Voyage Rerank 2.5

Voyage Rerank 2.5 sits in the second stage of hybrid retrieval pipelines, where a cross-encoder reranker rescores the candidates already merged from BM25 keyword search and dense vector retrieval through Reciprocal Rank Fusion. By operating on richer query-document interactions rather than independent embeddings, the model is designed to address the systematic blind spots each first-stage retriever leaves behind and to lift end-to-end RAG quality without changing the upstream fusion design.

In a 2026 hybrid search reference, the reranker is reported to deliver a +7.9% NDCG gain over Cohere Rerank v3.5 alongside a 32,000-token context that the same source frames as 8× that competitor, which makes it well suited to long-document passages typical of enterprise and legal corpora. On Isaacus's Legal RAG Bench it lands in the top tier of rerankers compared alongside Kanon 2 Reranker and Qwen 3 Reranker 8B, though it is outperformed by those newer models on that legal-domain benchmark. Practically, it is a strong default pick for general-purpose RAG pipelines that want a drop-in reranker after BM25-plus-vector fusion, while legal-specialized teams may want to benchmark newer domain-tuned rerankers first.