TL;DR – voyage-context-4 is our next-generation contextualized chunk embedding model – it produces chunk embeddings that capture the full document context without any manual metadata or context augmentation, so you can stop thinking about chunking. A new mixture-of-experts (MoE) backbone delivers better context-aware embeddings, while built-in auto-chunking, transparent handling of documents longer than 32K tokens, and native support for overlapping chunks together remove chunking as a design concern. Averaged across 39 datasets spanning 8 domains voyage-context-4 outperforms voyage-context-3 by 1.4% and 2.08% on document-level and chunk-level retrieval, respectively, and surpasses voyage-4-large, our best general-purpose embedding model, by 0.4% on single-embedding evaluation. It’s a drop-in replacement for voyage-context-3 and is priced at $0.12 per 1M tokens versus $0.18 per 1M tokens for voyage-context-3.
We’re excited to introduce voyage-context-4, the next generation of our contextualized chunk embedding model, where each chunk embedding encodes not only the chunk’s own content but also the contextual information from the full document. voyage-context-4 improves retrieval quality in nearly every domain while adding auto-chunking and transparent handling of documents of any length – together, these remove the need to engineer a chunking strategy at all.
