At BoxWorks 2026 on November 5, Box announced an expanded Nvidia partnership built around Box Context Retrieval: a Nemotron-powered search engine intended to give AI agents useful context from enterprise files without bypassing users’ existing permissions. Box also announced an AI workflow-building tool called Box Architect.
8
17
How Box Context Retrieval is designed to work
The engine uses Nvidia Nemotron Retriever embedding models. Box describes it as combining keyword matching with semantic search to find relevant passages for an agent’s task, while honoring the access permissions already attached to Box content.
23
25
36
Box and coverage of the announcement also describe retrieval across formats beyond ordinary text, including scanned documents, images, audio and video.
7
23 The stated goal is to give an AI agent relevant material to work with, rather than make it search through enterprise content without regard to relevance or access rights.
Some early coverage called the feature “Box Content Retrieval.” Box’s announcement uses the name Box Context Retrieval.
4
8
Box’s reported performance figures
Box said the engine could improve search accuracy by 20%, reduce token use by 90% and make question-answering four times faster.
17 These are company-reported figures. The available reporting does not provide enough benchmark detail to assess the comparison baseline or methodology, so the numbers should be read as Box’s claims rather than independently established results.
Availability and other announcements
Box said Context Retrieval was planned for general availability to Enterprise Advanced customers in early 2027; the announcement did not describe it as generally available at launch.
2
17
The wider Nvidia collaboration also includes bringing Nvidia’s open-weight Nemotron models into Box AI and extending security and governance to agent execution through Nvidia OpenShell.
8
26 Nvidia describes OpenShell as a runtime for autonomous AI agents.
18
Box separately unveiled Box Architect, a tool for building AI-driven workflows with content stored in Box.
5
17 Together, the announcements point to two connected enterprise AI needs: agents need relevant, permission-aware information, and teams need ways to put agents to work in content-based processes.