Researchers feeding papers into AI assistants run into the same problem: pasting text from a research PDF into a chat window loses equations, scrambles the order of two-column text, and collapses tables into unreadable rows. The PDF to Structured Data tool at Purplelink converts a paper PDF into Markdown that keeps equations as LaTeX, tables intact, and reading order correct.
Steps
- Go to purplelink.llc/tools/pdf-structure/. The tool accepts PDF files up to 20 MB. Most journal and arXiv papers fall well within that limit.
- Click Choose file and select your PDF.
[SCREENSHOT: file upload area with a paper PDF selected, showing the filename]
- Click Extract. Processing runs on the server and takes 5 to 30 seconds for most papers. A progress indicator appears while it works.
- When extraction finishes, two tabs appear: Markdown and JSON. Select Markdown for pasting into an AI assistant.
[SCREENSHOT: Markdown output tab showing a section of text with a preserved LaTeX equation]
- Click Copy to copy the full Markdown to your clipboard. Use Download to save a .md file if you plan to reuse the content.
- Paste into Claude, ChatGPT, or whichever assistant you use. Equations in the output are LaTeX, for example
$$E = mc^2$$or$$\int_a^b f(x)\,dx$$, which these assistants read and interpret correctly.
What's happening under the hood
The tool runs reading-order detection on the PDF so that multi-column text flows left to right in logical reading sequence rather than extracting column one then column two as separate blocks. Tables are identified structurally and rendered as Markdown table syntax. Math formulas are recognized as equation regions and emitted as LaTeX strings rather than garbled Unicode character sequences. Figure captions are detected and kept adjacent to their position in the document. The PDF is processed entirely in memory and deleted when the server response is sent. Nothing is stored and no data is sent to a third-party service. The tool requires a PDF with selectable text; scanned page images will not extract correctly.
Q&A
What if the paper has a two-column layout?
Reading-order detection handles this. The text flows in logical reading order across both columns, so sentences will not alternate between columns.
When should I use JSON output instead of Markdown?
The JSON output includes element type, content, and bounding-box coordinates for each block. Use it when building a retrieval pipeline or indexing system that needs element-level metadata rather than a flat text stream.
Is this suitable for preprints and arXiv papers?
Yes. Any PDF with selectable text works, including arXiv preprints, published journal articles, and conference papers.
Ready to try it: purplelink.llc/tools/pdf-structure/