How chatting with a PDF works
Upload a document and we extract its text, split it into passages of roughly 500 tokens, and turn each one into a vector embedding. Your question is embedded the same way, the closest passages are retrieved, and only those go to the language model.
That matters for long documents. Pasting a 200-page report into a chat window either does not fit or buries the answer; retrieval pulls the three paragraphs that actually address your question, so the reply is short, specific and traceable back to the page it came from.
PDF, Word (.doc and .docx), plain text and Markdown all work, up to 10 MB per file and three files per demo.
What people use it for
Contracts and leases — ask what the notice period is instead of scanning clauses. Manuals and spec sheets — ask for the torque setting rather than hunting the table. Research papers — ask what the method was and what the sample size was before deciding to read the whole thing.
Finance and operations teams drop in a policy document and check what it says about a specific case. Students summarise a chapter and then quiz themselves on it.
You can add several documents to one demo and ask across them — useful for comparing two versions of the same contract.
Limits, cost and privacy
Free, no account, no card: three files, three URLs and ten messages per session. Files are stored only long enough to index them and are deleted, along with the index, after 24 hours. We do not train models on them.
Scanned PDFs that contain images of text rather than real text will come back empty — we do not run OCR in the demo. If you need a document assistant that stays, sign up free: the same engine becomes a permanent chatbot you can embed on your site or connect to WhatsApp.