Box says a Nemotron retriever can cut the tokens an agent spends by 90 percent
Box Context Retrieval, announced Thursday at BoxWorks, is promised to Enterprise Advanced customers early next year. The release states a 20 percent gain on Recall@5 and answers four times as fast. TechCrunch calls the product Content Retrieval and quotes a 20 percent drop in errors.
Feature2 minIn-depth
Priya Shah2 min read
- Box
- Nvidia
- Nemotron
- Agents

Photo: BoxSource page
Box said on Thursday that it is introducing Box Context Retrieval, an engine that uses Nvidia’s Nemotron Retriever embedding models so an agent can pull passages from company files it is allowed to see. The announcement, posted on Box’s blog and datelined Redwood City, was made for BoxWorks 2026 in San Francisco. Box says customers will have the potential to improve search accuracy by 20 percent on a measure it names as Recall@5, to reduce token usage by 90 percent, and to answer questions four times as fast. The engine is scheduled for general availability to Enterprise Advanced customers early next year. Nvidia’s Nemotron models, starting with Nemotron 3 Ultra, are available now in Box AI for those same customers.
The files Box lists as newly practical for this index are the ones that have been awkward for a text search: scanned documents, images, audio, and video. The release also describes hybrid search, keyword match plus a search on meaning, and permission-aware retrieval, so an agent is supposed to see only what the person asking already can. Indexing, Box says, should make new and updated files available to agents in real time. The examples in the release are loans, contracts, clinical protocols, insurance claims, and university files. They are illustrations of the pitch, not results from named customers.
AI agents need more than intelligence. They need the knowledge and context of the business to do useful work.
— Jensen Huang, founder and chief executive of Nvidia, in Box’s 5 November release
A second name, and a second version of the 20 percent
Russell Brandom at TechCrunch, in a story timestamped 6:00 a.m. Pacific the same day, does not use Box’s product name. He calls the feature Box Content Retrieval. He writes that Kari Briski, Nvidia’s vice president of generative AI software, told him she expects queries to be answered four times faster, with 20 percent fewer errors and 90 percent fewer tokens. The speed and the token figure match the release. The error figure does not, on the page: Box states a 20 percent improvement in Recall@5, and TechCrunch’s paraphrase of Briski is fewer errors. Neither story says whether those are the same internal test. Brandom also writes that the system will reach all of Box’s Enterprise Advance customers early next year. Box’s release says Enterprise Advanced.
Aaron Levie, Box’s chief executive, told TechCrunch the change is that an agent no longer has to read each document. Box can send back the chunks relevant to the query or the task. In the company’s own release he says the partnership is meant to give customers a faster, more precise way to ground AI in their content, and that open-weight models sit alongside closed ones. A separate part of the announcement is about where the agent runs. Box and Nvidia say they are extending governance through Nvidia OpenShell, which they describe as a secure runtime with policy enforcement and supervision, paired with Box Mount for content controls. Both companies are members of the Open Secure AI Alliance, the release says. Brandom writes that a Nemotron release in July topped the RTEB retrieval benchmark. That comparison is in his story. It is not in the Box post.


