At its annual developer conference held on Thursday, Box unveiled a significant new capability built around Nvidia's Nemotron embedding technology, integrating it directly into the foundation of its platform. The announcement signals a deeper push by the cloud content management company into AI-driven enterprise workflows.

The new feature, called Box Content Retrieval, is designed to help AI agents respond to queries with greater precision and efficiency. It does this by drawing on structured metadata generated through Nvidia's Nemotron embedding system, allowing agents to locate and surface the most relevant content without having to process entire documents from scratch.

According to Box, the practical benefits for enterprise users include faster query responses, improved accuracy, and meaningfully lower computational costs. Kari Briski, Nvidia's Vice President of Generative AI Software, indicated that the performance gains are substantial - queries are expected to return results four times faster, with error rates dropping by around 20% and token usage reduced by as much as 90%.

The feature is slated to roll out early next year and will be accessible to all Box Enterprise Advance subscribers.

How the System Works

Box CEO Aaron Levie explained the core advantage of the new approach in straightforward terms. Rather than requiring an AI agent to read through each document in full to find relevant information, the system can return the most pertinent sections or chunks of content directly in response to a given query or task. This fundamentally changes the efficiency profile of agent-based workflows operating on large document repositories.

The underlying technology powering this feature is Nvidia's Nemotron Embed project, which provides open-weight embedding models specifically designed to enhance automated retrieval systems. These models work by using large language models to pre-process and index data in a way that makes it far more accessible and usable for AI agents operating downstream.

Embedding models of this type have historically received less attention from major AI research laboratories, which have tended to concentrate resources on generative and reasoning capabilities. Nvidia, however, has made this category a deliberate focus within its open-source model development efforts. A Nemotron embedding model released in July topped the BEIR and MTEB retrieval benchmarks, which are widely used to evaluate the quality and accuracy of information retrieval systems.

Real-World Applications

Box highlighted the legal industry as a prime example of where the technology is expected to have a meaningful impact. Lawyers working with large volumes of contracts, case files, and legal precedents could use the system to locate specific clauses or thematically related content far more quickly than traditional keyword-based search would allow.

A key advantage of Nemotron's embedding approach is that it goes beyond surface-level keyword matching. Instead, the models map the conceptual meaning behind text, which means a search can surface relevant content even when different terminology or phrasing is used across documents. In a legal context, this could mean identifying related clauses or analogous case law that would not appear in a standard keyword search because the exact wording differs.

This semantic understanding makes the system considerably more robust for professional environments where precise language varies widely across documents, jurisdictions, or time periods.

More Than Just Open Models

Although the Nemotron embedding models themselves are freely available for developers to use, Box emphasizes that the full Box Content Retrieval system is not easily replicated outside of its platform. Much of its value comes from deep integration with Box's existing infrastructure - specifically its metadata management capabilities and its granular permission controls, which govern who can access what content within an organization.

Levie noted that the final product bundles together several distinct technologies into a single, simplified API interface. This layered approach means that while the AI models at the core are open-source, the surrounding architecture that makes them enterprise-ready within a secure, permissioned content environment represents a significant proprietary advantage for Box.

The partnership reflects a broader trend of enterprise software companies embedding specialized AI components from hardware and infrastructure providers directly into their core products, rather than building such capabilities from the ground up. For Box, aligning with Nvidia's model development roadmap provides access to cutting-edge retrieval technology while allowing the company to focus on the enterprise integration layer where its existing strengths lie.

Box Content Retrieval is expected to become available to Enterprise Advance customers in the first part of next year, with the company positioning it as a foundational component for businesses looking to deploy AI agents across large internal content libraries.