A model small enough to run in a browser cannot hold two hundred pages in its context, so handing it the whole document is not an option. The document is cut into overlapping passages instead, and for each question the passages that share the most distinctive words with it are found and only those are read. Every answer lists the pages it drew on, and those passages are shown in full underneath, so a claim can be checked against the document rather than taken on trust. The model runs on your own hardware through WebGPU; the document and your questions stay on the device.
Features
The document is cut into overlapping passages with their page numbers kept
Retrieval weights rare words above common ones, so a distinctive term wins
Every answer cites the pages it used, with the passages shown underneath
Nine models from 350 MB to 2.2 GB, cached after the first download
Follow-up questions keep the recent conversation in view
Warns when a document has too little text to question
Copy the whole conversation as text
How to use the Chat with PDF
Drop the PDF onto the page
Pick a model — the first use downloads it once
Ask a question, or start from one of the suggestions
Read the answer and open the cited passages to check it
Example
Input
lease.pdf — "What notice period does either party have to give?"
Output
Two months in writing, by either party. (p. 3) · Answered from Page 3
The passage it used is shown underneath so the claim can be verified.
Common errors & troubleshooting
The browser cannot run the model. — It needs WebGPU, which desktop Chrome and Edge provide. Without it, the PDF Summarizer works and Extract Text gives you the words to use elsewhere.
Almost no text could be read. — The PDF is a scan. Run it through text recognition first; there is nothing to retrieve from a page with no text layer.
The answer is confidently wrong. — Small models do that. Open the cited passages — if the answer is not in them, it was invented, and a larger model usually helps.
The first question is very slow. — The model is downloading, which happens once and then it is cached. Later questions start immediately.
Frequently asked questions
Is my document sent to a server?
No. The text is extracted in this tab, the retrieval runs here, and the model runs on your own GPU. The model weights are downloaded once from a public host and cached by your browser.
Why does it only read a few passages?
Because a model of this size has a limited context. Feeding it the four most relevant passages gives better answers than truncating the document at an arbitrary point.
Which model should I pick?
Start with the 1B recommended one — around 750 MB, and good enough for straightforward questions. Move up if the answers feel thin, and note that the larger models need a capable GPU.
Can it answer questions that span the whole document?
Poorly. Retrieval finds passages, so "what does clause 7 say" works well and "summarise every obligation across 200 pages" does not. Use the summarizer for the second kind.
Do the citations guarantee the answer is right?
They let you check it, which is the point. Small models still misread a passage they were given, so the citation is a starting point for verification rather than a proof.
Related tools
PDF Summarizer — Summarise a long PDF section by section, on your device.
PDF to Text — Extract selectable text from a PDF as plain text or Markdown.
PDF OCR — Recognise the text in a scanned PDF and make it searchable.
Audio to PDF Transcript — Transcribe a recording on your device and export it as a PDF, SRT or VTT.
Compare PDFs — Diff the extracted text of two PDFs line-by-line.
PDF Read Aloud — Listen to a PDF read out loud, passage by passage.
PDF Privacy Scanner — Find the metadata and personal data hiding inside a PDF.
More PDF tools
Image to PDF — Combine images into a single PDF (auto / portrait / landscape).
PDF to Images — Render PDF pages to PNG or JPG images you can download.
Merge PDFs — Combine multiple PDFs into one, with reordering.
Split PDF — Extract pages or ranges (e.g. 1-3,5,7-9) into a new PDF.
Rotate PDF — Rotate all pages or a selection by 90/180/270°.
Organize PDF — Reorder, rotate and delete pages from thumbnails, then export.
Add Page Numbers — Stamp page numbers onto a PDF (position, format, start).
Watermark PDF — Add a diagonal text watermark to every page.
Crop PDF — Trim margins from every page via the crop box.
PDF to Text — Extract selectable text from a PDF as plain text or Markdown.
Compare PDFs — Diff the extracted text of two PDFs line-by-line.
Encrypt PDF — Password-protect a PDF with AES-256 or RC4 encryption.
Decrypt PDF — Remove password protection from an encrypted PDF.
Compress PDF — Shrink a PDF by re-rendering each page to a JPEG at a chosen quality and resolution.
PDF Metadata Editor — View, edit or strip a PDF’s title, author, subject, keywords and producer.
Resize PDF — Rescale every PDF page to A4, Letter or another standard size, content centred.
Flatten PDF — Lock filled form fields into the page so a PDF looks the same everywhere.
PDF N-up — Print 2, 4, 9 or 16 pages per sheet, with a booklet option for folding.
Interleave PDF Pages — Merge odd and even scan passes back into one document, reversing if needed.
PDF to Single Long Page — Stack every page onto one continuous sheet you scroll instead of paging.
PDF Form Data Extractor — Pull the answers out of a filled PDF form as JSON or CSV, field by field.
Blank Page Remover — Find and delete blank pages from a scan, with a tolerance you control.
Text to PDF — Turn typed or pasted text into a paginated, downloadable PDF.
Markdown to PDF — Render Markdown — headings, lists, tables and code — as a PDF.