PDF to Markdown with OCR and Tables

PDF is a presentation format, while Markdown is a content format. A good conversion therefore does more than copy visible characters. It must recover reading order, headings, lists, links, and tables, then make uncertainty easy to review. Scanned PDFs add a second problem: their pages may contain only images, so optical character recognition (OCR) is required before useful Markdown can be produced.

1. Identify the PDF type

Start by checking whether text can be selected in a normal PDF viewer. This simple test separates three common cases:

• Text PDF: characters are embedded and can usually be extracted quickly.
• Scanned PDF: each page is an image and needs OCR.
• Mixed PDF: some pages contain text while others are scans, screenshots, or photographed forms.

Do not enable OCR automatically for every document. Native text is normally faster and more accurate. Use OCR for pages that have little extracted text or when the preview clearly misses visible content.

2. Choose evidence before conversion

Markdown cannot reproduce every part of a PDF layout. Preserve source evidence when the document contains contracts, financial tables, diagrams, multi-column pages, or material that will be quoted later.

• Keep page references so Markdown sections can be traced to page numbers.
• Keep page images for charts, screenshots, formulas, and difficult layouts.
• Limit the page range when only one chapter or appendix is needed.
• Select the OCR language that matches the document.

Page images increase the output size, so they are best treated as verification material rather than a replacement for extracted text.

3. Review tables as data, not decoration

Simple tables often convert cleanly. Complex tables can fail when they contain merged cells, nested headers, blank spacer columns, footnotes, or values positioned visually instead of structurally. Review the Markdown table against the original page before using it in documentation or an AI knowledge package.

| Region | Revenue | Change |
| --- | ---: | ---: |
| West | 2.4M | 8% |
| East | 2.1M | -2% |

Check that every row has the same number of columns, negative signs are preserved, units remain attached to the correct header, and OCR has not confused characters such as 0 and O. When a table is too wide for readable Markdown, export it as CSV and link to it from the Markdown document.

4. Treat OCR output as a draft

OCR confidence is a signal, not a guarantee. Small fonts, low contrast, handwriting, rotated pages, stamps, and unusual column layouts can all reduce accuracy. Review names, dates, totals, legal clauses, code, and measurements even when the overall confidence looks high.

A practical review order is:

1. Confirm the page order and heading hierarchy.
2. Compare the lowest-confidence pages with their page images.
3. Search for suspicious replacements such as broken words and repeated punctuation.
4. Verify tables, numbers, proper names, and quoted passages.
5. Edit the Markdown before exporting the final package.

5. Prepare the result for reuse

Clean Markdown is useful because it can be versioned, searched, edited, and split into knowledge chunks. Remove repeated headers and footers, repair heading levels, and keep source markers near claims that need evidence. For RAG or Skill exports, choose chunk boundaries based on headings when the PDF has a meaningful structure. Use size-based chunks only when headings are missing or unreliable.