With the release of its latest features, MongoDB takes aim at two key issues that are preventing companies from moving their AI projects into production.
While not the only hurdles to successfully building agents and other AI applications, inaccurate data capture and infrastructure that prevent companies from meeting regulatory compliance requirements are among them.
New MongoDB features built on the vendor’s Voyage AI models, such as native reranking in MongoDB Atlas and Voyage Context 4 embedded models, work to improve data retrieval, while features like MongoDB Search and Vector Search are now available in the vendor’s Enterprise Advanced Edition, allowing customers to run AI workloads on-premises, in private clouds, and in hybrid environments.
The new tool was announced Tuesday at user conference MongoDB.local Bengaluru in India.
“The characteristics of accuracy deepen” [MongoDB’s platform]“But the bigger change is reach,” Moor Insights & Strategy analyst Mike Leone told TechTarget. “Bringing vector search and hybrid search on-premises and in private clouds means that companies that are subject to strict data residency and sovereignty rules, and cannot do this in the public cloud, can finally build accurate AI in one system.”
BARC US analyst Kevin Petrie similarly pointed to the value of MongoDB’s new features.
“This release is a great step forward for MongoDB,” he told TechTarget. “Enterprises of all sizes need to improve the accuracy of data capture, especially when it comes to text documents that provide critical context to agent AI.”
New York City-based MongoDB started as a NoSQL database vendor, but has expanded beyond its roots and now builds the Atlas data platform, which includes AI development capabilities. Competing vendors now include database specialists like Couchbase and Redis, as well as Databricks, Snowflake, and hyperscale cloud vendors.
take on trouble
Despite significant advances in AI model capabilities, many companies struggle to build agents and other AI applications that they can trust to run accurately in production.
This release is a great step forward for MongoDB. Companies of all sizes need to improve the accuracy of data capture, especially when it comes to text documents that provide critical context to agent AI.
Kevin PetrieBARC US Analyst
The problem that often hinders successful AI development is insufficient data capture.
Without proper context provided by a company’s own proprietary data and business logic, AI tools will not have the situational awareness to provide reliable output and development projects will fail. However, discovering and operationalizing the relevant information needed to successfully build AI tools can be difficult.
In response, this year various vendors introduced features aimed at connecting agents with the context they need. In June alone, AWS, Databricks, Microsoft, and Snowflake all made context for agents the focus of their product development initiatives.
MongoDB is similarly focused on improving data search for AI in 2026. In January, the vendor announced a set of five Voyage models that classify and make data discoverable, and in May it added new vector embedding capabilities that similarly aid in data discovery.
MongoDB is currently working to further improve data retrieval for AI based on customer feedback and MongoDB’s own awareness of issues that have stalled customers’ AI projects, said Benjamin Cefalo, the vendor’s chief product officer for core products.
“At most stages when a team takes an AI project to production, they will then hit one of two walls: either the retrieval is not accurate enough to be reliable, or the infrastructure does not meet the data residency requirements,” he said. “And in a literal sense, the accuracy problem is becoming more expensive.”
If a model tracks the wrong context, it increases development costs as the model loops back to try to find the correct context. Meanwhile, Cefalo continued, companies are trying to solve the problem by adding systems such as new search engines and vector databases.
“The conclusion was very simple,” he says. “If search capabilities are to make AI accurate and affordable, AI shouldn’t be pasted next to a database. It should reside within the database where the data already resides, whether it’s on-premises, in a private cloud, or in a hybrid environment.”
Specific new MongoDB features designed to improve data acquisition for AI include:
Native reranking of MongoDB Atlas. Built on Voyage AI embedding and ranking models to improve data retrieval accuracy.
Voyage Context 4 is a new embedding model for long documents designed to understand the full document context and use it to inform AI tools.
MongoDB’s hybrid search allows you to combine full-text and vector searches in a single query within a vendor’s production database.
Voyage Context 4 and Hybrid Search are generally available, but Native Reranking for MongoDB Atlas is in preview.
While some AI efforts have been stalled by insufficient data capture, others, particularly those by companies in highly regulated industries such as healthcare, finance, and energy, are bogged down by data sovereignty rules and regulatory requirements that make cloud-based AI development difficult.
To address the needs of these organizations, MongoDB is releasing Enterprise Advanced editions of Search and Vector Search to enable users to build AI tools behind their firewalls.
Additionally, search and vector search are now available in the free community edition of MongoDB, and the vendor has started supporting Apache Iceberg tables with Atlas Stream Processing.
Measuring
According to Leone, the re-ranking feature is probably the most important individual feature for MongoDB users. Meanwhile, from a competitive perspective, integrating features rather than forcing users to combine parts of the AI development workflow will help differentiate MongoDB from competitors, he continued.
“Individual competencies are becoming common across fields; [but] “What makes MongoDB unique is that it brings everything together in one place: the Voyage model, the operational database where the data for the app already resides, and the searches that are run against that live data,” he said.
But even though MongoDB has evolved to become a data and analytics platform that drives AI tools as well as competitors, without governance and analytics at scale, MongoDB is still not a fully-featured data management provider, Leone said.
“MongoDB is right there and definitely ahead of the curve when it comes to performing accurate searches directly on live application data rather than copies,” he said. “What’s even more challenging is that these platforms already hold a lot of customer data, giving companies a head start if they want to buy everything from one vendor.”
Petrie similarly said that as MongoDB expands beyond its roots, it’s finding ways to differentiate itself from similarly evolving vendors.
Specifically, we highlighted the ability for users to combine different data types to train models and connect agents to the right context.
“Now is MongoDB’s time to shine,” he said. “In order to deploy AI, we need context to ensure the adoption of AI. [models] The agent then produces reliable output. MongoDB allows you to organize and make use of multimodal data, especially text, with rich context. Integration of operational and analytical workloads also supports agent AI. ”
Looking to the future
As MongoDB evolves, its purpose is to move from a system of record that allows organizations to store data to an intelligence system for AI, Cefalo said.
Toward that end, vendors’ roadmaps include improving the memory and acquisition layers of agents so they can act on data they can trust, providing capabilities that allow organizations to run workloads in their preferred environment, and integrating previously disparate tools to simplify development.
“Every agent workload has three tiers: harness, model, and data, and the data tier is the one we’re going to own,” Cefalo said.
Meanwhile, after working on data retrieval, Leone suggested adding functionality to MongoDB that would allow customers to test the accuracy of their AI application’s retrieval.
“They’ve made search more accurate, and the natural next step is to allow customers to see and prove how accurate it is in their own environment. That’s exactly what regulated buyers have to show auditors,” he said. “Having a first-party assessment layer for acquisition completes the compliance story and gives those teams one less thing to source elsewhere.”
Eric Avidon is a senior news writer at Informa TechTarget and a journalist with more than 30 years of experience. He is responsible for analysis and data management.
