Text layers, reading order, and Markdown limits
The converter uses PDF.js in the browser to read a PDF's embedded text layer, then maps extracted text and basic structure to Markdown. A PDF is a visual page description, not a semantic document, so reading order can differ from the way a person sees columns, positioned labels, or tables. A scan made only of images has no usable text layer unless an OCR step is present; this tool must not imply that it can recover one reliably.
Four documents to test
- Heading and paragraph: a one-page text PDF with “Quarterly Report” as an 18pt title becomes
# Extracted followed by the extracted text in the verified fixture.
- Simple table: check row and column order in the output; positioned text may arrive as a sequence rather than a Markdown table, so edit and verify before publishing it.
- Two columns: visual left-to-right reading is not guaranteed by the PDF text layer. Compare the Markdown order with the source page and treat unexpected interleaving as a document limitation.
- Scan-only PDF: verify the empty or error result rather than inventing text. OCR requires a separate, evidenced capability and changes the privacy and accuracy contract.
Tables, links, and page boundaries
Markdown has a smaller vocabulary than PDF layout. A heading may be inferred from font size in a simple fixture, but a visually aligned table is still a sequence of positioned text unless the converter has an explicit table rule. Check links, list markers, hyphenation, repeated headers, and page breaks before feeding the result to an indexer or an AI pipeline. Keep the original PDF beside the Markdown for auditability.
Test the conversion as a document workflow
Start with a small text PDF, compare the output with the source, and only then process a longer document. Note page count, whether the source has selectable text, and any repeated header or footer. If the output will be committed to a repository, retain the source fixture and review the diff so a reordered paragraph or lost link is visible rather than silently accepted.
What “offline” does not mean
Offline processing reduces the tool's upload surface; it does not make an extracted document automatically accurate, accessible, or safe to publish. Browser memory, malformed PDFs, embedded fonts, and positioned glyphs can all affect results. A human should review important contracts, research, and policy documents.
Decide whether the output is fit for purpose
Plain paragraphs are a good candidate for direct copy; dense financial tables, footnotes, columns, and scanned pages are review-heavy. Mark uncertain sections in your downstream notes instead of presenting reconstructed Markdown as the original layout. For AI ingestion, preserve page separators and source filenames so a claim can be traced back to the PDF.
CommonMark syntax is only a representation: validate fenced code, links, tables, and escaped characters in the destination editor. A successful download confirms file creation, not semantic fidelity. Keep conversion settings with the output when repeatability matters. For a regulated or shared workflow, store a checksum or source identifier rather than relying on a filename alone.
FAQ
Will it work with scanned PDFs?
Not reliably. The converter reads the text layer that pdfjs extracts from the document. If your PDF is a photograph of a page with no embedded text (a typical scanned PDF), there is nothing for the tool to read — you will need to OCR the file first.
Does my PDF get uploaded anywhere?
No. The file is parsed by pdfjs-dist inside your browser and never leaves the tab. There is no upload endpoint, no temporary cloud storage, and the resulting Markdown stays local until you copy or download it.
How does it pick which lines are headings?
By font size. The largest distinct size in the document becomes a #, the next-largest ##, and the next ###. It is a heuristic — most reports and academic papers work well; magazines and heavily-designed PDFs may not.
Does it handle tables?
Yes, best-effort. Any block of consecutive lines with aligned columns becomes a GitHub-Flavored Markdown pipe table with a separator row. Complex layouts with merged cells or nested tables may still need a manual cleanup pass.
Can I include the images from the PDF in the Markdown?
Yes, opt-in. Toggle Include images and each embedded image is rasterized to a base64 PNG and inlined into the Markdown. We keep this off by default because it can balloon the file size of the output.
Why are there Send to Claude and Send to ChatGPT buttons?
Because most users converting PDFs to Markdown are doing it to paste the result into an LLM. The buttons copy the Markdown to your clipboard with a Summarize this document: prompt prepended, then open the chat in a new tab. We make no API call; your conversation is entirely with the chat provider.