How it works
Everything happens inside your browser
Unformation is a web page that behaves like a desktop program. Your document is opened, searched and rewritten by code running on your device. There is no upload endpoint to send it to.
The four steps
Upload, choose a rule, review, download: in technical detail
- 1
1. Upload or paste
Drop up to 10 files (50 MB each) or paste text. Your browser reads the file into memory; no request is made.
- The file is read with the browser's FileReader API into an ArrayBuffer, memory that belongs to the page, not to any server.
- The buffer is handed to a Web Worker, a background thread of the same page, so the interface stays responsive while large files are processed.
- DOCX, PPTX and XLSX are ZIP archives; the worker unpacks them with JSZip and locates the XML parts that hold text. PDFs are opened with pdf.js, which extracts the text layer page by page. TXT, MD, CSV, JSON and HTML are read as plain text.
- 2
2. Choose the rule
Decide how each category is rewritten: numbered labels such as Person 1 / Company A, realistic fake values, masking or full redaction. Optionally enable smart detection and add words to your dictionary.
- A strategy is set per category (person, organization, email, phone, IBAN, ID, address, date and so on) and stored only in the page's memory for this run.
- Smart detection is off by default. When you enable it you pick a model: about 110 MB for English or about 180 MB for the multilingual model used for Turkish. It is downloaded once from the Hugging Face Hub, cached by the browser, and runs locally after that.
- Your dictionary (words that must always be replaced) is kept in localStorage on this device and applied on every run; nothing about it is sent anywhere.
- 3
3. Review what will change
Every detected item is listed with its category and planned replacement. Switch items on or off, add words that were missed, correct a category.
- Layer 1, patterns and checksums: emails, phone numbers, IBANs (mod‑97), card numbers (Luhn), Turkish national IDs and tax numbers, URLs, IP addresses, plates and address hints.
- Layer 2, your dictionary: exact words you always want replaced, matched whole-word and case-insensitive; Turkish suffixes are kept intact.
- Layer 3, optional smart detection: the named-entity model you chose in step 2 runs in the browser through ONNX Runtime and adds people, organizations and places that patterns cannot see.
- The three layers are merged, overlaps are resolved and consistent replacements are planned: the same name always becomes the same placeholder. The review list is exactly what will be applied.
- 4
4. Download and use
Get the same file back with formatting intact, a Markdown version for pasting into an AI, and a mapping table to reverse the answer later.
- For Office files the replacements are written back into the very same XML runs, so styles, headers, footers, comments, tables and images stay untouched. The archive is re-zipped in the worker.
- For PDFs the extracted, anonymized text is assembled into a new DOCX and a Markdown file; the original PDF is not modified.
- The result becomes a Blob, a file in memory, and is saved with a normal browser download. Closing the tab discards everything except your dictionary and the setting that hides the first-visit note.
Data flow
One box: your device
The diagram shows every place your document goes. All of them are inside your browser.
Verify it yourself
Do not take our word for it
Three checks that take under two minutes. They work in Chrome, Edge, Firefox and Safari.
Network tab
- Open developer tools (F12 or Cmd+Option+I) and choose the Network tab.
- Load the page, then upload and anonymize a document.
- Look at the list of requests. You will see the page's own scripts and, if enabled, one download of the model. No request carries your file, there is no POST with your document in it.
Wi‑Fi off
- Open the tool once so your browser caches it.
- Turn off Wi‑Fi or switch to airplane mode.
- Anonymize a document. Pattern and dictionary detection, rewriting and download all keep working. A server-side tool would stop here.
Optional: service worker and storage
- In developer tools open the Application tab (Firefox: Storage).
- Under Service Workers you will see the worker that caches the app for offline use; under Cache Storage the app shell and, if enabled, the model files.
- Under Local Storage you will find only your dictionary and a flag for the first-visit notice. There are no cookies.
Formats
What is preserved, what comes back
| Format | What is preserved | Output |
|---|---|---|
| DOCX (Word) | Styles, headers and footers, comments, footnotes, tables, images, tracked-change text | Anonymized DOCX + Markdown |
| PPTX (PowerPoint) | Slide layouts, notes, tables, charts and images | Anonymized PPTX + Markdown |
| XLSX (Excel) | Sheets, formulas, number formats, shared strings, comments | Anonymized XLSX + Markdown |
| Text content and reading order (layout is not reproduced) | DOCX + Markdown | |
| TXT / MD | Text as-is, line breaks | Same format |
| CSV | Columns, delimiters, quoting | Same format |
| JSON | Structure and keys; only string values change | Same format |
| HTML | Tags and attributes; only text nodes change | Same format |
Every job can also produce a mapping table (CSV or JSON) that the de-anonymizer uses to put the original names back into an AI's answer.
Limits
What it does not do: honestly
Scanned PDFs
A scanned PDF is a picture of a page. There is no text layer to read, so nothing is detected. Run OCR first (many scanners and PDF apps include it) and try again.
Detection is not perfect
Checksummed patterns are exact. Names, companies and places depend on your dictionary and the optional model, which can miss unusual names or flag ordinary words. The review list exists precisely so you can correct it before download.
Complex PDF layouts
Multi-column pages, forms and tables may come out in a different order in the DOCX and Markdown output. The words are anonymized; the layout is not reproduced.
Legacy .doc, .ppt and .xls
The binary Office formats from before 2007 cannot be edited reliably in a browser. Save them as .docx / .pptx / .xlsx first.
Text inside images
Names embedded in pictures, screenshots or scanned signatures are not read and not changed.
Password-protected files
Encrypted documents cannot be opened. Remove the password, anonymize, then protect the copy again.
Ready when you are
Open the tool, drop a document and watch the Network tab stay quiet.