May 2025 Summaries

Voyage-3.5 and voyage-3.5-lite are new embedding models that improve retrieval quality over their predecessors, voyage-3 and voyage-3-lite, while maintaining the same price and size. These models offer embeddings in various dimensions and utilize Matryoshka learning and quantization-aware training to enhance performance. They outperform the OpenAI-v3-large model in multiple domains, reducing vector database costs significantly with their smaller embedding dimensions and quantization options. Evaluations across 100 datasets in domains such as technical documentation, code, law, finance, and multilingual content show that voyage-3.5 and voyage-3.5-lite achieve superior retrieval quality, demonstrating state-of-the-art cost-performance ratios. The models introduce a binary rescoring method that further enhances retrieval quality while minimizing database costs. Both models are currently available, with the first 200 million tokens offered for free.

Since joining the MongoDB team two months ago, Voyage AI has been making significant strides in its mission to develop advanced embedding models and rerankers, ensuring these models remain accessible to all teams regardless of their use of MongoDB. The partnership has accelerated Voyage AI's progress through increased investment in engineering and research, enhanced compute resources, and a commitment to hiring new talent. This synergy has already benefited teams like Continue, which successfully utilized Voyage AI models to enhance their open-source AI code assistant. Looking ahead, Voyage AI is focused on scaling its APIs to handle increased traffic, developing new models for AI applications, and integrating automated text embedding into MongoDB Atlas Vector Search for a seamless semantic search experience. Additionally, MongoDB customers will soon have direct access to Voyage AI's model APIs, simplifying the adoption of these cutting-edge tools. The Voyage AI team is eager to engage with developers through events and platforms, offering opportunities to explore their models and stay informed about ongoing developments.