rerank-2.5
Rerank 2.5 introduces instruction-following ranking to Voyage's reranker family. Instructions can express which properties make a document relevant.
Model design
The release combines revised training data with distillation from larger internal instruction models. A ranking instruction can be appended or prepended to the query. The model supports a 32,000-token query-document pair.Instruction-following rerank-2.5 models
Practical use
Practical guidance. Use explicit criteria when relevance depends on more than topic, such as a document type or required evidence. Keep hard access restrictions in application filters. A ranking instruction is not an authorization rule.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API |
| Atlas lifecycle | Legacy since 29 September 2026MongoDB Voyage AI model lifecycle: Atlas scope and release history |
| Query limit | 8,000 tokensVoyage AI reranking API and earlier models |
Related models
- rerank-2.5-lite
Voyage AI · Voyage Rerank 2.5
Smaller instruction-following reranker with long-document support.
Launch date API Instruction-following reranker with a 32,000-token pair limit.
Launch date APISecond-generation reranker with a 16,000-token pair limit.
Launch date API