Ask a PDF anything — with an AI that never uploads it

Bindery reads the document’s text on your device, indexes it into page-aware passages and retrieves the ones relevant to your question. A small instruct model then answers in your browser on WebGPU — Qwen 2.5, Llama 3.2 or SmolLM2, downloaded once into the browser cache — and every claim is cited with a page number so you can check it.

Prefer a bigger model? Point Bindery at an OpenAI-compatible endpoint with your own key (OpenAI, Groq, OpenRouter, Mistral or a local Ollama). Then only the retrieved excerpts and your question are sent, never the whole file, and the key stays in your browser. Other “chat with PDF” tools upload the entire document to their servers first. Bindery runs entirely in your browser — your file is never uploaded, so there is nothing to delete from a server afterwards.

How it works

1. Drop the PDF and open Ask this PDF — its text is indexed locally
2. Load the on-device model once (or save your own endpoint)
3. Ask, summarise or extract fields as JSON — answers cite pages

Chat with a PDF that stays on your device

Bindery reads the document’s text on your device, indexes it into page-aware passages, and retrieves the ones relevant to your question. A small instruct model then answers in your browser on WebGPU — Qwen 2.5 0.5B, Llama 3.2 1B or SmolLM2 360M, downloaded once (roughly 250 MB to 700 MB) into the browser cache — and every claim cites a page number so you can check it. Ask what a clause means, summarise a 60-page report, or extract fields such as parties, dates and amounts as JSON. Ask across several open files at once when the answer spans documents.

Bring your own model without giving up the file

Small models are excellent at finding and quoting the right passage and weaker at long reasoning. For harder analysis, point Bindery at an OpenAI-compatible endpoint with your own key — OpenAI, Groq, OpenRouter, Mistral, or a local Ollama. Then only the handful of retrieved passages and your question are sent to the endpoint you chose, never the whole document, and never anything to Bindery. The key stays in your browser. Other “chat with PDF” tools upload the entire file to their servers first; the difference matters for contracts, medical records and anything under NDA.

On-device AI versus upload-based chat

Getting good answers

• Scans need OCR first so the pages have text to read.
• Ask specific questions; the retrieval step finds passages that match your wording.
• Use the citations. The model can misread; the page reference lets you verify in seconds.
• WebGPU needs a recent Chrome, Edge or Safari; without it, use your own endpoint.

Questions people ask

How good are the on-device answers?

Small models are excellent at finding and quoting the right passage and weaker at long reasoning. Because retrieval and citations are done by Bindery, you always see which pages the answer came from; for hard analysis, use your own key with a larger model.

What exactly is sent when I use my own key?

The handful of passages Bindery retrieved for your question (typically a few thousand characters) plus the question itself, to the endpoint you configured. Never the full document, never to Bindery.

Does it work on scanned PDFs?

Run OCR first — it is on-device too — so the pages have a text layer to read.

Which on-device model should I choose?

Qwen 2.5 0.5B is the fast default. Llama 3.2 1B answers better at the cost of a larger download and a slower first load. SmolLM2 is the smallest for low-memory devices.

Is my document used to train anything?

No. On-device models run locally and learn nothing. With your own key, your provider’s terms apply to the excerpts you send.

Can it answer questions across several PDFs?

Yes. Open the files together and ask across them; answers cite the file and page.

Guides for this job

More Bindery tools