Voyage's rerank-2.5 beats Cohere by 7.94% on search accuracy and follows instructions
A reranker in a RAG pipeline takes roughly a hundred retrieved documents and moves the most relevant ones to the top. Voyage AI has released rerank-2.5 and rerank-2.5-lite, which outperform Cohere Rerank v3.5 by 7.94% and 7.16% on result quality across 93 datasets. The models understand hints like "look for statutes, not court cases" and read up to 32 thousand tokens. Pricing is unchanged, so teams should test the swap on their own data.
AI-processed from Voyage AI; edited by Hamidun News
Imagine searching a knowledge base for "Jaguar." The first retrieval stage returns a hundred documents about cars, cats, and the operating system of the same name. To get the right one to the top, RAG systems use a reranker: a model that takes the query and the hundred candidates and reorders them by actual relevance.
Voyage AI has released a new generation of these models: rerank-2.5 and the lighter rerank-2.5-lite.
The 7.94% in the headline means that on average result quality (the NDCG@10 metric, which measures how high the relevant documents rank in the top ten), rerank-2.5 beats Cohere Rerank v3.
5 by 7.94%, and the lite version beats it by 7.16%.
The measurement covers 93 datasets. The second novelty: the reranker can now be told the task in ordinary words. This is worth reading for anyone building search or RAG and choosing what to reorder results with.
What changed in the models
The company compared the models on 93 datasets across nine domains: technical documentation, code, law, finance, internet reviews, multilingual search, long documents, medicine, and dialogue. The multilingual portion includes 51 datasets in 31 languages. The evaluation works like this: a first-stage retriever (BM25, OpenAI's text-embedding-3-large, and others) returns up to 100 candidates, the reranker reorders them, and the top 10 are taken.
Averaged across the four first-stage retrievers, rerank-2.5 outperforms Cohere Rerank v3.5 by 7.94%, and rerank-2.5-lite by 7.16%. They also beat the open Qwen3-Reranker-8B, by 2.25% and 1.47% respectively, and Voyage stresses that its models are more than an order of magnitude smaller. The previous Voyage generation trails by 1.85% for the larger model and 1.08% for the smaller one. Voyage also claims both models are better than LLMs used as rerankers, but it promises the details in a separate article, so for now this is an assertion without figures.
The company also notes that Cohere Rerank v3.5 can degrade results when paired with the strongest first-stage retriever, while the new models improve results with any first-stage method. For practitioners, this is a signal: test a reranker on your own data rather than adopting one by default.
The context window has grown to 32 thousand tokens. That is eight times Cohere Rerank v3.5's and twice Voyage's previous models. Long documents such as contracts, specifications, and regulations can now be submitted whole, without truncation. The price per token has not changed.
Instructions instead of fine-tuning
The main feature is instruction following. A natural-language phrase is added to the query (or placed before it), and the model adjusts its relevance scores accordingly. Voyage cites three types of tasks:
- Emphasis. In a search over scientific papers, you can specify "weigh the title, ignore the abstract."
- Document type. A legal service can enforce "look for regulations and statutes, not court cases."
- Context clarification. The phrase "this is an app about cars" helps the model understand that "Jaguar" is a car brand, not an animal.
The examples in the article show the difference. Asked which group is at risk for hand, foot, and mouth disease, the older model returned a generic answer, "children." With the instruction "focus on age," the top result became a passage about children under 10 and seasonal outbreak patterns. In another example, a user asks what "unlocking a device" means and explains that they are a businessperson with a work phone. With that hint, the model prioritizes an answer about switching carriers while traveling under corporate rules.
In numbers: across 24 internal datasets from seven domains, instructions yield a gain of 8.13% for rerank-2.5 and 7.55% for the lite model. On three datasets built from real applications, the gains were 11.48% and 7.83%. On the MAIR academic benchmark, which includes task-specific instructions, the models beat Cohere Rerank v3.5 by 12.70% and 10.36%.
What to do in practice
The models are available now on a pay-per-token basis, and new users get 200 million tokens free. Customers already on the previous generation should upgrade, Voyage advises: quality is higher, the context is twice as long, and the price is the same. The older models will remain available for those who do not want to switch.
RAG developers should try two things. First, move standing rules into the instruction: document type priorities, language, industry. This is cheaper than retraining a model or building filter layers. Second, recheck document chunking settings: with a 32K-token window, some texts no longer need to be split so finely.
There are limitations too. All figures come from Voyage itself, including those on internal datasets, and independent measurements are not yet available. The 7–8% gain in NDCG@10 is an average across a set and may differ on your corpus. The company also recommends pairing the reranker with its own embeddings, so part of the material is marketing. Still, the combination of higher accuracy, long context, and unchanged pricing makes rerank-2.5 a serious candidate to replace Cohere in production.
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
- Nov 21, 2026Developer Skips Pinecone: Searching 72,450 Code Chunks in 50 ms
- Nov 21, 2026Gemini 4 Neon shrank the TS Rust binary from 47.44 to 7.97 MB in under 10 hours
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
Key Questions Answered
What changed in the models?
The company compared the models on 93 datasets across nine domains: technical documentation, code, law, finance, internet reviews, multilingual search, long documents, medicine, and dialogue. The multilingual portion includes 51 datasets in 31 languages. The evaluation works like this: a first-stage retriever (BM25, OpenAI's text-embedding-3-large, and others) returns up to 100 candidates, the reranker reorders them, and the top 10 are taken.
What are the key details on Instructions instead of fine-tuning?
The main feature is instruction following. A natural-language phrase is added to the query (or placed before it), and the model adjusts its relevance scores accordingly. Voyage cites three types of tasks: Emphasis. In a search over scientific papers, you can specify "weigh the title, ignore the abstract." Document type. A legal service can enforce "look for regulations and statutes, not court cases." Context clarification.
What to do in practice?
The models are available now on a pay-per-token basis, and new users get 200 million tokens free. Customers already on the previous generation should upgrade, Voyage advises: quality is higher, the context is twice as long, and the price is the same. The older models will remain available for those who do not want to switch. RAG developers should try two things. First, move standing rules into the instruction: document type priorities, language, industry.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.