Box uses Nvidia's Nemotron for agent content search

Box announced at BoxWorks 2026 a partnership with Nvidia that puts Nemotron at the center of its content management platform. The result is Box Context Retrieval, an engine that transforms corporate files into a structured knowledge layer, ready for AI agents to query with precision.

The proposal is simple to understand. AI agents need context to be useful, and that context is scattered across contracts, policies, customer history and internal documents. Today, many agents need to read document by document to answer a question. Box's new engine uses Nemotron's embedding models to index content semantically, allowing the agent to find relevant passages without needing to process everything.

In this article you will see:

• What Box Context Retrieval is and how Nemotron works inside it
• The numbers Box and Nvidia promise for search and token consumption
• Who benefits from the news and when it arrives
• How this fits into Box's strategy of offering model choice

What is Box Context Retrieval

Box Context Retrieval is a semantic search engine powered by Nemotron Retriever, a collection of open models from Nvidia that converts business data into a knowledge layer that agents can reason over. Instead of relying on exact keywords, the engine maps concepts and relationships between documents.

The big difference is the scope. The engine can index not only text, but also scanned documents, images, audio and video. This means an agent can find information in digitized contracts, meeting recordings or property images without anyone needing to transcribe or describe them manually.

Retrieval is also permission aware. The agent only accesses content the user is permitted to see. This is essential in corporate environments, where sensitive information needs to be protected even when AI is involved.

The numbers Box and Nvidia promise

According to Box, customers who adopt the engine can expect measurable improvements. The company projects a 20 percent increase in search accuracy, measured by Recall at 5, a 90 percent reduction in token usage and a four times acceleration in response speed to questions.

The token reduction is particularly relevant for those using agents in production. The fewer tokens an agent needs to process to find an answer, the lower the operational cost and the faster the response time. This makes it viable to scale agents for tasks that were previously economically unfeasible.

It is worth remembering that these numbers are company projections, not independent benchmarks. Still, they indicate the direction Box wants to go: making context retrieval cheaper and more accurate for agents.

Who benefits and when it arrives

Box listed concrete use cases for different professional profiles. Financial analysts can query loan applications with answers based only on authorized content. Legal professionals can find clauses and precedents in contracts, even when terminology varies. Clinical researchers access updated policies and protocols in real time. Insurance auditors analyze claims packages that include scanned forms, images and videos. University administrators query policies and research materials.

Box Context Retrieval will be generally available for Enterprise Advanced customers at the start of next year. Nvidia's Nemotron models, starting with 3 Ultra, are available now in Box AI for Enterprise Advanced customers.

How this fits Box's strategy

Box adopts a multi-model approach. The platform gives customers the flexibility to choose the right model for each task, considering performance, cost and control. With the addition of the open Nemotron models, customers gain access to frontier-scale intelligence with greater cost efficiency for specialized and high-volume workloads.

The partnership also includes security in agent execution. Box and Nvidia are part of the Open Secure AI Alliance and are collaborating so agents can access content with security and governance controls. NVIDIA OpenShell provides a secure execution environment, with policy enforcement and oversight.

For those already following Nvidia's tools for agents, it's worth comparing with the Nvidia Agent Toolkit for corporate data context, which takes a complementary approach to the same problem.

For those building AI tools, Box's move illustrates a trend: the value of agents is not only in the model, but in the ability to retrieve relevant context quickly and securely. Companies that already invest in content management can gain an advantage by connecting their agents to this knowledge layer.

See also how the partnership between Mistral and Nvidia expands AI engineering tools in Mistral and Nvidia expand AI engineering partnership.