Box unveils an agent-first retrieval engine built on Nvidia's Nemotron embedding models

Box used its BoxWorks 2026 conference in San Francisco on November 5 to present a new retrieval engine designed for AI agents rather than human searchers. The company calls it Box Context Retrieval (some press coverage uses the name Content Retrieval) and says it is powered by Nvidia's Nemotron Retriever, a family of open embedding models. The announcement is part of a broader expansion of the partnership between the two companies.

How the engine works

The idea is that an agent should not have to read every file it touches. Instead, the engine indexes enterprise content and returns the passages most relevant to a request. Box says it combines exact keyword matching with semantic search, so it can surface related material even when a query and a document use different terms. It indexes scanned documents, images, audio and video as well as text, and Box says new or updated files become available to agents with low latency. Results are permission-aware: an agent working for a user only draws on content that user is allowed to see. Box CEO Aaron Levie described the product to reporters as several technologies bundled behind a single interface.

Claimed gains and availability

Box lists potential outcomes of a 20% improvement in search accuracy, measured by Recall@5, a 90% reduction in token usage and question answering up to four times faster. Nvidia's Kari Briski cited similar figures in press coverage. Box does not say what these numbers are compared against, and they are vendor projections rather than results from customer deployments.

Context Retrieval is not available yet. Box says it will become generally available to Enterprise Advanced customers early next year, and its own release warns that unreleased features may arrive late or not at all. Separately, Box is adding Nvidia's open-weight Nemotron models to Box AI, starting with Nemotron 3 Ultra, which Enterprise Advanced customers can already choose. The companies also plan to pair Box's governance controls with Nvidia OpenShell, a secure runtime for running agents.

Why it matters

For enterprise software vendors, the race is shifting from chat assistants to agents that act on company data, and retrieval quality and token costs are becoming selling points. By building on open embedding models and its existing permission system, Box is betting that it can serve as the content layer other agents plug into. How well the engine performs on real corporate archives will only be clear once customers can test it next year.