MongoDB gives on-premises developers hybrid search and vector features

MongoDB Inc. today said its search and vector search features that had previously been confined to its Atlas cloud platform are now available for self-managed deployments, including MongoDB Community Edition and MongoDB Enterprise Server.

The feature puts advanced artificial intelligence-building tools in the hands of millions of developers operating local and on-premises environments. Now in public preview, the additions enable developers to test and build AI-powered applications without relying on third-party search engines or vector databases. They provide access to full-text, semantic and hybrid search capabilities natively embedded within MongoDB’s architecture.

“This is the same technology we built inside of Atlas,” said Ben Cefalo, senior vice president and head of core products at MongoDB. “That’s important because we want the applications you’re building to react the same way, whether you’re running them inside your own data centers or on Atlas.”

In the past, implementing these features often required bolting on external systems, which can create synchronization issues, increase operational overhead and add cost. Developers can now run hybrid queries combining keyword and vector search, manage unstructured data such as chat messages, documents and videos locally and build agentic AI systems with persistent memory.

As AI agents become increasingly central to software development, databases are evolving from passive storage systems into intelligent memory engines, MongoDB said. “A large language model provides reasoning, but it’s the database that gives the agent its real power, memory and state,” Cefalo said.

Version 8.2 debuts

Alongside the new features, MongoDB also introduced MongoDB 8.2, adding what the vendor said are significant performance improvements, including up to 49% faster unindexed queries and nearly triple the throughput for time-series bulk inserts. The update also adds substring support to MongoDB’s queryable encryption, a client-side technology that encrypts sensitive data before it reaches the database and allows for expressive queries to be run on the encrypted data without the need for decryption.

“None of this matters if you can’t trust your database,” Cefalo said. “We have a battle-tested foundation with years of enterprise application road miles pushing on performance, availability and security.”

Acquisitions have played a major role in MongoDB’s AI push. Its purchase of Voyage AI Inc. early this year brought a set of embedding and re-ranking models that enhance the platform’s ability to serve high-context, domain-optimized data to LLMs. AI embeddings are numerical representations of data, or vectors, that capture the meaning and context of all types of data in a form that computers can understand and compare. They are critical to enabling AI models to understand semantic meaning, perform similarity searches and power recommendation systems.

“A good embedding model versus a great embedding model is the difference between fuzzy recollection you have after waking up from a dream and the crisp, vibrant memory you have from the most important events of your life,” Cefalo said.

MongoDB claims more than 75% of Fortune 100 companies are customers, including seven of the world’s 10 largest banks and 13 of the 15 largest auto manufacturers.

Photo: MongoDB