During its developer conference, Box demonstrated a new capability that builds Nvidia's Nemotron embedding system directly into the core of its product. Termed Box Content Retrieval, this upcoming feature enables AI agents to address user queries with significantly higher effectiveness by leveraging the embedded metadata generated by Nemotron.
Faster Speeds and Lower Query Costs for Enterprises
According to Box, the primary benefits for users include faster, more precise, and less expensive query processing. Kari Briski, Nvidia's vice president of generative AI software, stated that she anticipates queries being resolved four times faster, accompanied by a 20 percent reduction in errors and a 90 percent drop in token usage. The system is scheduled for rollout early next year, becoming accessible to all of Box's Enterprise Advance customers.
Streamlining Document Review for AI Agents
Box CEO Aaron Levie emphasized that the upgrade dramatically boosts the relevance of the platform's search technology for agents. Instead of requiring an agent to read through entire documents sequentially, the system can extract and transmit specific document chunks that directly align with the user's query or the ongoing task managed by the agent.
Rooted in Nvidia's Open-Weight Embedding Initiative
This functionality is a direct implementation of Nvidia's Nemotron Embed project, which supplies open-weight embedding models designed to optimize automated retrieval systems. While major research labs have historically overlooked such models, which utilize large language models to pre-process data for agent consumption, Nvidia has made this sector a key focus. Its recent Nemotron release in July secured the top spot on the RTEB benchmark, which evaluates retrieval performance.
Enhanced Contextual Search for Legal and Complex Fields
Box notes that the system shines brightest when paired with active agents, citing legal professionals as a prime example. Lawyers will theoretically be able to unearth relevant clauses and legal precedents much more efficiently under the new setup. Because Nemotron embeddings map underlying concepts rather than relying strictly on basic keywords, the search utility can surface related legal themes even when different phrasing or vocabulary is employed in the documents.
Combining Proprietary Controls with Open Models
Although the underlying Nemotron models are freely accessible, the overarching system heavily integrates Box's pre-existing metadata and granular permission controls, making it difficult for competitors to replicate from scratch. Levie noted that the platform effectively packs around five distinct technologies into a single straightforward API.
