Extracting PDF Text for LLMs

This skill provides tools and guidance for extracting text from PDFs in formats suitable for language model consumption.

Quick Decision Guide

| PDF Type | Best Approach | Script | |----------|--------------|--------| | Simple text PDF | PyMuPDF | scripts/extract_pymupdf.py | | PDF with tables | pdfplumber | scripts/extract_pdfplumber.py | | Scanned/image PDF (local) | pytesseract | scripts/extract_with_ocr.py | | Complex layout, highest accuracy | Mistral OCR API | scripts/extract_mistral_ocr.py | | End-to-end RAG pipeline | marker-pdf | pip install marker-pdf |

Recommended Workflow

Try PyMuPDF first - fastest, handles most text-based PDFs well
If tables are mangled - switch to pdfplumber
If scanned/image-based - use Mistral OCR API (best accuracy) or local OCR (free but slower)

Local Extraction (No API Required)

PyMuPDF - Fast General Extraction

Best for: Text-heavy PDFs, speed-critical workflows, basic structure preservation.

uv run scripts/extract_pymupdf.py input.pdf output.md

The script outputs markdown with preserved headings and paragraphs. For LLM-optimized output, it uses pymupdf4llm which formats text for RAG systems.

pdfplumber - Table Extraction

Best for: PDFs with tables, financial documents, structured data.

uv run scripts/extract_pdfplumber.py input.pdf output.md

Tables are converted to markdown format. Note: pdfplumber works best on machine-generated PDFs, not scanned documents.

Local OCR - Scanned Documents

Best for: Scanned PDFs when API access is unavailable.

uv run scripts/extract_with_ocr.py input.pdf output.txt

Requires: pytesseract, pdf2image, and Tesseract installed (brew install tesseract on macOS).

API-Based Extraction

Mistral OCR API

Best for: Complex layouts, scanned documents, highest accuracy, multilingual content, math formulas.

Pricing: ~1000 pages per dollar (very cost-effective)

export MISTRAL_API_KEY="your-key"
uv run scripts/extract_mistral_ocr.py input.pdf output.md

Features:

Outputs clean markdown
Preserves document structure (headings, lists, tables)
Handles images, math equations, multilingual text
95%+ accuracy on complex documents

For detailed API options and other services, see references/api-services.md.

Output Format Recommendations

For LLM consumption, markdown is preferred:

Preserves semantic structure (headings become context boundaries)
Tables remain readable
Compatible with most RAG chunking strategies

For detailed comparisons of local tools, see references/local-tools.md.