Box partners with Nvidia to build faster, cheaper AI agents into its platform

Box is embedding Nvidia's Nemotron models directly into its content platform, a move that will change how enterprise search and AI agents retrieve information across millions of documents. The new feature, Box Content Retrieval, was announced Thursday at Box's developer conference and rolls out to Enterprise Advance customers in early 2024.

The integration uses Nemotron's embedding system to pre-process and map the conceptual meaning of stored files, rather than relying on simple keyword matching. Box CEO Aaron Levie said the system packs "probably five different technologies into one simple API," combining Nvidia's open-weight models with Box's existing metadata and permission controls.

Faster answers with fewer resources

Kari Briski, Nvidia's VP of generative AI software, told TechCrunch that queries should run four times faster, with 20% fewer errors and 90% fewer tokens consumed. Those efficiency gains matter for teams running frequent, complex searches across large document libraries.

The system works by sending back relevant chunks from documents rather than forcing an agent to read through entire files. "Instead of an agent having to kind of read through each document, we can send back chunks from the documents that are most relevant to the user's query or whatever the task is that the agent is doing," Levie said.

Beyond keywords to conceptual search

Nemotron's embeddings go past surface-level keyword detection to map underlying concepts. A legal team searching for liability clauses could surface related themes even when different terminology appears across contracts. Box cited lawyers as a primary use case, where finding relevant clauses and precedents quickly translates to billable hours saved.

Nvidia has made retrieval systems a focus of its open-source model program. Its Nemotron release in July topped the RTEB benchmark, which measures retrieval quality. Major AI labs have often overlooked embedding models despite their role in making agents more effective.

While the Nemotron models are freely available, recreating Box's implementation from scratch would be difficult. The feature depends on the company's existing metadata structure and permission controls built over years of enterprise document management.

Why this matters for creatives, marketers, and writers

For anyone who spends hours hunting through shared drives, project folders, or content libraries, a search tool that understands concepts rather than just matching words changes the speed of research. A marketer looking for past campaign assets could find relevant work even if the original files used different terminology. Writers and creatives pulling reference material from company archives will spend less time guessing which keywords colleagues used when naming files. The reduction in token consumption also means lower costs for teams running frequent AI-assisted searches - a practical concern as more creative workflows incorporate agent-based tools.