Chat with PDF: Summary and Q&A

Open a PDF to get a summary or ask questions about it. Answers cite the page they come from, like (p. 4), and everything runs on your own device, so private documents stay private.

  • Downloads the AI model from Hugging Face once; your text never leaves this device
  • Free, no sign-up
  • Works on phones

Looking for an AI model on this device…

Choose a PDF file
or drag and drop here

What goes online

  • Downloads the AI model from Hugging Face once; your text never leaves this device.
  • Everything else happens inside this browser tab.

Often used next

Looking for something else?

DeskTray has hundreds of free tools that work the same way.

How to chat with a PDF

  1. Download an AI model once in Settings (the tool takes you there; the recommended one suits your device).
  2. Open a PDF. Its text is read page by page on your device.
  3. Choose Summary for an overview, or Ask questions and type what you want to know.
  4. Check the cited pages for anything important, then copy the summary or the answers.

Features

  • Summaries in three lengths, with page references
  • Answers taken from the PDF, citing pages like (p. 4)
  • See the passages each answer is based on
  • Follow-up questions in a short chat
  • Spots scanned PDFs and hands them to OCR PDF
  • Opens password-protected PDFs

What you can do with it

Open a PDF and choose Summary for an overview, or Ask questions to find something specific. It suits the long documents people rarely read in full: rental and loan agreements, insurance policy wordings, company circulars, exam and job notifications, annual reports, research papers and course notes.

Every answer ends with the pages it came from, such as (p. 4), and you can open “Passages the answer was based on” to read the exact text the model used. That lets you check a point in seconds instead of searching the whole file.

How the answers are found

First the tool reads the text of every page with the same PDF engine as PDF to Text. It then cuts each page into short passages of about 120 words; a passage never crosses a page, so citations stay exact.

The AI model can read only about 3,000 words at once, so it cannot look at a 50-page document in one go. When you ask a question, a keyword search (the BM25 method used by search engines) ranks every passage by how well it matches your words, and the best ones, up to about 2,000 words, are given to the model with their page numbers. The model is told to answer only from those passages and to say so when the answer isn’t there. If your question names a page, such as “what does page 6 say”, that page is included first.

For a summary, the document is read in parts of a few pages each. The model writes notes on every part, keeping page references, and then combines the notes into one summary. You can limit the summary to a page range such as 1-10.

Limits to know

  • Scanned PDFs have no text, only pictures of pages. The tool spots them and offers to open the file in OCR PDF, which adds the text; then open the result here.
  • Tables, columns and forms may come out in a jumbled order, and charts and images are not read at all.
  • Keyword search needs words that appear in the document. If you ask about “rent” but the agreement says “licence fee”, try that word too.
  • Small models can misread a passage or mix up numbers. Always check important answers on the cited page, and don’t treat answers as legal or financial advice.
  • A summary covers up to 40 parts at a time, roughly 150 pages of dense text. For longer documents, enter a page range and summarise in sections.

Private by design

Most “chat with PDF” websites upload your document to their servers. Here, the PDF is opened inside your browser and the AI model runs on your device, so bank statements, Aadhaar copies, medical reports and agreements never leave it. The only download is the model itself, once, from Hugging Face, after which the tool also works offline. Closing the tab clears the document and the chat.

Models and devices

It uses the same on-device models as the AI Writer, from Qwen3 0.6B (337 MB) to Qwen3 4B (2.1 GB), and a model downloaded in one tool works in the other. For questions about long documents, Qwen3 1.7B or larger gives noticeably better answers than the smallest model. The models need WebGPU: Chrome or Edge on a computer, Chrome on a recent Android phone, or Safari on iOS, iPadOS or macOS 26 or later.

Frequently asked questions

Is my PDF uploaded?

No. The text is read from the PDF inside your browser, and the AI model runs on your own device. Only the model itself is downloaded from Hugging Face, once. That makes the tool safe for bank statements, agreements and other private papers.

Why does it say my PDF has no text?

Scanned PDFs and photos of paper contain pictures of pages, not text. Run the file through OCR PDF first (there is a button for it), then open the searchable PDF here.

Does the AI read the whole document?

For a summary, yes: it reads the document in parts and combines its notes. For a question, it reads only the passages that best match your words, because the model can take in about 3,000 words at once. If an answer is missing, ask again using words that appear in the document.

Can it answer questions in Hindi or other Indian languages?

You can ask, and Qwen3 models will try, but small models are much weaker in Indian languages than in English, and keyword search works best when the question uses the same language and words as the PDF.

Does it remember earlier questions?

Yes, the last few questions and answers are kept, so you can ask follow-ups like “and what about the deposit?”. Clear chat starts afresh. Nothing is saved after you close the page.

More AI tools

All 4 AI Tools