Voyage Rerank 2

Rerank-2 is Voyage AI's second-generation reranking model: a cross-encoder that jointly scores query-document pairs to reorder vector-search candidates by true relevance, with a 16K-token context window. It sits in the standard two-stage RAG pattern — fast embedding retrieval first, precise reranking second — and is a stable, well-integrated general-purpose choice with first-party support across LlamaIndex, LangChain, and other frameworks.

The honest picture: rerank-2 is no longer the latest — rerank-2.5 (32K context, instruction-following, multilingual) and rerank-2.5-lite have superseded it as Voyage's recommended models, and the 2.5 line is priced per token ($0.05/M for full, $0.02/M for lite). If you are starting fresh, evaluate 2.5 first; rerank-2's value today is as the stable, battle-tested fallback with wide ecosystem defaults still pointing at it. For self-hosted reranking, BGE and Jina cross-encoders remain the open alternatives.