Nothing leaves your device. Here is how to check.

1. The airplane-mode test

1. Open Sumi once while online, so your browser keeps a copy of the app.
2. Turn on airplane mode, or switch off Wi-Fi and unplug any cable.
3. Reload the page, drop a PDF, review the findings and click Redact & download.

Everything works offline. A tool that needed a server to do the work would fail here.

2. Watch the Network tab

1. Open your browser's developer tools: F12 on Windows and Linux, Cmd Option I on a Mac.
2. Open the Network tab, tick Preserve log, then clear the list.
3. Drop a PDF, toggle a few findings, and export.

The list stays empty. If you opened the app for the first time a moment ago, you may see the app's own files (scripts, fonts) load from this site. Those are requests for code, never requests carrying your document, and they disappear on the next visit because the app is cached.

3. Read the security policy

The page ships with a Content-Security-Policy. It instructs your browser to refuse any connection to a server other than this site. That closes the usual ways a script talks to a server (fetch, XHR, WebSockets, beacons). It is one layer of protection, not a guarantee by itself, which is why the two tests above matter most. To see it, open developer tools, go to Elements (or Inspector) and look for the Content-Security-Policy tag inside <head>. Its connect-src entry lists every place the page may talk to.

4. The live counter

While a document is open, the panel shows how many network requests the page made since you opened it, measured by your browser. It should read 0. It only sees requests made by the page itself, not by background workers, so treat it as a convenience; the Network tab is the reference.

Scanned pages and photos

When a page is a scan or you drop a photo, Sumi reads it with text recognition (Tesseract) running in this tab. The recognition engine and its language files come from this site and are kept for offline use: English and French are included, and other languages (about 1 to 4 MB each) are fetched from this site the first time a document needs them. If you turn on Find handwriting and signatures, a small text-detection model and its runtime are also downloaded once from this site (up to 32 MB). It boxes text the reader could not read, such as handwriting, signatures and stamps, so you can cover it. Photos are redrawn before anything else happens, which also drops their hidden metadata, such as the GPS location where they were taken.

The optional AI detection

AI detection is off until you turn it on. When you do, your browser downloads a language model once, from huggingface.co (about 110 MB). That request names the model and nothing else. The model then runs on this device: your document is analysed locally and never sent. In the Network tab you will see the model files load the first time, and nothing on later documents. The security policy allows exactly two destinations: this site and Hugging Face's model servers.

What the download contains

• Each page is redrawn as an image, and the black boxes are painted into the pixels. Nothing sits under them.
• The new file holds only those images: no text layer, no fonts, no title, author or creation dates, no attachments, no scripts. Filled form fields are flattened into the picture, and their text is checked like the rest of the page.
• We do not use analytics, cookies, or error reporting tools.

Automatic detection can miss things, especially in scanned pages, handwriting or unusual formats. Always review the findings, and open the result before you share it.