What conversion actually does
A PDF is not text. It is drawing instructions: glyphs at coordinates, sometimes sitting on a text layer, sometimes with none at all. Every tool in this space starts by reading the text layer, then guessing structure from geometry and font metadata. Large bold type high on the page becomes a heading. Indented lines with bullet glyphs become list items. Whatever remains gets stitched back into paragraphs, in reading order.
That sets an honest ceiling for the output. Headings, paragraphs, and lists survive a PDF to Markdown conversion well. Tables mostly do not, because a table in a PDF is text arranged in a grid with lines drawn over it. Converters that recover tables reliably are running ML models on top of extraction, and that difference is exactly where the three routes split.
Route 1: a local library in your own script
PyMuPDF ships pymupdf4llm, a PDF to Markdown module built for RAG ingestion, and Marker is a popular open-source alternative that also handles DOCX, PPTX, and EPUB. You install them, write twenty lines of Python, and point them at a directory.
This is the route for recurring work. Hundreds of files a week, a CI job, a corpus you rebuild nightly. Scripting wins because the pipeline is versioned along with your code and nothing ever leaves the machine. The costs sit elsewhere: a Python environment to maintain, ML-based tools that want real hardware, and per-document fiddling. Stray headers and footers on every page are the classic annoyance; cleaning them is your code, your edge cases.
I've spent more evenings tuning header-detection heuristics for one weird document template than I care to admit. If your documents arrive from five different sources, budget for that.
Route 2: a hosted extraction API
Adobe's PDF Services has an official PDF to Markdown endpoint aimed at LLM ingestion, and services like Nanonets pitch RAG-ready output with tables and code blocks preserved. You trade files leaving your network for quality and scale: these pipelines handle messy layouts, rotated scans, and tables better than anything you can assemble in an afternoon, and there is a vendor to complain to when output regresses.
Two things to check before committing a batch. Whether their pricing model fits your volume, since metered per-document costs add up in ways flat-rate tools do not. And whether your legal or security posture allows the upload at all. For NDA material, medical records, or anything regulated, this route may simply be closed.
Route 3: convert PDF to Markdown in the browser
Browser tools need no install and no account. The strongest ones, including AmiPDF's PDF to Markdown tool, run the whole extraction locally: the file is opened by your browser, parsed with a JavaScript PDF engine, and never uploaded. You drop a document in, pick Markdown or plain text output, click the pages you want or type a range like 1-5,8, and preview the result with a character count before downloading anything.
The limits of browser PDF to Markdown extraction are worth knowing up front. It reads the text layer, so scanned PDFs need OCR first. AmiPDF's OCR also runs locally, handles up to 50 pages or 50 MB per file, and adds an invisible text layer that the Markdown step can then read. Structure detection covers headings, paragraphs, and lists; tables come through as text, not as Markdown tables. Encrypted files are rejected until you decrypt them. For raw words without any formatting, the same tool doubles as a plain-text extractor.
For one-off documents, confidential material, or quick chapter extraction from an e-book, this route is hard to beat on setup cost: the price is zero and the privacy story is verifiable. Open your browser's network tab while converting and watch nothing leave.
Which route for which job
Recurring pipelines with stable document templates belong in Route 1, where the heuristics live in version control. High-volume or layout-heavy work where tables must survive belongs in Route 2, if uploads are permitted. Everything else — single documents, private documents, chapters, quick text rescue — is Route 3 territory, and starting there costs you nothing but ten minutes.
FAQ
Does PDF to Markdown conversion work on scanned PDFs?
Not directly. A scan is an image with no text layer, so extraction has nothing to read. Run OCR first to add a text layer, then convert. Browser-based OCR exists for this, with the 50-page per file ceiling noted above.
Can I convert only part of a document?
Yes, in the browser route. Page pickers let you click individual pages in a grid or type ranges like 1-5,8. That is the natural way to pull one chapter out of a long e-book without converting 300 pages.
How can I verify a browser tool keeps files local?
Open the developer tools network tab, run a conversion, and watch the requests. A local tool shows no upload during processing. Claims on a landing page are marketing; the network tab is evidence.
Markdown or plain text — which do I need?
Markdown when structure matters: docs sites, wikis, notes that keep headings. Plain text when you only need the words for search, quoting, or a word count. When in doubt, extract Markdown; stripping formatting later is trivial, recovering it is not.
Top comments (0)