Box, a California-based content management firm, is teaming up with NVIDIA to give artificial intelligence (AI) agents access to the knowledge held in company files.
The companies will build a retrieval engine on NVIDIA’s Nemotron open models, bring the chipmaker’s open-weight language models into Box AI and extend security controls to agents as they run tasks. Box unveiled the plans on November 5 at BoxWorks 2026, its annual conference in San Francisco.
“Enterprise AI reaches its full potential when models and agents can securely access the unique context of a business,” said Aaron Levie, chief executive and co-founder of Box. “Open weight AI is an important part of that vision, complementing closed models while driving innovation and broader adoption across the industry.”
“AI agents need more than intelligence. They need the knowledge and context of the business to do useful work,” said Jensen Huang, founder and chief executive of NVIDIA.
Founded in 2005, Box is headquartered in Redwood City and counts JLL, Morgan Stanley and Nationwide among its customers.
Searching beyond plain text
The new engine, Box Context Retrieval, uses NVIDIA Nemotron Retriever embedding models to turn stored content into a structured knowledge layer that agents can reason over. It is designed to index material that AI has long struggled to interpret, including scanned documents, images, audio and video.
Retrieval is permission-aware, so agents draw only on files that a user is allowed to access.
Box said customers could improve search accuracy by 20%, based on whether the right result appears among the top five (Recall@5), cut token usage by 90% and answer questions four times faster.
Box Context Retrieval also pairs keyword matching with semantic search, which reads for meaning and context, so users can find material without knowing the exact search terms.
Possible uses include:
Legal professionals finding clauses and precedents across contracts and matter files, even when different terminology is used;
Insurance auditors analyzing claim packages that mix scanned forms, property images, recorded statements and video evidence;
Clinical researchers working with newly added or updated policies, care protocols and research in real time.
The company is also adding NVIDIA’s open-weight Nemotron models, starting with Nemotron 3 Ultra, to Box AI, which lets customers choose models by performance, cost and control. It said the open models offer a more cost-efficient option for specialized, high-volume workloads.
Box and NVIDIA, both members of the Open Secure AI Alliance, are extending the partnership to agent safety.
They are combining Box Mount, which carries Box’s content security and governance, with NVIDIA OpenShell, a secure runtime that enforces policies and supervises how agents access content, use tools and execute tasks.
Nemotron 3 Ultra is available now to Enterprise Advanced customers, with Box Context Retrieval to follow early next year.