Box Puts Nvidia's Nemotron Retrieval Inside Its Content Platform

Box has unveiled Box Content Retrieval, a feature that builds Nvidia's Nemotron embedding system into the core Box product. The company announced it at its developer conference. The system uses embedded metadata created by Nemotron so that AI agents can answer queries against stored content. TechCrunch

The feature is described as an implementation of Nvidia's Nemotron Embed project, which offers open-weight embedding models (models whose trained weights are published for anyone to download and run) intended to improve automated retrieval systems. Box is therefore not licensing a closed API for this layer. It is adopting open-weight models as the retrieval substrate inside its own platform, which gives the company a defined model lineage to version, evaluate and, if it chooses, host differently over time.

Kari Briski, Nvidia's vice president of generative AI software, said she expects queries run through the system to be answered four times faster, with 20% fewer errors and 90% fewer tokens used. She offered these as expectations for the system's performance. Latency, accuracy and token consumption each matter for a different reason. Fewer tokens (the chunks of text a model reads and writes, typically billed by volume) lower per-query inference cost. Fewer errors cut the number of retrieval passes an agent needs before it reaches a usable answer. Faster responses matter most in multi-step agent loops, where each retrieval call sits on the critical path, the step every later step waits for.

Box Content Retrieval will be available to all Enterprise Advance customers, with rollout beginning early next year.

The Nemotron family covers more than embeddings. Nvidia describes its retrieval models as capable of parsing multimodal documents (files that mix text, images and tables), generating high-quality embeddings and reranking results. Nemotron Parse 1.1 uses bounding boxes to retain document layout and to classify content types such as headers, paragraphs and captions. Nvidia Developer Blog A separate model, nemotron-page-elements-v3, is an object detection model designed to locate key page elements, including tables and charts. For repositories full of scanned PDFs, slide decks and spreadsheets exported as images, that layout-aware parsing determines whether a chunk of text keeps its structural meaning when it is embedded.

On the benchmark side, Nvidia's Nemotron release in July topped the RTEB benchmark, which measures retrieval quality. Leaderboard position does not guarantee production performance on a customer's corpus, but it gives the embedding layer a public, reproducible reference point, which enterprise buyers increasingly ask for.

Nvidia's roadmap extends beyond the current models. In June, Nvidia described Nemotron 3 Ultra as a smaller, faster open model built for long-running agents, and said it delivers 5x faster inference. NVIDIA Investor Relations On August 11, Reuters reported, citing The Information, that Nvidia is developing Nemotron 4 open-source models. According to that report, Nemotron 4 is a 1-trillion-parameter model intended to rival open AI models. Reuters The report says the model is in development and has not been released.

The broader context here is that an agent's answer is only as good as the passage it retrieves. Model reasoning gets most of the attention, but for document-heavy workflows the retrieval stage decides which contract clause, engineering spec or financial table reaches the model at all. Putting embeddings and document metadata inside the content platform, rather than in a separate pipeline the customer must stitch together, changes who owns that stage and how it is governed.

In my view, the figures Briski cited should be tested against real corpora before they become planning inputs. Enterprise content is messier than benchmark datasets, with permissions, versioning and inconsistent scans. The question for IT teams is less whether retrieval improves in the abstract and more whether it holds on their own repositories, under their own access controls.

In my view, if the efficiency claims hold at scale, agents can work across larger document estates without the token budget growing at the same rate. That would make routine questions over internal archives cheaper to run and faster to answer. Rollout to Enterprise Advance customers in early 2027 will be the first point at which that can be measured in production.