rerank-2.5 Reranker

Overview

Rerankers are neural networks that predict the relevancy scores between a query and documents and rank them based on the scores. They are used to refine search results in semantic search/retrieval systems and retrieval-augmented generation (RAG). rerank-2.5 significantly improves upon the rerank-2 performance while also introducing instruction-following capabilities for the first time. On the Massive Instructed Retrieval Benchmark (MAIR), rerank-2.5 outperform Cohere Rerank v3.5 by 12.70%. The accuracy of rerank-2.5 is increased by an average of 8.13% on 24 domain-specific instruction-following datasets across 7 domains (web, tech, legal, finance, conversational, medical, and code). It supports a 32K-token context length, an 8x increase over Cohere Rerank v3.5. Latency is 1.5 s for 25K tokens, and throughput is 60M tokens per hour at $0.05 per 1M tokens on an ml.g6.xlarge. Learn more about rerank-2.5 here: https://blog.voyageai.com/2025/08/11/rerank-2-5/ 

Highlights

• First reranker with instruction-following capabilities, allowing users to dynamically steer the reranking process by providing explicit instructions alongside their query. These instructions can define the user notion of relevance or specify the desired characteristics of the documents to be retrieved.

• rerank-2.5 is 7.94% more accurate than Cohere Reranker v3.5, additional 8.13% performance gain if with instruction.

• rerank-2.5 improves retrieval quality over rerank-2 by 1.85% while increasing to a 32K context length. Latencies are 1.5 s (1 GPU), 415 ms (4 GPUs), and 245 ms (8 GPUs) for 25K tokens. We recommend using multiple GPUs to reduce latency. The supported 12xlarge and 24xlarge instances come with 4 GPUs each, while the 48xlarge instances are equipped with 8 GPUs. 60M tokens per hour at $0.05 per 1M tokens on an ml.g6.xlarge