Introducing voyage-context-3: focused chunk-level details with global document context

Voyage-context-3 is a novel contextualized chunk embedding model that enhances retrieval accuracy by capturing both local and global document context without requiring manual metadata or context augmentation. As a drop-in replacement for traditional context-agnostic embeddings, it simplifies retrieval-augmented generation (RAG) systems by reducing sensitivity to chunking strategies and offering improved performance compared to established models like OpenAI-v3-large and Cohere-v4. By leveraging Matryoshka learning and quantization-aware training, voyage-context-3 maintains high retrieval quality across multiple dimensions while significantly lowering vector database storage costs. The model excels in both chunk-level and document-level retrieval tasks, outperforming rivals across various domains and real-world datasets, making it especially effective for handling long, unstructured documents and high-sensitivity retrieval tasks in fields such as finance, medical, or legal domains.