Ask a PDF questions with on-device AIAI
Get answers about a document from AI that runs in your browser. The file never uploads.
How to ask a PDF questions privately
- 1
Add your PDF
Drop or select the PDF. Its text is read on your device; nothing is uploaded.
- 2
Type your question
Ask about clauses, figures, parties, dates, or anything in the document.
- 3
Answer on your device
A model in your browser reads the text and answers locally. The document never leaves your device.
Why ask your PDF on your device?
Chat-with-PDF tools are everywhere, but they work by uploading your document to a server. For a contract, a report or anything sensitive, that is exactly what you are trying to avoid. Running the model inside your browser means you get the answers and the document never leaves your machine.
Frequently asked questions
Does this upload my PDF to answer questions?+
No. The answer is produced by a model that runs inside your browser, so the document text is never sent anywhere. There is no endpoint that accepts your file. The live "0 B uploaded" indicator on the page measures outbound data in real time, so you can watch that it stays at zero. Every other chat-with-PDF tool uploads your file; this one does not.
Exactly which AI model answers, and with what settings?+
It depends on your browser. If your browser has WebGPU, Ask PDF uses Qwen2.5-1.5B-Instruct: a 1.5-billion-parameter instruction model, 4-bit quantized (q4f16_1), with a 4,096-token context window, run through MLC web-llm on your GPU inside a Web Worker. This is the same model the Summarize PDF tool uses. If instead your browser has a ready built-in AI with a Prompt API (for example recent Chrome with its on-device model provisioned), it uses that. The page shows which one applies and its details before you start.
How is the model installed, and from where?+
For the WebGPU model, pressing Download fetches the weights (about 850 MB) from this site over your own connection, same-origin, never from a third party. Your browser stores and caches them (via IndexedDB), so it is a genuine one-time download shared with the Summarize tool. For the built-in path, your browser downloads and manages its own model. In both cases only model weights are fetched; your PDF is never part of the download.
Where is the Qwen model hosted, and is it from a trusted source?+
We host it ourselves. The weights (about 850 MB) sit in our own AWS S3 bucket and are served over HTTPS through our CloudFront CDN, same-origin with this site, so at the time you use it nothing is fetched from a third party. The model is Qwen2.5-1.5B-Instruct, an open-weights model released by Alibaba’s Qwen team under the Apache-2.0 license, in the standard MLC web-llm build; we serve those published weights as they are and do not modify them.
How is this kept secure, and what stops the model from sending my data anywhere?+
The whole page runs under a strict Content-Security-Policy. Its connect-src is limited to our own origin plus a couple of privacy-preserving analytics endpoints that carry only anonymous page-view pings and never any file data, and there is no endpoint anywhere that accepts your document. The model runs in a sandboxed Web Worker that is bound by that same policy, so even the model code has no server it could send your PDF to; its open-source web-llm runtime is served from our own site, and the weights load over HTTPS. Your PDF text is only ever handed to the local model in memory. The live "0 B uploaded" indicator measures the page’s outbound data so you can watch it stay at zero, and our code and infrastructure run automated security checks on every change.
How do I know the model is already installed?+
After the first download it is cached in your browser, and the page detects that: the model card shows "Downloaded and ready on this browser" instead of a Download button, and answering starts immediately with no network. If you already downloaded the model for Summarize PDF, Ask PDF reuses it with no second download.
Can I choose the model, and where is each one stored?+
Yes. The tool has a Model selector with Auto, Built-in (Gemini Nano), and Downloadable (Qwen), plus a "Check my system" button that shows whether your device supports each. Gemini Nano is downloaded and stored by your browser (once, reused across every site). The Qwen model is stored in your browser’s storage for this site (IndexedDB), about 850 MB, downloaded once from this site and cached, and shared with Summarize PDF so it never downloads twice. For a full write-up of both models and when each runs, see the blog post linked on the tool.
When is Gemini Nano used versus the downloadable Qwen model?+
On Auto, the tool uses Gemini Nano when your browser already has it downloaded and ready (instant, no download), otherwise the Qwen model if your browser has WebGPU, otherwise it offers to download the built-in model. Since a ready built-in model is relatively rare and WebGPU is common, most people use Qwen. You can also force either one with the Model selector.
Does anything get sent while or after answering?+
No. Once the model is on your device, answers are generated entirely in your browser with no network at all, during or after. The only network event the feature can cause is the one-time model download coming to you; your document never goes the other way.
How good are the answers, and what are the limits?+
The model answers from the document you loaded. When the exact answer is not spelled out it gives the closest relevant detail rather than just refusing, and it replies naturally to greetings or off-topic messages. It is a small model (the 1.5-billion-parameter Qwen model on WebGPU, or your browser’s built-in model), so it is best at direct questions such as facts, clauses, figures, and short explanations rather than long multi-step reasoning. It has a limited context, so for very long PDFs it answers from the opening section; split a long file first for full coverage.
Can I ask more than one question about the same PDF?+
Yes. Ask as many questions as you like about the loaded document; each is answered on your device with no re-download, since the model stays loaded for the session. Load a different PDF any time to start fresh.
What are the requirements, and does it work on phones?+
It needs a desktop or laptop. The downloadable model needs WebGPU (recent Chrome, Edge, or other WebGPU-capable browsers) and a GPU with enough memory (about 1.6 GB of VRAM); the built-in path needs a browser with an on-device Prompt API. Phones and tablets are not supported: the model is a 1.5-billion-parameter AI (about 850 MB) that a mobile GPU cannot run, and loading it can exhaust memory and crash the device, so the tool does not offer it there. If a device cannot run it, the tool says so and does nothing; it never falls back to uploading your file.
What about scanned PDFs?+
A scanned PDF has no selectable text, so there is nothing to read. Run OCR PDF first (also on-device) to add a text layer, then ask about the result.
Common uses
Task-specific guides for the situations people reach for this tool for.