MongoDB adds Voyage 4 AI models & automates vector search

MongoDB has extended its AI product line-up with new Voyage 4 embedding and reranking models and new database features aimed at running retrieval-based AI applications in production.

The company said it has integrated Voyage AI's models into MongoDB's platform infrastructure. It positioned the move as a way to keep operational data and retrieval in one system, rather than splitting workloads across an operational database, a separate vector store, and external model APIs.

The update includes five embedding models from Voyage AI and a new Voyage 4 series. MongoDB also announced an Atlas Embedding and Reranking API, and an automated embedding feature for its community vector search. It also launched an AI-powered data operations assistant for MongoDB Compass and Atlas Data Explorer.

MongoDB framed the announcements around a shift in focus among customers. It said teams now face pressure to run AI systems reliably at scale, rather than trialling proofs of concept.

"The biggest challenge customers face with AI isn't experimentation, it's operating reliably at scale," said Fred Roma, Senior Vice President of Product and Engineering, MongoDB.

Voyage 4 models

The Voyage 4 series adds several options aimed at different trade-offs between accuracy, latency, and cost. MongoDB named voyage-4 as a general-purpose embedding model and voyage-4-large as a model positioned for the highest retrieval accuracy. It also listed voyage-4-lite, which the company said targets latency and cost, and voyage-4-nano, which it described as open-weights for local development, testing, or on-device use.

MongoDB also made the voyage-multimodal-3.5 model generally available. The company said the model expands interleaved text and image support to include video. It said the model targets context extraction from a mix of content such as tables, graphics, slides, and PDFs.

MongoDB said the models are available through MongoDB Atlas via API. It also said Voyage AI remains available as a standalone platform independent of MongoDB.

Database integration

Alongside the new models, MongoDB introduced what it calls Automated Embedding for MongoDB Vector Search. The company said the feature generates and stores embeddings whenever data is inserted, updated, or queried. It presented the approach as a way of keeping embeddings aligned with changing data without a separate pipeline.

MongoDB said Automated Embedding is in public preview. It said the feature has support in drivers such as JavaScript, Python, and Java. It also said it integrates with AI frameworks including LangChain and LangGraph in Python. It said the feature is available for MongoDB Community and will also come to MongoDB Atlas.

MongoDB also announced an "intelligent assistant" for MongoDB Compass and Atlas Data Explorer. It said the assistant provides natural-language support for data operations such as query optimisation.

Customer usage

MongoDB named Tavily and TinyFish as customers using MongoDB for AI-powered features and workloads. TinyFish referenced its evaluation of embedding models and the developer experience around APIs.

"We were looking for extremely accurate embedding models, and Voyage AI provided accuracy at scale," says Sudheesh Nair, Cofounder and CEO of TinyFish. "The Python APIs that Voyage comes out of the box with are also extremely lightweight and very fast."

Tavily's CEO described a focus on product delivery with limited resources.

"Today, companies need to move extremely fast, and at very lean startups, you need to only focus on what you are building," said Rotem Weiss, CEO of Tavily. "MongoDB allows us to focus on what matters most, our customers and our business."

Startup programme

MongoDB also announced an expansion to MongoDB for Startups. The company said the programme now represents more than USD $200 billion in combined valuation across participating companies. It cited Pitchbook data for the figure.

The company described new launch partners as Fireworks AI and Temporal. It said the startup programme will include matched credits across a set of complementary technologies, coordinated onboarding content, and joint events.

MongoDB tied the programme to growing infrastructure choices faced by founders building AI-driven products. It also cited a market forecast for AI growth as context for the pace of change.

"Startups building in the AI era can't waste time in their early years untangling infrastructure mistakes. They need a robust data foundation and a stack that works from day one and scales with their business," said Suraj Patel, VP, MongoDB Ventures & Corporate Development, MongoDB.

Fireworks AI and Temporal also commented on their roles in the programme. Fireworks AI described model and workload changes over time. Temporal referenced workflows and developer experience.

"Fireworks provides a future-proof foundation, allowing customers to evolve models and workloads without rebuilding their AI infrastructure as they grow," said Lin Qiao, CEO and co-founder of Fireworks.

"Temporal is dedicated to helping developers build resilient, scalable applications without the boilerplate," said Samar Abbas, CEO of Temporal.