Chat with a PDF — Ask Questions, Get Answers With Page Numbers

Because it is a scan — a photograph of pages rather than a document. There is nothing to read and no amount of indexing will help. Run it through the PDF OCR tool first to add a text layer, then chat with the result.

Yes, and it always does. Passages keep their page number through splitting and indexing, so every answer arrives with the page and the exact text it was drawn from. A local model of this size is wrong often enough that an answer you cannot check is not worth much.

A few hundred pages is realistic on a laptop. Indexing time grows with length — it is roughly linear, at a second or two per ten pages once the model is loaded — and it all happens once, before the first question. Very long documents are better split first.

No. The file is read, split, indexed and answered entirely inside this browser tab. The only thing downloaded is the model itself, from Hugging Face, and that happens once and is then cached. Nothing about your document goes the other way — which is the point of using this rather than a service on a contract, a payslip or a medical letter.

The document is split into passages and each one is turned into a vector that captures its meaning. Your question is turned into a vector the same way, the closest few passages are found by comparing them, and only those passages are given to the answer model. This is the same retrieval-augmented approach the hosted services use; the difference is that here the retrieval and the answering both happen on your device.

It is a model between 65 and 400 megabytes, which is one to four orders of magnitude smaller than a hosted assistant. It is good at pulling out a fact that is stated plainly in one place and much weaker at questions requiring several parts of the document to be combined, or at anything needing judgement. That is exactly why the passages behind every answer are shown with it — the answer is a shortcut to the right paragraph, not a substitute for reading it.

It says so rather than inventing one. Two separate checks make that possible: if no passage is even topically close to the question, the question is never put to the model at all, and when the extractive model is used its confidence is thresholded, because that model always returns its best guess and only its low confidence distinguishes a guess from an answer.

Neither. There is no key to obtain, no sign-up, no quota and no per-question cost. The trade is that the models are small enough to run on your own hardware, and answer quality reflects that.

After the first run, yes. The model is cached by the browser, so a second visit works with no network at all. The first visit has to download it.