BGE Reranker Base - RAG Precision Booster for SageMaker

BGE Reranker Base from BAAI is a cross-encoder model that reranks a list of retrieved passages against a query, consistently improving nDCG@10 by 3-8 points over first-stage embedding retrieval alone. It is the missing second stage in most RAG pipelines: your vector search narrows candidates to 20-100 passages, the reranker scores each (query, passage) pair directly and returns a precise ranked order.

Cross-encoders see both query and passage together in a single forward pass -- unlike bi-encoders that embed each independently -- which lets them capture fine-grained relevance signals that embedding similarity misses. The trade-off is higher per-passage latency, which is why rerankers are applied to a small candidate set, not the full corpus.

Deploy as a SageMaker endpoint. The endpoint accepts {"query": "...", "passages": ["...", "..."]} and returns a relevance score per passage. Rerank your top-20 vector search results and pass only the top-3 to the LLM -- better answers, lower token cost.

Primary use cases: RAG pipeline precision improvement, enterprise document search, e-commerce product ranking, legal case retrieval, and any two-stage retrieval architecture where first-stage recall is high but precision needs improvement.