Box has deepened its partnership with NVIDIA to give AI agents reliable access to the knowledge held in company files. The expanded collaboration was unveiled on 5 November at BoxWorks 2026, the company's annual conference in San Francisco.

The centrepiece is Box Context Retrieval, a new engine powered by NVIDIA Nemotron Retriever embedding models. It turns stored content into a structured knowledge layer agents can search and reason over, including scanned documents, images, audio and video. It combines keyword matching with semantic search and is permission-aware, so agents draw only on files the user is allowed to see.

What Box claims, and when it ships

Box says the engine can improve search accuracy by 20 percent, measured at Recall@5, while cutting token consumption by up to 90 percent and delivering answers about four times faster. Those are company figures from the announcement, not independently verified benchmarks, according to the Box press release.

The partnership also brings NVIDIA's frontier open-weight models into Box AI, starting with Nemotron 3 Ultra, which is now available to Enterprise Advanced customers. Box Context Retrieval itself is scheduled for general availability to the same customer tier early next year. The two companies are also extending security and governance to agent execution through NVIDIA OpenShell, and Box made its agent security controls generally available at the event.

The announcements land as companies move from chatbots to agents that act on their behalf. Those agents are only as useful as the context they can reach, and most of it sits in unstructured files. Box Architect, a builder that designs agentic workflows from a plain-language description, enters beta for Enterprise Advanced customers in the fourth quarter.