Box builds Nvidia Nemotron embeddings into agent-first Content Retrieval
At its developer conference Box said Enterprise Advance customers will get the feature early next year; Nvidia's Kari Briski expects ~4× faster answers, 20% fewer errors and 90% fewer tokens.
Why it matters
Embedding retrieval into the content layer is how enterprise agents avoid reading every file on each turn; Box is wiring Nvidia's open Nemotron embedders into its metadata and permissions rather than exposing a raw vector store. Claimed gains of roughly 4× speed, 20% fewer errors and 90% fewer tokens, if they hold in customer traffic, would change the cost of agentic search over large corpora. Availability is limited to Enterprise Advance early next year, so the proof is in production latency and citation quality.
Key facts
- Box Content Retrieval announced at Box developer conference
- Uses Nvidia Nemotron embedding models
- Aimed at agent queries over Box content
- Nvidia VP Kari Briski cites ~4× faster answers
- Also cites 20% fewer errors and 90% fewer tokens
- Rolling out to Enterprise Advance customers early next year