PDF to Markdown Converter

How this PDF to Markdown converter reconstructs structure

A PDF records where marks were painted on a page. Markdown records what a document means: this line is a heading, these items form a list, and these values belong to a table. Most PDFs do not store that meaning. A PDF to Markdown converter therefore cannot merely swap one file extension for another. It has to reconstruct a plausible reading order and label every place where the page does not provide enough evidence for certainty.

This converter starts with the complete PDF text layer and the positions of its text. It detects page-spanning headings before looking for column gutters, so a two-column article is read down the left column and then down the right rather than stitched together across each baseline. Font size, indentation and list markers provide conservative structure clues. When a PDF exposes a complete tagged structure tree, that explicit evidence takes precedence over geometric inference. Tagged heading levels, lists and table rows become their matching Markdown structures; other roles keep their readable text. Page numbers marked as PDF artifacts do not invalidate that structure or leak into the result. The report says which path each page used.

On geometry-inferred pages, list indentation becomes nested Markdown instead of being flattened into its parent item. Small raised footnote markers are attached to the nearby baseline before reading order is decided. A table-shaped region that is too wide or ambiguous to serialize remains readable text, and its measured column anchors are named in the report rather than disappearing silently.

On an untagged page, adjacent wrapped lines are grouped into paragraphs from their spacing, indentation and line length instead of becoming one Markdown paragraph per printed line. A lowercase word split at the line edge is conservatively rejoined when the evidence is strong; short or ambiguous hyphenated compounds keep their mark. The conversion report names both paragraph inference and this repair because a PDF does not store either decision.

Repeated running heads, footers and page numbers are kept rather than silently deleted. When the same margin line appears on at least three pages and 60 percent of the document, each affected page names it in the report. Full-width margin lines are excluded from the body-column test, so a header crossing the gutter cannot turn two columns of prose into a Markdown table.

Tables only become tables when the anchors agree

Wide spaces do not prove that a page contains cells. Two columns of prose, an address block and a price aligned at the right margin can all resemble a table to a coordinate parser. This page writes a Markdown table only when at least three rows share stable column anchors and the region does not have the signature of narrative columns. Pipes and backslashes inside cells are escaped, so a value such as A | B remains one cell instead of splitting the table in two.

A table-shaped region that does not clear that threshold stays as ordinary reading-order text. It is also named in the loss report. Nothing is silently promoted into a grid, and the same table text is not duplicated again as paragraphs. If the grid itself is the work you need to review, PDF to Excel shows candidate regions as page images and provides an editable cell matrix before writing a workbook.

Links come from annotations, not blue pixels

A real PDF link is an annotation placed over a rectangle. The converter matches that rectangle to overlapping text and writes a Markdown link only for an absolute HTTP, HTTPS or email destination. Script URLs, data URLs and other unsafe schemes are left as visible text and counted in the report. A word that merely happens to be blue and underlined is not called a link because appearance alone cannot prove where it goes.

This distinction matters because a PDF can contain an annotation whose destination is not printed anywhere on the page. Copying every annotation into Markdown without checking its scheme would turn an inert document into an active link. Keeping the label while refusing the destination preserves what the reader saw without carrying unsafe behavior forward.

Images are rendered crops, and the report says so

PDF image objects are often reused, clipped, masked, recolored or combined with vector marks. Copying their compressed bytes can therefore produce something different from the picture a reader actually saw. This tool renders the final page and crops the painted image region instead. The resulting PNG is deterministic for the same decoded pixels and receives a stable name under assets/. The loss report calls it a rendered crop rather than pretending it was the untouched source photograph.

Vector artwork has no general Markdown equivalent. It is counted as review evidence and is not invented as an icon name or alt text. Image alt text is deliberately generic when the PDF exposes no description. The complete ZIP contains document.md, every referenced image and conversion-report.json; the report has no timestamp, so identical input and engine behavior produce an identical truth artifact.

A scanned page never becomes imaginary text

A scan is a picture of words, not stored characters. A page with no text layer is kept as a local page image and receives the explicit finding “OCR was not run.” Sparse selectable text is not enough to call a page a scan: a title page remains text. A visual fallback is added only when that sparse text sits over page-sized raster evidence, as a stamped page number on a scan might. This prevents both an empty Markdown file and a readable title being misrepresented.

Optical character recognition is a different process that guesses letters from shapes. No OCR model is downloaded here, no remote recognition service receives the pages, and no guessed words are mixed with real PDF text. If selectable text matters, run OCR in a tool that names its language model and confidence, then return with the searchable PDF.

Use the raw source and the loss report together

The Raw tab shows a bounded excerpt of very long Markdown output. A note gives the exact cut and total; Copy raw, the Markdown download and the ZIP always use the complete result. Styled applies the same display-only boundary and is a read-only review surface whose local image references are replaced with in-memory blob URLs; relative paths are never allowed to trigger a request to this site or another host. Copy raw is appropriate for a code editor. Copy styled supplies both HTML and plain Markdown for applications that understand rich clipboard content for text-only results. When the document has images, styled copy is withheld because temporary browser image URLs would break after paste. The ZIP is the portable result because a standalone Markdown file cannot carry its neighboring asset folder.

For a text-only result that needs editorial changes, take the Raw output to the Markdown editor. Keep the loss report beside it: the editor can revise inferred headings and reading order, but it cannot recover structure the PDF never exposed or replace the converter's evidence about those inferences.

The opposite direction is a separate job. Markdown to PDF starts with known headings, lists and tables and lays them onto paper. It does not face the inference problem described above. For a quick extraction where formatting is irrelevant, PDF to text keeps the same column-aware reading order without adding Markdown syntax. For book-length material intended to reflow on an e-reader, PDF to EPUB removes repeated running heads and rebuilds chapters.

Your document stays in the browser

The PDF reader, structure inference, page renderer, PNG encoder and verified ZIP writer all run in this tab. The heavy PDF and archive code is fetched only after you choose a file. The document itself is never uploaded, and no preview image path is allowed to become a network request. You can verify that claim directly in the browser Network panel while the pages are read and the archive is assembled.