MongoDB adds AI retrieval tools for regulated firms

MongoDB has introduced new artificial intelligence retrieval and compliance features for its database platform, extending search tools to cloud and self-managed environments.

The launch targets two common obstacles in enterprise AI projects: unreliable information retrieval and infrastructure limits in regulated settings. The update includes Voyage Context 4, Hybrid Search and Native Reranking. Search and Vector Search are also now available for MongoDB Enterprise Advanced and Community Edition.

MongoDB aims to make AI retrieval work inside the same database that stores operational data, rather than through separate search systems. The approach is intended to keep retrieval tied to current data and let customers run the same functions in public cloud, private cloud, on-premises and local environments.

Ben Cefalo, Chief Product Officer, Core Products, MongoDB, said enterprises often hit infrastructure limits rather than model limits when they try to move AI into production.

"The biggest barrier to enterprise AI in production and at scale isn't the LLM. It's memory, retrieval, accuracy, and compliance. Most enterprises aren't blocked by ambition. They're held back by infrastructure that wasn't designed to provide AI with trusted access to enterprise data. Bolting on more systems to solve those problems only creates more vendors, more latency, and more points of failure," Cefalo said.

"Whether you're running in the cloud, private cloud, or behind a firewall, MongoDB gives you the same production-grade retrieval capabilities wherever your data lives," he said.

Search changes

At the centre of the announcement is a set of features intended to improve how AI systems find and rank information. Native Reranking, in public preview in MongoDB Atlas, works on top of existing search results inside the database. MongoDB said it can improve retrieval quality by up to 30%, based on Voyage instruction-following rerankers on the MAIR benchmark.

Voyage Context 4 is now generally available and is built for long documents. It processes long documents in full context rather than breaking them into isolated chunks, which can help preserve meaning across complex material.

Hybrid Search is also generally available. It combines full-text search and vector search in a single query, allowing systems to match both exact terms and semantic meaning without separate products.

These tools are powered by Voyage AI models. MongoDB said embeddings can remain current automatically, so AI agents retrieve from the current state of operational data rather than from a separate copy.

Compliance focus

The other part of the launch centres on where AI workloads can run. Search and Vector Search are now generally available as an add-on for MongoDB Enterprise Advanced, which targets organisations running infrastructure in their own data centres or private cloud estates.

That matters for businesses in finance and other regulated sectors, where data residency, sovereignty and internal compliance rules can restrict the use of public cloud services. MongoDB said more than 20 large banks and financial institutions had been evaluating Search for Enterprise Advanced before the release.

Search and Vector Search are also now generally available for MongoDB Community Edition. That gives developers access to full-text, vector and hybrid retrieval in self-managed environments without a licence fee, offering a local route for experimentation before moving workloads to other MongoDB deployments.

Customer example

MongoDB used Emergent Labs as an example of how the retrieval changes are meant to work in practice. The AI-native application development platform first tested its platform on PostgreSQL. According to MongoDB, agents became stuck in schema migration loops as users refined ideas.

On MongoDB Atlas, agents can create and modify data structures as applications change, while search and embeddings remain in the same database as the underlying data, according to the company.

"Our agents write code, modify data structures, and act on what they read back millions of times a day. If retrieval returns something stale or wrong, the agent builds on it, and the error compounds. MongoDB gives us the retrieval accuracy to keep agents working from the current state of the data, and that's what lets us run two million applications at scale," said Mukund Jha, CEO of Emergent Labs.

The broader announcement also included other database updates, including Apache Iceberg support in Atlas Stream Processing, new dedicated cluster infrastructure on AWS for Atlas M30 and above, and asymmetric search node deployment for multi-region Atlas clusters. MongoDB said the latter can lower total search node costs in multi-region clusters by 25% to 40% or more.

MongoDB also set out an education target in India through an expansion of its MongoDB for Academia programme. The initiative, delivered with the All India Council for Technical Education, HCL GUVI and the ICT Academy of Kerala, has reached more than 650,000 students since 2023, according to the company.