ask

Browser-native question answering over a PDF, using a local language model. Your document never leaves your device.

Live: https://ask.xjmani.com

Part of the xjmani tools collection.

What it does

Drop a PDF in the page, ask questions in the chat, get answers with citations that point back to the source paragraph in the document.

• PDF parsing runs in the browser via pdfjs-dist.
• Embeddings are computed locally with all-MiniLM-L6-v2 through transformers.js.
• Inference runs in your GPU through @mlc-ai/web-llm using Llama 3.2 3B Instruct (q4f16_1).
• Indexed chunks are persisted in IndexedDB so reopening a previously loaded document is instant.

The first time you load the tool, your browser downloads about 2 GB of model weights from Hugging Face. After that, it runs offline.

Requirements

WebGPU. Chrome 121+, Edge 121+, or Safari 18+ on macOS Sonoma or newer. Firefox does not yet ship WebGPU on by default.

Self-host

The whole tool is one index.html and a vendor/ directory. Serve it as static files behind a CSP that allows connect-src 'self' https://huggingface.co https://cdn-lfs.huggingface.co so the model can download. The included nginx.conf and Dockerfile run on any container host.

git clone https://github.com/xjmani/ask
cd ask
docker build -t ask .
docker run --rm -p 8080:8080 ask

Then open http://localhost:8080 in a WebGPU-capable browser.

License

MIT. See LICENSE.