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

Box announced Box Content Retrieval at its developer conference, embedding Nvidia’s Nemotron system into the core of its product so agents can answer queries from pre-processed metadata rather than grinding through whole files.

Nvidia’s Kari Briski told TechCrunch the company expects queries answered four times faster, with 20% fewer errors and 90% fewer tokens. The feature rolls out early next year to Box Enterprise Advance customers.

That is not a demo toy. It is enterprise content infrastructure betting that retrieval quality — not just the chat model on top — will decide whether agents are useful inside companies that already live in Box.

We’re watching the agent stack split into pieces that actually matter. Generation gets the headlines. Retrieval decides whether the agent is guessing or grounded. Box’s move is a bet that open embedding models plus enterprise metadata and permissions beat “just let the agent read everything.”

Aaron Levie framed it cleanly: instead of an agent reading each document, Box sends back the chunks most relevant to the query or task. That sounds obvious until you remember most enterprise agent demos still treat the file store like a flat folder and burn tokens proving they can open PDFs.

Nemotron Embed is open-weight, and Nvidia’s July release topped the RTEB retrieval benchmark. Freely available models do not make this trivial to recreate. Levie said Box is packing roughly five technologies into one API — embeddings, existing metadata, and permission controls chief among them. The moat is not the weights. It’s the governed context.

The legal-workflow example is the right stress test. Concept-level retrieval that surfaces related clauses even when wording differs is where keyword search always failed lawyers. If Box ships that at the claimed latency and token savings, agents stop being a cost center for search and start looking like a productivity layer.

We’re for useful agents that respect permissions. The risk if this scales poorly is false confidence: a tidy chunk that looks relevant and is subtly wrong, delivered four times faster. Retrieval quality metrics and audit trails matter as much as the speed claim.

Major labs have often treated embedding and retrieval models as secondary to generative releases. Nvidia’s open-source Nemotron project has made retrieval a deliberate focus, and Box is one of the first large content platforms to put those embeddings at the center of an agent-facing product surface rather than a side experiment.