How the document assistant works
Upload a PDF, a scan or a photo of a document. The assistant reads its text — the PDF text layer directly, scanned pages through OCR — and then does one of two things: writes a summary with the key points, or answers a question you type about the document.
Every answer is built from passages of your document only, and the passages it relied on are quoted with their page numbers, so you can jump to the source instead of trusting a paraphrase. The language model runs on FileConvertLab's own servers; see the privacy policy for how uploaded files are handled.
Summaries and key points
Summarize produces a few sentences on what the document is, who it concerns and what it states, followed by up to eight key points: figures, dates, obligations and conclusions. It is written in the language of the page you are using, whatever the language of the document.
The summary covers the first 10 pages and up to 10,000 characters of text; for longer documents it says that only the beginning was summarised. Ask questions to reach content further in. To count words and pages before reading, use the PDF word counter.
Asking questions
For questions the assistant indexes up to 30 pages: the text is split into overlapping passages, each passage is turned into a multilingual embedding, and your question is matched against them. The six closest passages go to the language model, which answers from them alone and says so when the document does not contain the answer.
Because matching works on meaning rather than exact words, a question can be asked in a different language from the document. Keep questions specific — "What is the notice period?" finds the right clause more reliably than "Tell me about the contract". You can save the whole conversation as Markdown notes.
Limits
Files up to 50 MB; summaries read the first 10 pages, questions search the first 30 pages; a question can be up to 500 characters. Each summary or question counts as one heavy operation against your daily limit.
Language models can misread tables, miss details spread across passages or state something the passage only implies. Treat answers as pointers to the right place in the document and check the quoted page. For fields you want in a table, the AI data extractor is the better tool.