Box is putting Nvidia’s Nemotron embedding models at the heart of a new feature called Box Context Retrieval, which is meant to help AI agents find the right passages inside company files. The cloud content company showed it off at its developer conference on Thursday, November 5. Instead of making an agent read every document, Box Context Retrieval sends back only the chunks that matter for the task.
The pitch is simple: faster answers, fewer mistakes, and a smaller AI bill. It is aimed at businesses that already store contracts, policies, and research in Box and now want agents to work with that content safely.
What Box Context Retrieval Does
Box says the feature automatically builds and maintains a secure index over the content a company keeps in Box. Agents and apps can then ask for focused context rather than pulling whole files into a model. Box describes it as one less custom pipeline or external vector database for teams to build and govern.
Box CEO Aaron Levie explained the idea in plain terms. Rather than having an agent read through each document, Box can return the pieces of documents that best match the user’s question or the job the agent is doing. Levie also said the company is packing about five different technologies behind one simple API.
Hybrid search with Nemotron embeddings
Under the hood, the system pairs old-fashioned keyword matching with semantic search powered by Nemotron embedding models. That means you can search for an exact project code or just ask a question in your own words. Because embeddings map meaning instead of exact terms, the search can surface related ideas even when the wording or language is different.
Permissions checked on every query
Box says every request checks the person’s or agent’s live Box permissions, so nobody sees files they are not allowed to open. Access changes in Box carry over to retrieval automatically, without syncing separate access lists. That reliance on Box’s existing metadata and permission controls is a big part of why the setup would be hard to copy from scratch, even though the Nemotron models themselves are freely available.
How Nvidia’s Nemotron Fits In
Nemotron Embed is Nvidia’s family of open-weight embedding models built to improve automated retrieval. Embedding models use AI to pre-process data so that AI agents can use it more easily, an area that big labs have often paid less attention to. Nvidia’s Nemotron release from July topped the RTEB benchmark, which measures retrieval quality.
The partnership comes as Nvidia keeps growing its open Nemotron family, including the Nemotron 3.5 Super model we covered earlier. Nvidia’s VP of generative AI software, Kari Briski, gave some expected gains for Box customers.
| Metric | Nvidia’s expected result |
|---|---|
| Query speed | About 4 times faster |
| Errors | About 20% fewer |
| Tokens used | About 90% fewer |
| Availability | Enterprise Advanced customers, rolling out early next year |
Keep in mind these are company projections, not independent test results. Real gains will depend on how much content a business has and how its agents are built.
Where Box Context Retrieval Could Help Most
Box frames the feature around industries that live inside documents. Lawyers are the headline example, since they could jump straight to relevant clauses and precedents instead of scrolling through long files.
- Legal teams: search leases, agreements, and NDAs, including scanned PDFs, and limit each matter to authorized users.
- Finance teams: point agents to approved policies, filings, and reports rather than outdated drafts, with results linking back to sources.
- Life sciences: search standard operating procedures, study documents, and regulatory submissions by meaning, not just keywords.
- Developers: plug the index into tools like LangChain, LlamaIndex, and Semantic Kernel through Box’s REST APIs or its MCP server.
The MCP support matters because more assistants now connect to outside data that way. A company building a custom copilot could feed it Box content without standing up a separate search stack, similar to how document review tools are being built straight into everyday apps.
What This Means for You
If your company keeps its files in Box, this could make internal agents cheaper and more reliable once it ships. If you are evaluating enterprise AI, it is another sign that the battle is shifting from the model itself to who controls the trusted business data underneath it.
- Check your plan: the feature is slated for Box Enterprise Advanced customers, starting early next year.
- Clean up permissions now, since retrieval will follow whatever access rules already exist in Box.
- Treat the speed and token savings as targets to test, not guarantees.
- Watch your token spend, because sending smaller chunks to a model is where much of the cost saving comes from.
Security-minded teams should also note that permission checks on every query reduce one common AI agent data security risk, where an agent surfaces something a user should never see.
Sources
Primary source: Box Context Retrieval product page. Additional reporting from TechCrunch. More AI updates are on our AI news page.
Conclusion
Box is betting that AI agents work best when they get a few precise passages instead of entire documents. By building Nvidia’s Nemotron embeddings into Box Context Retrieval, the company wants to make enterprise search faster, cheaper, and permission-aware by default. The numbers Nvidia shared are promising, but they are still projections. The real test comes early next year, when Enterprise Advanced customers start putting it to work.
