How it works
Add files in any mix, all at once or a few at a time. Each one is read in your browser and turned into Markdown, the plain format that language models read best: headings stay headings, lists stay lists and tables stay tables. Every file shows its status and token count as it finishes, and the preview shows exactly what will be exported.
The steps on the right apply to every file. Switch one off and the preview updates straight away, without opening the files again. Removing a file or changing what is removed can be undone with Ctrl Z.
What each step does
- Convert to Markdown keeps the structure of PDFs, Word files and web pages, turns CSV files into tables and formats JSON as a code block.
- Clean up tidies spacing, removes the headers, footers and page numbers that repeat on PDF pages, joins words broken across lines and strips invisible characters such as zero-width spaces.
- Remove personal details replaces emails, phone numbers, card and bank numbers, ID numbers, street addresses, names, IP addresses, API keys and other secrets with labels. The same value always gets the same label, so the text still makes sense.
- Split into chunks cuts the text at sentence ends into pieces of the size you choose, with a little overlap so no idea is cut in half. Each chunk remembers the heading it sits under.
Tokens and context sizes
Tokens are counted with the o200k or cl100k tokenizer, the two most common today. The bar shows how much of a chosen context size your files fill together, so you can tell at a glance whether everything fits in one prompt or needs to be split. Other models count a little differently, so treat the number as a close guide and leave some room for your question and the answer.
Personal details: what is found and what is not
The check looks for patterns, such as the shape of an email address or a card number that passes its checksum. Names are found from clues like greetings, sign-offs, titles and labelled fields, plus any words you add yourself. Open Personal details to see every finding in context and keep the ones that are fine to share.
No automatic check catches everything. Read sensitive documents yourself before you share them, especially names written in running text, handwriting in scans and anything said in a recording.
Exports
Copy all puts every file into one Markdown block with a heading for each file, ready to paste into a chat. Download saves the same as one .md file, a ZIP with one .md per file, or a JSONL file with one chunk per line for embedding and search tools. The manifest lists each file with its token count, chunks and what was removed, and records the steps you used.
Images, recordings and downloads
Text in images and scanned PDFs is read with an on-device text reader, and speech in audio and video is transcribed with an on-device speech model. Each downloads once the first time you need it, and the workspace tells you its size and waits for your go-ahead before fetching anything. After that it works offline. Summaries can be picked from your own sentences, which needs no download, or written by your browser's built-in summarizer where one is available.
Privacy
Everything happens in your browser. Your files, the text taken from them and your choices never leave this device and are not stored anywhere, so they are gone when you close the page.
Questions
How many files can I add?
As many as your device can hold in memory. Text files and PDFs are quick. Images and recordings take longer because they are read on your device, one at a time.
Can it read old .doc files?
Not directly. Open them in a word processor and save them as .docx first.
Is the Markdown the same as the original layout?
It keeps the structure, not the look. Columns, page layout, fonts and pictures are left out, because a language model does not need them. Text inside pictures is only read for image files and scanned PDFs.