Every answer shows the matching passages with their page numbers, so you can check the source before relying on it.
Three steps, all in your browser:
- Open your PDF with the file picker - it is read in your browser and never uploaded.
- Type a question in plain language and press Ask the document.
- Read the result: the most relevant passages with their page numbers, plus a written answer when a built-in or on-device model is active.
Chat with PDF - Ask Your PDF Questions Privately
Open a PDF and ask questions about it in plain language, right in this page. The file is never uploaded anywhere - the text is read out of the PDF in your browser, the matching passages are found on your device, and every answer cites the page numbers it came from, so you can check the source before relying on it.
What you get without downloading anything
Without any AI model, the tool still finds and shows the most relevant passages for your question, each labeled with its page number. On a long contract, manual, or report, that alone answers the everyday job of locating where the document covers something - type the question, jump to the cited page. A built-in sample document and sample question render a live result the moment the page loads, so you can see the output shape before opening your own file.
Optional on-device upgrades
Two upgrades are offered, and both stay on your device. A small matching model (about 25 MB, one-time download, cached for later visits) upgrades passage search from keyword overlap to matching by meaning, which helps when your wording differs from the document's. Separately, a small language model turns the matching passages into a written answer; its download size is read from the model list and shown before anything is fetched. If your browser ships built-in AI, the page uses it for written answers with zero download instead. The status line under the question box always names which engine produced the latest answer and which matching method ran.
Check the pages, not just the answer
Written answers come from a small on-device model drafting over the matching passages. That is useful for a quick first read, but it is not a professional summarizer - the passages and page numbers are printed with every answer precisely so you can verify the claim against the document itself before acting on it.
Scanned PDFs need OCR first
This tool reads the text layer of a PDF. A scanned or photographed document has no text layer, so it returns nothing here - the page tells you so instead of failing silently. Run the picture through the image-to-text OCR tool first, then ask your questions against the recognized text. Very long documents are read up to the first 300 pages, and the limit is stated in the tool itself.
When a different tool fits better
If you want the complete text of a PDF rather than answers to specific questions, the PDF-to-text converter extracts the whole document in one pass. If you want an open-ended AI conversation that is not grounded in a document, the private AI chat is the site's conversational surface - and the two pages share the same on-device model cache, so a model downloaded on one is instantly available on the other.
Frequently Asked Questions
Is my PDF uploaded to a server?
No. The PDF is opened and read entirely in your browser, the passage search runs on your device, and the answers are generated on your device too. There is no server behind this page, so there is nothing for the document to be uploaded to.
What do I get if I never load an AI model?
The tool still finds and shows the most relevant passages for your question, each labeled with its page number. That works instantly with no download, and for locating where a long document covers something it is often all you need. Loading a model adds a written answer on top of those passages.
Why does my scanned PDF return nothing?
Scanned or photographed documents have no text layer, and this tool reads the text layer. The page warns you when a file looks scanned instead of failing silently. Run it through the image-to-text OCR tool first, then ask your questions against the recognized text.
How reliable are the written answers?
They are drafted by a small on-device model working over the matching passages - a solid first read, not a professional analyst. Every answer prints the passages and page numbers it used, so check the cited pages before relying on the answer.
How large are the downloads, and do they repeat?
The optional matching model is about 25 MB. The language models are larger, and each one's size is read from the model list and shown before the download starts. Both are one-time downloads cached by the browser, and the language-model cache is shared with the private AI chat page.