Chat with PDF100% ON-DEVICE
Ask questions about any PDF — summaries, clauses, numbers, deadlines. The AI model runs on your device: your document and your questions are never uploaded anywhere.
Loading…
Chat with a PDF Without Uploading It
Every popular "chat with PDF" service works the same way: your document goes to their servers, your questions go to a cloud AI, and you trust their retention and training policies. That's a hard no for contracts, medical records, financial statements, or anything under NDA — the documents people most want to interrogate.
This tool flips the architecture: a compact language model (Llama 3.2 class) downloads once into your browser's cache and runs on your own hardware via WebGPU. The PDF is read, indexed, and searched locally; the model answers from the most relevant passages and cites the pages it used, which you can expand and verify. Disconnect from the internet after the model loads — it keeps working.
Honest trade-off: an on-device model is smaller than ChatGPT. It's good at summaries and finding what a document says, less good at complex reasoning — that's why every answer shows its sources. For sensitive documents, private-and-verifiable beats smart-but-uploaded.
You might also need
FAQ
Is anything sent to a server?
The model weights are downloaded once (like any web asset) and cached. After that, everything — reading the PDF, searching it, generating answers — happens on your device. Your file and your questions are never transmitted.
Why does it need a download, and which model should I pick?
Because the AI runs on your machine instead of a server. "Fast" (~880 MB) is the sweet spot for most laptops; "Best" (~2.3 GB) gives better answers on capable machines; "Tiny" (~370 MB) suits older devices. The download happens once per browser.
What are the requirements?
A desktop/laptop browser with WebGPU — recent Chrome, Edge, or Safari — and a few GB of free memory. Phones generally aren't supported yet.
Can it read scanned PDFs?
Scans have no text layer. Run the document through OCR with the "Searchable PDF" output first, then chat with the result — still fully local.
How accurate are the answers?
Good for summaries, definitions, and "what does it say about X" — but it's a compact model, so always check the cited pages (every answer lists its sources). Don't rely on it alone for legal or medical decisions.
Found this tool helpful? Share it