JEV Document Classification
September 20, 2026 · View on GitHub
Local-first Vite/Express app that files root-folder documents into configured category folders using TypeSafe Jev (via Vercel AI Gateway) for category, confidentiality, prompt-injection risk, and subject choices.
| At a glance | Details |
|---|---|
| Source | Source |
| Tags | Open source · Free source build · BYOK |
| Product homepage | Project homepage — source-built local web app; no separate hosted product required. |
| Pricing and access | MIT source has no app purchase fee, checked 2026-09-20. Classification needs a Vercel AI Gateway API key. Provider usage can incur charges. Preview/undo stay local to the selected folder. |
| Jev evidence | Inspected server/classification.ts: createGateway + experimental_evaluate with model typesafe-ai/jev; typed Choice questions for category, confidentiality, injection risk, and subject. Live classification not run. |
| Disclosure | Free source access does not include Gateway usage. AI-assisted, independently curated listing; no commercial relationship was declared. Inclusion is not endorsement. Implementation reviewed from public source. |
| Maintainer | Charlyhno-eng. |
| Format | Vite + React front end with Express local server (jev-document-classification 0.1.0). |
| Platform and availability | Source build with Node.js and npm; npm run dev opens http://localhost:5173. Experimental personal tool. |
| Jev's role | Supplies typed filing decisions over locally extracted document profiles; application code owns extraction, folder moves, audit, preview, undo, and reserved folders (Need review, Not processable, Suspected prompt injection). |
| Requirements | Node.js + npm; Vercel AI Gateway key for live classification. |
| License | MIT. |
When to use
Use it to sort a messy local document folder into named categories with an audit trail and undo. Prefer tax-doc-classifier when the problem is IRS form/page identity from PDF page text. Prefer doc-router when routing PDF pages between local extraction and OCR rather than filing into folders.
How it works
The local server extracts readable text and builds a bounded document profile (headings, subject candidates, excerpts). server/classification.ts asks Jev through the Gateway; low category confidence routes to Need review, injection flags to Suspected prompt injection, and unreadable files to Not processable without a Jev call. Short documents can batch; files move only after a successful decision and are never overwritten.
Get started
git clone https://github.com/Charlyhno-eng/jev-document-classification.git
cd jev-document-classification
git checkout 17cdc20acb407e7cfb623ef47d32e1b8568a5c86
npm install
npm run dev
Open http://localhost:5173, enter a Gateway key, choose a folder, and start classification. Live runs send document profiles to Vercel AI Gateway / Jev and can incur charges. This listing did not run npm test or live classification.
Examples and demos
- README screenshots of the audit UI and filing run.
- Offline tests under
tests/server/(local moves without Jev) — not executed on the review host.
Limits and data handling
Bounded profiles (not full originals beyond the profile cap) leave the host for the Gateway when classifying. Keys live in local config; do not expose a public deploy with a shared key. Upstream cost/speed claims were not independently measured.
Review and maintenance
Reviewed on 2026-09-20 at commit 17cdc20: 0.1.0, MIT. AI-assisted source review of README, LICENSE, server/classification.ts, and package.json. No live Gateway calls.
Related: tax-doc-classifier, doc-router.