Ask PDF

Ask a PDF a question and see exactly where the answer is

Ask PDF answers questions about a PDF by pointing at the lines that contain the answer. Drop a contract, a manual, a policy document or a long report, type a question in plain language — “What is the notice period?”, “Is water damage covered?”, “When is the deadline for applications?” — and you get a verdict (answered, partly answered or not in this document), the best-matching lines with their page numbers, a preview of the page with the lines highlighted, and the lines around them for context.

It is built to be checked, not trusted blindly. The AI never writes an answer of its own: it can only choose among the lines of your document, so every answer is a quotation you can verify on the page. When the document does not cover your question, it says so instead of guessing.

Text extraction, and OCR for scanned pages, run in your browser. Nothing leaves your computer until you press the AI button.

How to ask a PDF a question

1. Drop a PDF onto the upload box. The text of every page is extracted in your browser and split into lines, each tagged with its page number.
2. Optionally set Pages to search (for example 1-20) — only those pages are sent. For scanned documents, pick the OCR language under Scanned pages.
3. Type your question and press Find the answer with AI (or Enter).
4. Read the verdict and the matching lines. Tap a match to see its page highlighted and the lines around it; tap a context line to add it to the quote.
5. Edit the answer box if you like, then press Copy answer with page reference. Earlier questions stay in the history list for this session.

Ask PDF features

How the answer is found

After extraction, each visual line of the selected pages becomes a numbered option. Your question and those lines go to Galaxy of Tools’ server, which asks TypeSafe’s Jev model two things: which line contains the answer (a probability for every line), and whether any line answers the question at all. For long selections — more than 250 lines — the server first ranks blocks of 20 lines and then looks inside the three best blocks.

The verdict comes from that second judgment: 70% or more is shown as answered, under 35% as not in this document, and anything in between as partly answered. Up to five lines are listed, best first. Because the model selects lines rather than generating sentences, it cannot misquote figures or invent a clause that is not there.

Where it works well — and where it doesn’t

• Good fit: contracts, terms and policies, product manuals, HR handbooks, research papers, tender documents, insurance wordings and long reports where you need “where does it say…?”.
• Answers spread over a page (a total that depends on several rows, or a yes that follows from two clauses) may show as “partly answered”. Read the context lines, and add them to the quote.
• Scans with poor OCR — faint, skewed or handwritten pages — can hide an answer simply because the words were misread. Check the page preview, or improve the scan with PDF OCR first.
• Non-English documents work, but less reliably than English. Ask in the document’s language for the best match.
• Very long PDFs: one question searches up to 1,200 lines (roughly 30–50 pages of dense text). Use a page range to search the part that matters, then ask again for the next part.
• Tables are read row by row; for a question about a whole column, try the PDF Table Extractor.

What is sent, and what is not

Opening the file, OCR, line splitting, the page preview and the history all stay in your browser. When you press the AI button, one request carries your question and the text lines of the pages you selected; the PDF file itself, its images and its layout are never uploaded. The server does not store the text or your question. For reading or searching a PDF without any AI, use the PDF Viewer; to get the key points of a whole document, try the PDF Summarizer.