Local PDF assistant
An assistant that reads your document without ever letting it leave your device.
Your files never leave your deviceHow to use the assistant on a PDF, step by step
The assistant runs a small language model directly on your device’s graphics card, to summarize, answer questions about, extract from, or translate a passage of a PDF, or to turn it into flashcards.
Drop in the PDF
Each page’s text is extracted in your browser, as with the text-extraction tool. If too few pages contain text, the document is treated as a scan and the assistant stays unavailable.
Choose and download a model
If your browser exposes WebGPU, choose between the “Standard” model (≈850 MB) and “More precise” (≈2.2 GB, offered only if the graphics card has enough memory); the download shows its progress and then stays in the browser cache.
Choose a function
Four tabs are on offer: Summarize, Ask a question, Dates and amounts, Translate.
To revise, choose “Flashcards”
The “Make flashcards” button draws questions and their answers from the document. Each answer is checked against the text: those not found in it are marked “check in your notes”. “Download for Anki (.txt)” exports them, one card per line.
Run it and read the answer
The answer text appears as it is generated; a “Stop” button interrupts the ongoing generation at any time.
Delete the model if needed
“Delete the model from this device” removes the downloaded file from the browser cache.
A password-protected PDF is refused, as everywhere else in VellumPDF. For date and amount extraction, any item the model proposes that cannot be found word for word in the document is automatically discarded before it is shown.
Frequently asked questions
Does my PDF’s content get sent to a server for the assistant to answer?
No. The language model runs in your browser, on your device’s graphics card, once downloaded. The network witness shown under each answer proves it: 0 bytes sent while the assistant is answering. Only the initial model download, announced before your click, comes from another site (huggingface.co and raw.githubusercontent.com).
Why do I need to download something before I can start?
The language model itself (about 850 MB by default, or 2.2 GB for the "more precise" option) is not bundled with VellumPDF: it downloads once, on your click, and then stays in your browser’s cache. A "Delete the model from this device" button erases it whenever you want.
What happens if my browser does not support WebGPU?
The panel explains it and offers no loading button: there is no fallback to a server. WebGPU is available in recent versions of Chrome, Edge and Firefox, and in Safari from version 26 onward. On phones it takes Chrome, Edge or Samsung Internet on Android 12 or later, or Safari on iOS 26; Firefox for Android does not expose it yet.
Are the assistant’s summaries and answers 100% reliable?
No, and the interface says so at every step: it is a small model running on your device, capable of a faithful summary but not a legal opinion. Always check important details in the document itself, especially extracted amounts and dates.
How does the assistant handle a document too long for a single request?
The text is split by page into chunks of about 2,400 tokens. A summary spanning more than one chunk is done in two passes — a summary of each chunk, then a summary of those summaries — and the interface says so in one sentence. For a question, the model, on your device, reads only the two most relevant chunks.
Can I make flashcards from my course notes?
Yes. “Flashcards” draws questions and their answers from the document, from up to four parts of a long course. Each answer is checked against the text: if its key words are not in it, the card is marked “check in your notes”, because a small model can be wrong. “Download for Anki (.txt)” exports the cards, one per line, for Anki, Quizlet or Mnemosyne.