Chat with your PDF, without uploading it

What you can drop in

It reads PDF files with citations linked to the exact page, Word .docx and legacy .doc with section headings preserved, PowerPoint .pptx slide decks, and plain Markdown or text files. A single session holds up to 25 documents or 300 MB combined, and a PDF is read up to its first 500 pages, past that it says so instead of quietly cutting off. A scanned PDF is a photograph of a page, not text: if a file has no real text layer, the app tells you plainly instead of pretending to have read it. Recognizing text in scans is planned, not built yet.

Finding the right paragraph, not just any paragraph

A forty-page lease might mention "landlord" four dozen times, in the boilerplate, the notice clause, and the insurance clause, and actually define who it is in exactly one place. A search that just matches similar wording will happily hand back the wrong four dozen. So before it searches, the app tries to build a table of contents for your document. Word files already carry real heading styles, so that map comes for free. PDFs don't carry that kind of metadata, so the app looks for the pattern contracts and reports actually use: a number, a run of capital letters, a colon, like "7. SECURITY DEPOSIT DETAILS:", and only trusts it once it's seen at least two of those in a row; one lucky match doesn't count as structure.

When a document has that kind of outline, the app shows it to the same on-device process that's about to write your answer and asks it to guess which two or three sections are worth searching first. That guess only ever nudges the ranking; it never replaces the actual search So a bad guess costs nothing, and a document with no detectable structure, like an email or a plain-text note, is simply searched the ordinary way.

Two ways to get an answer

Every browser can do passage search: ask a question and within a couple of seconds you get back the exact sentences that answer it, tagged with which document and page they came from, no download, no setup, works the moment you drop a file. Written answers (a sentence or two in plain language, built from those same passages) need a browser that can put your device's graphics hardware to work; most current Chrome, Edge, Firefox, and Safari qualify, but not every install does.

On a capable browser you get a real choice before anything downloads: a smaller download that answers faster, or a larger one that tends to be more careful, sizes shown up front. Decline both and passage search keeps working the same; you just don't get the paragraph on top. If a device that looked capable turns out short on graphics memory once it's running, the app steps down to the smaller option and says so, never a stuck composer.

Why you can verify every answer

If the best passage it can find doesn't clear a minimum relevance bar, the app doesn't guess: it says plainly that it couldn't find anything about that in your documents, and shows you the two closest passages anyway in case they help. That check runs before any answer gets written, on purpose: skipping a bad generation is cheaper than catching one after the fact.

After an answer is written there's a second check: each sentence is compared back against the passage it cited. Sentences that hold up are marked verified; ones that don't, or that claim something with no citation attached, get a quiet dashed underline instead of being deleted; you still see everything the app wrote, with a hint about what to double-check yourself. If more than half a turn's sentences fail that check, it offers to show you the passages instead.