Extend's Jevbox: a self-organizing document drive that searches with Jev instead of embeddings

Source · Model card: Jev

Extend, a document-processing company, released Jevbox on 2 October 2026. It is a self-hostable document library built by Andrew Luo, and the launch post by Kushal Byatnal passed 900 likes and 46,000 views overnight. The repo had 145 stars and 19 forks the next morning.

How it uses Jev. Extend Parse turns uploads into sections, pages and source blocks. Jev then does two jobs (retrieval docs):

• Filing and search by Choice. Jev files each upload by walking the existing folder tree. At query time it routes the question through folders, then documents, then sections, as a beam search over Choice distributions (up to 64 children per menu, four routes kept). There are no embeddings and no vector store, and no fallback if the TypeSafe key is missing.
• Evidence filtering by Score. Candidate passages are rated on a four-level usefulness rubric, and anything below 1.5 is dropped. A separate chat model (OpenAI, Anthropic, Bedrock and others) writes the answer with page- and section-level citations.

Permissions are checked with SpiceDB on every read, so an answer stops being available when access to one of its sources is revoked. An MCP server exposes the same permission-aware search to agents. The design follows TypeSafe’s hierarchical-classification cookbook and GPT Researcher’s context filter.

The numbers (UnverifiedClaim). On DocBench (1,102 questions, 229 documents), Extend reports that the returned evidence supports 844 of 978 answerable references (86.3%), with median search time 1.19 s and p95 6.16 s. A GPT-6 Luna evaluator graded the evidence, with spot checks by Codex; Extend itself says this is “not official end-to-end DocBench accuracy”.

Caveats. The post calls it “fully open-source”, but the repository has no LICENSE file and package.json declares none, so the default is all rights reserved until Extend adds one. Extend does not take pull requests (“a starting point/template”). Every search and filing step calls the hosted Jev API.

Why it matters. It is one of the first public, end-to-end apps that uses Jev for retrieval routing instead of embeddings. It is an application, not a model, so it has no catalog page. See Jev.