIndex Evaluation benchmarks relative to prior versions. Specific scores are not listed in the model card, but the model is recognized as a state-of-the-art cross-encoder for its size.
# rerank documents by relevance; billed per document curl https://gigarouter.ai/v1/rerank \ -H "Authorization: Bearer $GR_KEY" \ -d '{"model":"BAAI/bge-reranker-base","query":"capital of France", "documents":["Paris is the capital of France.","Bananas are yellow."]}'