Box taps Nvidia’s Nemotron for new agent-first content retrieval system

At its developer conference on Thursday, Box announced a new feature that will build Nvidia’s Nemotron embedding system into the core of its product. Called Box Content Retrieval, the new feature allows agents to answer queries far more effectively, relying on the embedded metadata created by Nemotron. 

The upshot for users is faster, more accurate and less expensive queries. Nvidia’s VP of Generative AI software Kari Briski told TechCrunch that she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens used.

The system will be available to all Enterprise Advance customers, rolling out early next year.

“This dramatically boosts both the relevance of our search technology for agents,” Box CEO Aaron Levie said. “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 the task is that the agent is doing.”

The feature is an implementation of Nvidia’s Nemotron Embed project, which offers open-weight embedding models to improve automated retrieval systems. Major labs have often overlooked embedding models, which use LLMs to pre-process data to work better with agents. While largely overlooked by frontier labs, the sector has been a focus for Nvidia’s open-source model project: Nvidia’s most recent Nemotron release in July ranked #1 on the RTEB benchmark, which measures retrieval quality.

The system will be most useful when paired with more active agents, which can operate with more fluency over the Nemotron-processed data. Box gave the example of lawyers, who can find relevant clauses and precedents more easily under the new system. Because Nemotron’s embeddings go beyond simple keywords to map underlying concteps, a search tool can surface related legal concepts even when different language is used.

While the Nemotron models are freely available, the system also relies heavily on Box’s pre-existing metadata and permission controls, making it challenging to recreate from scratch. “we’re probably packing in five different technologies into one simple API,” Levie told TechCrunch.

View Bio