Pdf2md

PDF -> Markdown (pdf2md)

Extracts PDF content into Markdown: text as-is + image/diagram meaning via a Vision Language Model.

When to Use

• User asks to "extract from PDF", "convert PDF to text/Markdown", "transcribe slides"
• Presentation decks (text + graphics + diagrams)
• Conference talks, workshops, technical reports
• PDFs where plain text extraction produces garbled output

When NOT to Use

• PDFs that are purely text (use pdfplumber directly — faster, no VL needed)
• One-page forms or invoices (use pdfplumber tables)
• Filling PDF forms (use well known pdf skill instead)
• Scanned documents without selectable text (this skill works, but OCR-only tools may be faster)

Dependencies

Install once per environment:

pip install pdfplumber pypdfium2 pillow requests

VL Model Setup

Default: LM Studio with qwen/qwen3-vl-8b loaded on localhost:1234.

Any OpenAI-compatible endpoint works — Ollama, vLLM, cloud providers. To switch:

python scripts/extract_pdfs.py --source "<path>" --model "llama3.2-vision" --api-url "http://localhost:11434/v1/chat/completions"

Workflow

Step 1: Verify VL model is reachable

python -c "import requests; r = requests.get('http://localhost:1234/v1/models'); print([m['id'] for m in r.json()['data']])"

If you don't see your VL model, load it in LM Studio before proceeding.

Step 2: Test on 1-2 PDFs first

python scripts/extract_pdfs.py --source "<path-to-pdfs>" --first 2

Check the output in _markdown/. If quality is poor (too much "water", missing text, or style descriptions leaking through), edit the prompt in references/vl-prompt.md and the VL_PROMPT variable in the script, then re-run with --force.

Step 3: Run on everything

python scripts/extract_pdfs.py --source "<path-to-pdfs>"

Already-processed PDFs are skipped automatically. Use --force to overwrite.

Step 4: If something fails

• VL call fails (3 retries exhausted): page is left with fallback text from pdfplumber, or marked [VL error]
• Rendering fails: page is skipped, counted as error
• Output file corrupted: delete the .md and re-run — the script will re-process it
• Model runs out of VRAM: reduce RENDER_SCALE in the script (1.5 or 1.0)

Script Arguments

Output Structure

_source/
  _markdown/
    subfolder/
      presentation1.md
      presentation2.md

Folder structure mirrors the source. Each .md file:

# Presentation Title

---
## Page 1 / 40

**[VL slide description]**

<content extracted by VL model>

---
## Page 2 / 40
...

VL Prompt Rules (hard-won lessons)

1. Never describe fonts, colors, margins, backgrounds — only content
2. Never write intros like "This slide shows..." — start with content
3. Never invent "Conclusions" or "Overall assessment" sections
4. Extract ALL text verbatim
5. For diagrams/charts: describe MEANING, not appearance
6. Concise, dense, no filler — every sentence must carry information
7. Do NOT limit output tokens — let the model decide

See references/vl-prompt.md for the full prompt template and tuning guide.

Performance

See Also

• VL Prompt (references/vl-prompt.md) — prompt template + iteration history + tuning tips
• Script (scripts/extract_pdfs.py) — the extraction tool

Evolution rule: If you are dissatisfied with the extraction quality or the workflow, identify what to change (prompt, script parameters, dependencies) and offer to update this skill. The prompt lives in references/vl-prompt.md; the extraction logic in scripts/extract_pdfs.py.