Box plans AI search that returns relevant passages within file permissions

Box announced Context Retrieval on November 5 as part of its expanded NVIDIA partnership, with general availability planned for Enterprise Advanced customers in early 2027. The engine is designed to give enterprise AI relevant passages from company files.

It combines exact keyword matching with semantic search. Box's product description says it creates and maintains an index over content already in Box. Developers can rank passages and group them by source file, giving a model focused context for a question. The design reduces the separate retrieval pipelines and external vector databases that teams would otherwise maintain.

Retrieval can span enterprise content or be narrowed to selected folders and Hubs. Returned passages carry source-file attribution. The support notice says live user permissions and Shield classification policies are checked before content is passed to downstream models. Even when applications share an index, the requester's access determines which content is returned. Connections use Box REST APIs or MCP, with no separate data movement or external vector database required, according to the support notice.

The announced scope includes scans, images, audio and video. Box gives the example of analyzing forms, photographs, recordings and video in an insurance claim package. That is a company-described application.

What the performance figures compare

The 20% accuracy claim compares hybrid with semantic-only retrieval. Recall@5 checks whether a target answer appears among five results. Internal testing compared keyword, semantic and hybrid retrieval with enterprise queries over public datasets: an 11-percentage-point gain, roughly 20% relatively.

The approximately 90% token reduction concerns feeding relevant passages rather than full documents. Its example contrasts a 100,000-token handbook with a 10,000-token section. Fourfold turnaround reflects smaller inputs and fewer tool calls. Fixed runtime overhead and output generation remain. Results are approximate and not guaranteed.

The methodology cites Peerify, a September 2 paper on verifying peer-review claims using manuscript evidence. Its appendix compares full-manuscript inputs with retrieved passages across six benchmarks: average consumption falls from roughly 21,080 to 2,180 tokens per claim, a 9.67-fold difference. Those research results concern scholarly claim verification, a different task from testing Box customer deployments.

Existing MCP controls let administrators disable a tool category, enable read-only tools, permit reads and writes, or customize individual tools. Categories include search, files and folders. Category changes take effect immediately; individual configurations display confirmation after saving. These controls determine the available tools alongside content access checks.

Current paid Box AI features use AI Units, with consumption depending on the action and Standard or Premium model choice. The documentation lists 20,000 monthly units for Enterprise Advanced, while the order form specifies the contracted allocation. Unused monthly allowances do not roll over, and paid features are blocked when their entitlement is exhausted. Allowances reset at the start of each month, and consumption across paid actions is aggregated. Only administrators can inspect the organization’s overall entitlement and usage. These existing consumption rules do not establish a separate Context Retrieval price.

Box says Nemotron 3 Ultra is already available in Box AI for Enterprise Advanced. General availability of Context Retrieval remains planned for early 2027.