Box builds Nvidia Nemotron embeddings into new retrieval feature for AI agents

What happened

Box unveiled Box Content Retrieval at its developer conference on Thursday, a feature that places Nvidia's Nemotron embedding system at the center of its platform, TechCrunch reported. The tool lets AI agents draw on metadata generated by Nemotron when responding to queries, and Box said users should see quicker, more accurate and cheaper results.

Kari Briski, Nvidia's vice president of generative AI software, told TechCrunch she anticipates queries being answered four times as fast, with errors down 20% and token usage cut by 90%. Box CEO Aaron Levie said the approach means agents no longer need to work through entire documents. Instead, the system returns the specific passages most relevant to a user's request or to the agent's task.

The feature will reach all Box Enterprise Advance customers, with the rollout beginning early next year. Box said it works best alongside active agents and cited lawyers searching for relevant clauses and precedents as an example. Because the embeddings capture concepts rather than just keywords, related themes can surface even when the phrasing or language differs.

Context

The feature builds on Nvidia's Nemotron Embed project, which publishes open-weight embedding models designed to improve automated retrieval. Such models use large language models to pre-process data so agents can work with it more effectively, an area major AI labs have often overlooked. Retrieval has been a focus of Nvidia's open-source model work, and its Nemotron release in July ranked first on the RTEB retrieval quality benchmark.