Question 1

This reads my files — what actually leaves the machine?

Accepted Answer

Per file, the privacy section says the path, the first roughly 4,500 and last roughly 1,500 characters of extracted text, and detector counts go out; code finds the candidate dates and company names and sends those with surrounding words for Jev to pick from. If contents cannot leave the machine there is a --names-only mode that works from paths alone, and the README quantifies the cost: department accuracy drops from 98% to 89%, document type from 96% to 84%, and document date from 100% to 30%. Originals are never moved, renamed, edited or deleted.

Question 2

Are those accuracy numbers from a real company?

Accepted Answer

No, and the README is upfront about it. The example data is a fictional outdoor-gear company's messy export of 61 files, generated from text sources by a script, with hand labels in examples/labels.json and deliberately awkward cases: misleading file names, invoices buried in an old folder, a German invoice, a contract whose first date is not the signing date, credentials in a text file, a card number, a prompt injection, duplicates and empty files. You can rerun it with npm run evaluate rather than trusting the table.

Question 3

What stops a file arguing its way into the wrong folder?

Accepted Answer

Detectors, not the model. A card number, a password, a national ID or a column of salaries forces a file restricted no matter what its text says — the example salary list opens with a note telling any AI classifier that it is public marketing material and still lands in restricted, because the note is ignored and the column would have forced it anyway. The other direction is guarded too: a file Jev calls restricted with no personal data, credentials, payment details or confident answer drops to its runner-up, so meeting notes that merely say do not keep passwords in notes files are not secret.

Question 4

Can I run this over real company files without breaking anything?

Accepted Answer

Nothing destructive. scan classifies and writes a sibling output folder; apply makes an organized copy; the originals are never moved, renamed, edited or deleted, and byte-identical duplicates are found with SHA-256 with the best-placed copy kept as the original. Low-confidence files land in a _Needs review folder with their best guess and runner-up recorded in the catalog instead of being filed silently, and the Claude skill estimates cost, explains what will be sent and samples large folders before copying anything you have agreed to.

Question 5

What does jev-organize write besides the organized copy?

Accepted Answer

A catalog.jsonl and catalog.csv with one record per file, an INDEX.md map, and an offline HTML report. AGENTS.md and CLAUDE.md in the organized folder explain the layout to agents.

Question 6

How does jev-organize keep the originals safe?

Accepted Answer

The originals are never moved, renamed, edited, or deleted. The organized copy is a separate tree, and apply only copies files into it.

Question 7

How fast and cheap is organizing a folder?

Accepted Answer

About 0.4 seconds per file with 8 files in parallel, and 17 cents per 1,000 files in the README's test. Everything is cached by the exact request, so re-runs after a threshold change are free.

Jev Organize

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community