Question 1

Does my document text get sent to the API?

Accepted Answer

Not the whole thing. Each document is reduced locally to a structured profile capped at 4,500 characters, containing likely headings, locally extracted subject candidates and representative excerpts from the beginning, middle and end. Short documents keep their full text only when that is already under the limit. The README states the original text never leaves the app beyond that bounded context.

Question 2

What happens when Jev is unsure about the category?

Accepted Answer

It goes to a `Need review` folder, and the cutoff is a category probability below 0.75. The audit records both where the file actually landed and what Jev originally suggested, so a reviewer can move it quickly. If no probability comes back at all, the audit shows `Unavailable` and the file is still filed in the selected category.

Question 3

Can this destroy my files if a category is wrong?

Accepted Answer

Not by overwrite. Every moved file keeps a View action that renders PDFs directly, DOCX as extracted text and PNG, JPEG, GIF and WebP as images, and Undo restores the file to the selected root folder, renaming it with a number when the original name is taken. Each file only moves after its own successful decision, not because a batch call succeeded.

Question 4

How many API calls does a big folder of files cost?

Accepted Answer

Longer documents each get their own request, but very short ones are grouped into a shared request of up to 8 documents and 4,500 context characters. Every run reports its duration, input tokens and estimated cost, which is how you get a number before committing a whole folder, since the README does not quote a per-document figure.

Question 5

How do I set it up, and where does the API key come from?

Accepted Answer

npm install, then npm run dev, and open http://localhost:5173. The model is typesafe-ai/jev reached through Vercel AI Gateway, so you need a Vercel account: sign in, open the JEV model page at vercel.com/ai-gateway/models/jev, enable AI Gateway and create a key, then paste that key into the app. The application creates config/config.toml when it is missing, and you can optionally save an absolute default source-folder path in Gateway settings, which the local server then reopens automatically.

Question 6

What does Jev actually decide about each document?

Accepted Answer

Four typed decisions per document: category, confidentiality, a prompt-injection risk score and subject. The category drives the filing destination, and where a category probability comes back below 0.75 the document goes to Need review instead. Unsupported, unreadable and empty files go to Not processable without a model call, and documents with a suspicious injection score get their own folder, Suspected prompt injection. Those three names are reserved by the app, so they cannot double as category names.

Question 7

What ends up in Not processable?

Accepted Answer

Three things, and none of them reach a model call: a file type the local extractor does not support, a file it cannot read, and an empty one. Everything else is classified. The audit still records the destination and, for reviewable cases, the category Jev originally suggested, so a reviewer can move a file quickly without re-running anything.

Jev Document Classification

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community