The local PDF to Markdown converter reads a PDF visually, one page at a time. It identifies text, tables, and other content, then puts the results together as Markdown.
Processing happens in your browser, and your PDF is not uploaded to an application API. The tool downloads the models and runtime files it needs and requires a browser with WebGPU support.
How it works
1. Turn each PDF page into an image
The tool first renders each page as an image. This gives scanned PDFs and digitally created PDFs the same starting point: the visible page. It does not bypass recognition by extracting an existing text layer.
2. Identify the page layout
The tool uses PaddlePaddle's PP-DocLayoutV3 layout analysis model in ONNX format to locate text, tables, pictures, and other page elements and determine their reading order. This helps the tool choose how to process each region and assemble the results in the right sequence.
3. Process each region by content type
Each type of content is handled differently:
- Headings: PP-DocLayoutV3 identifies title regions, and the PP-OCRv6 models described below recognize their text. The results are formatted as Markdown headings. You may need to adjust the heading levels to match the document's structure.
- Text: PPOCRv6-small recognizes the words in the region. You may need to adjust the text alignment and spacing to match the document's structure.
- Tables: SLANet_plus recognizes the table structure. The pipeline combines this structure with text detected and recognized by PP-OCRv6_small_det and PP-OCRv6_small_rec, placing the text into the corresponding cells.
- Pictures and charts: These regions are cropped from the page and kept as images.
- Formulas: A quantized TexTeller3 ONNX model converts formulas into editable LaTeX.
- Headers and footers: These are omitted from the final Markdown to keep the main content uncluttered.
4. Assemble the Markdown
The tool combines the recognized text, tables, and image references into a Markdown document. The result provides a starting point for editing, while image crops preserve content that is not converted into text.
What still needs review
Tables that continue across pages
The tool can merge tables that appear to continue onto the next page, but it does so conservatively. Repeated headers, changes in column widths, or interrupted layouts can make a continuation difficult to identify. Review the result and use the undo toggle if a merge is incorrect.
Tables with images inside cells
Table recognition focuses on text and cell structure. Pictures embedded inside cells may not retain their exact placement or relationship to the surrounding text. These tables can require manual editing.
Difficult scans and complex layouts
Blurry text, unusual fonts, dense pages, and complicated reading order can lead to recognition or layout errors. Check important text, numbers, and table structure against the original PDF before using the output.
Conclusion
The conversion pipeline is still in beta, so results may vary depending on your PDF. It generally works best with digitally created PDFs; scanned documents may require more manual corrections.