Box plugs NVIDIA Nemotron into enterprise search for AI agents

Most enterprise AI agents waste money reading files they should never have opened. Box announced Box Context Retrieval on Wednesday morning at BoxWorks 2026 in San Francisco. It is an engine built on NVIDIA Nemotron embedding models. The engine gives an agent the exact chunks of a company's files it needs for a given question.

The engine sits between an agent and the content it reasons over. Box CEO Aaron Levie said the point is to stop an agent "having to kind of read through each document". Instead, the engine sends back only the chunks that match a task. The models come from NVIDIA Nemotron Retriever, a family of open weight embedding models.

What Box says the engine does

Box says customers can improve search accuracy by 20 percent at Recall@5. Token usage drops 90 percent. Answers come back four times faster. Those are Box's own numbers for its own engine. They describe the retrieval step, not the full agent run.

NVIDIA's VP of generative AI software, Kari Briski, gave TechCrunch the same figures framed a second way. She said queries will be answered four times faster. She said errors will drop 20 percent. She said token use will drop 90 percent.

The engine indexes more than text. Box says it will make scanned documents, images, audio and video searchable with multimodal indexing. It will combine keyword search with meaning-based search in one hybrid query. Every result respects the file permissions a user already has in Box. The agent only sees chunks from files the user is allowed to open.

Who gets it and when

Box Context Retrieval will be generally available to Enterprise Advanced customers early next year. NVIDIA's Nemotron 3 Ultra model is available to Enterprise Advanced customers now. It is the first of several Nemotron open weight models coming to Box AI.

Box and NVIDIA also said they are extending the partnership to agent execution. The vehicle is NVIDIA OpenShell, a secure runtime. OpenShell enforces policy on what an agent can call and what content it can touch. Both companies named the Open Secure AI Alliance as the frame for the work.

Box gave five example workflows in its post. A financial analyst asking about loan applications sees only content they have access to. A lawyer finds clauses and precedents across contracts even when the wording differs. A clinical researcher gets newly added care protocols in real time. An insurance auditor inspects claim packages that include scanned forms, property images, recorded statements and video. A university administrator asks questions across policies, research materials and student services documents.

The retrieval models are open weight. Another team could in theory wire up the same embedding models themselves. Box's position is that the engine's value comes from pairing Nemotron with Box's own permission controls and metadata. Levie described the result as "five different technologies into one simple API". A team that already has its content in Box gets the retrieval layer without re-plumbing permissions.

Source

Box: Box Partners With NVIDIA to Deliver the Trusted Content Layer for Enterprise AI. TechCrunch: Box taps Nvidia's Nemotron for new agent-first content retrieval system.