voyage-context-3: Contextual Retrieval Without the LLM
9 months ago
- #Retrieval-Augmented Generation
- #AI
- #Embedding Models
- Introduction of voyage-context-3, a contextualized chunk embedding model that captures full document context without manual metadata augmentation.
- voyage-context-3 outperforms OpenAI-v3-large, Cohere-v4, Jina-v3 late chunking, and contextual retrieval by significant margins in both chunk-level and document-level retrieval tasks.
- Supports multiple dimensions and quantization options, reducing vector database costs while maintaining high retrieval accuracy.
- Addresses chunking challenges in RAG by capturing both focused detail and global context without tradeoffs.
- Seamless drop-in replacement for standard embeddings, compatible with existing vector databases and workflows.
- Less sensitive to chunking strategies, reducing system sensitivity and improving retrieval performance.
- Evaluation shows superior performance across various domains and real-world datasets.
- Offers Matryoshka embeddings and quantization options, significantly cutting costs while preserving accuracy.
- Available now with a free tier for the first 200 million tokens, ideal for long documents and high-sensitivity retrieval tasks.