Question 1

Does my personal data get sent anywhere?

Accepted Answer

Only the labels do. The request to Jev contains field label fragments and your profile keys, never names, addresses or numbers, and a dialog lists exactly what will be sent so nothing goes out until you click. Confirmed mappings store keys only rather than values, which the README says was tested, and Jev is reached through the Vercel AI Gateway, which declares zero data retention and no training for this model.

Question 2

What happens when Jev is not confident about a field?

Accepted Answer

Three bands, and the low-confidence case never fills silently. At or above 0.8 the field is filled and shown green, between 0.5 and 0.8 it is filled and asks you to check it in yellow, and below that it is left blank and grey so you can tap to pick the item yourself. The mapping, not the answer, is what gets remembered, keyed to a form fingerprint built from field names and rounded positions.

Question 3

Is the accuracy good enough to trust the green fields?

Accepted Answer

Not with plain item names, which is the useful finding. On 10 OCR-noisy fields through the production relay at 1.4 seconds, plain names agreed 7 of 10 but the tax district field was confidently mistaken for the tax signature field at 0.89, which would have filled it. Adding aliases to the descriptions dropped agreement to 6 of 10 and produced zero confident mistakes, with that field falling to 0.56 and asking you to check. The aliases ship because green has to be trustworthy.

Question 4

Which forms does this actually handle?

Accepted Answer

Three kinds: a text layer with no fields, where city hall Word exports put boxes in as closed vector paths that pdf.js expands under a tracked CTM; a text layer plus AcroForm, which some bank forms use; and an image plus AcroForm, which is most 国税庁 forms and needs on-demand OCR. Encrypted no-changes forms are decrypted in the browser with qpdf-wasm. Not supported: skew correction, handwritten forms, XFA, vertical writing into fields and linked fields across pages.

Question 5

How does the fill actually work?

Accepted Answer

Drop a form; if it is encrypted, as most NTA forms are, it is decrypted in the browser with qpdf-wasm. Fields are found as AcroForm widgets or closed vector paths, nearby text comes from the text layer or OCR, a dialog shows exactly what will be sent, and Jev answers one choice per field in a single request before the form is filled and flattened.

Question 6

What is the preset path?

Accepted Answer

Popular forms ship as presets, so they fill with no OCR and no API call at all. A form's fingerprint, built from field names and rounded positions, keys both the presets and the mappings you confirm by hand. On the NTA opening-notification form the preset path filled 25 fields with zero relay calls, splitting a date and phone number correctly.

Question 7

What kinds of forms will it refuse to handle?

Accepted Answer

It will not correct skew on bad scans, read handwritten forms, handle XFA forms, write vertically into fields, or link fields across pages. The client is a single HTML file with a thin relay in api/, and the project is MIT licensed.

Jev Fill PDF

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community