Box Builds Nvidia's Nemotron Embeddings Into New Retrieval System for AI Agents

The system is designed for AI agents rather than human searchers. Instead of an agent reading each document in full to answer a question, Box will return only the passages judged most relevant to the query or task, drawing on metadata generated by the Nemotron models. The feature will be offered to all Enterprise Advance customers, with a rollout starting early next year.

The pitch to enterprise buyers is cost and accuracy. Kari Briski, Nvidia's vice president of generative AI software, told TechCrunch she expects queries to run four times faster, with 20% fewer errors and 90% fewer tokens consumed. Token consumption is a direct driver of what companies pay to run agents on large language models, so a cut of that size, if it holds in production, would change the economics of agent workloads that sit on top of large document stores.

Box chief executive Aaron Levie said the change "dramatically boosts both the relevance of our search technology for agents". He described the approach as sending back "chunks from the documents that are most relevant to the user's query" rather than having an agent work through each file.

Embedding models convert text into numerical representations that capture meaning, which lets a search system match concepts rather than exact keywords. Frontier AI labs have paid relatively little attention to them, but they have been a focus of Nvidia's open-model programme. According to TechCrunch, the Nemotron embedding release Nvidia put out in July ranked first on the RTEB benchmark, which measures retrieval quality.

Box gave legal work as an example of where the system should help. A lawyer looking for relevant clauses or precedents could surface related material even when it is written in different language from the query, because the embeddings map underlying concepts rather than matching words.

The Nemotron models themselves are freely available, but Box argues the product would be hard to copy because it also depends on the company's existing metadata and permission controls, which determine what each user and agent is allowed to see. "We're probably packing in five different technologies into one simple API," Levie told TechCrunch.

The launch extends a series of tie-ups between the two companies. Box has previously added support for Nvidia's Llama Nemotron reasoning models in its Box AI Studio and has worked with Nvidia on security controls for autonomous agents that access Box content. Box has not said whether Content Retrieval will be priced separately or extended beyond the Enterprise Advance tier when it ships early next year.