Latest MongoDB tools tackle top AI development hurdles

Taking on trouble

Despite significant advancements in AI model capabilities, many enterprises struggle to build agents and other AI applications they can trust to perform accurately in production.

This release is a nice step forward for MongoDB. Companies of all sizes need to improve the accuracy of their data retrieval, especially when it comes to the text documents that provide critical context for agentic AI.

Often, the problem preventing successful AI development is poor data retrieval.

Without the right context, provided by an enterprise's unique proprietary data and business logic, AI tools won't have the situational awareness to deliver outputs that can be trusted, and development projects will fail. However, discovering and operationalizing the relevant information required to successfully build AI tools is difficult.

In response, this year an array of vendors have introduced capabilities aimed at connecting agents with the context they need. In June alone, AWS, Databricks, Microsoft and Snowflake have all made context for agents a focal point of product development initiatives.

MongoDB has similarly focused on improving data retrieval for AI in 2026. In January, the vendor launched a set of five Voyage models that categorize data to make it discoverable, and in May it added new vector embedding capabilities that similarly aid data discovery.

Now, MongoDB is attempting to further improve data retrieval for AI in a move driven by both customer feedback and MongoDB's own recognition of the problems that stalled customers' AI projects, according to Benjamin Cefalo, the vendor's chief product officer of core products.

"Teams would get an AI project most of the way to production and then hit one of two walls: the retrieval wasn't accurate enough to trust, or the infrastructure couldn't meet their data residency requirements," he said. "And the accuracy problem has gotten more expensive in a very literal sense."

When models chase the wrong context, development costs increase, with models looping back to try to find the right context. Meanwhile, enterprises attempt to fix the problem by adding more systems such as a new search engine or vector database, Cefalo continued.

"The conclusion was pretty simple," he said. "If retrieval is what makes AI accurate and affordable, it shouldn't be stitched on beside the database. It should live inside it, where the data already is, whether that’s on-premises, in a private cloud, or in a hybrid environment."

Specific new MongoDB features designed to improve data retrieval for AI include the following:

• Native Reranking in MongoDB Atlas, which is built on the Voyage AI embedding and ranking models to improve data retrieval accuracy.
• Voyage Context 4, a new embedding model for long documents that is designed to understand the context of full documents so they can be used to inform AI tools.
• Hybrid Search in MongoDB to combine full-text and vector search in a single query in the vendor's operational database.

Voyage Context 4 and Hybrid Search are generally available, while Native Reranking in MongoDB Atlas is in preview.

While some AI initiatives are stopped by poor data retrieval, others -- particularly those by enterprises in highly regulated industries such as healthcare, finance and energy -- are stalled by data sovereignty rules and regulatory requirements that make cloud-based AI development difficult.

To address the needs of those organizations, MongoDB is enabling users to build AI tools behind firewalls with the launch of Search and Vector Search for its Enterprise Advanced edition.

In addition, Search and Vector Search are now available in MongoDB's free Community Edition, and the vendor launched support for Apache Iceberg tables in its Atlas Stream Processing.

Measuring up

The reranking capabilities are perhaps the most significant individual feature for MongoDB users, according to Leone. Meanwhile, from a competitive standpoint, bringing capabilities together rather than forcing users to piece together the pieces of an AI development workflow helps distinguish MongoDB from its peers, he continued.

"The individual capabilities are becoming common across the field, [but] what sets MongoDB apart is having all of it in one place -- the Voyage model, the operational database the app data already lives in and the search running on that live data," he said.

However, even as MongoDB evolves and its competitors become data and analytics platforms similarly pushing into AI tooling, without governance and large-scale analytics, MongoDB is not yet a full-featured data management provider, according to Leone.

"MongoDB is right there with them, and arguably ahead on one running accurate search directly on the live application data instead of a copy," he said. "Its tougher climb is that those platforms already hold a lot of a customer's data, so they get a head start when a company wants to buy everything from one vendor."

Petrie likewise noted that as MongoDB expands beyond its roots, it is finding ways to stand apart from vendors that are similarly evolving.

Specifically, he highlighted capabilities that allow users to combine different data types to train models and connect agents with proper context.

"This is a time for MongoDB to shine," he said. "AI adopters need context to ensure their [models] and agents generate trustworthy outputs. MongoDB enables them to organize and consume the multimodal data -- especially text -- that contains rich context. The convergence of operational and analytical workloads also … supports agentic AI."

Looking ahead

As MongoDB evolves, its aim is to move from a system of record that enables organizations to store data to a system of intelligence for AI, according to Cefalo.

Toward that end, the vendor's roadmap includes improving the memory and retrieval layer for agents so they act on trusted data, providing capabilities that enable organizations to run workloads in their preferred environments, and consolidating previously disparate tools to simplify development.

"For any agentic workload there are three layers -- the harness, the model and the data -- and the data layer is the one we intend to own," Cefalo said.

Leone, meanwhile, suggested that after addressing data retrieval, MongoDB add capabilities that help customers test the accuracy of retrieval for AI applications.

"They've made retrieval more accurate, and the natural next step is helping customers see and prove how accurate it is in their own environment, which is exactly what a regulated buyer has to show an auditor," he said. "A first-party evaluation layer for retrieval would round out the compliance story and give those teams one less thing to source elsewhere."

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.