Reads the document's text layer and rebuilds the structure: type size becomes heading levels, bullets and indentation become lists, aligned columns become tables where they really are tables.
Result
pages
Because it is a scan. A scanned page holds a picture of text, not text, so there is nothing to extract — no converter can recover a character that was never in the file. Reading letters out of an image needs OCR, which this tool does not do. The tool says so plainly and names the pages instead of handing you an empty file.
It is inferred, not read. PDFs do not record that something is a heading; they record glyphs at a size and a position. Type size, weight, indentation and column alignment are strong signals and they are usually right, but check headings and tables before relying on them.
A table in a PDF is often just text at aligned positions, with no grid in the file at all. Columns are recovered where several consecutive lines start at the same horizontal positions; a table with merged or wrapped cells will come out flatter than it looks on the page.
No. Both are excellent and both are Python — they run on a server, which would mean uploading your document. This converter runs in your browser, so it uses the browser's own PDF text layer and its own heuristics.