Textractor

Structure-aware markdown extraction from documents for LLM consumption.

Textractor parses PDF, DOCX, PPTX, HTML, and EPUB files into clean GitHub-flavored markdown, preserving document structure (headings, tables, code blocks, lists, images). It can split output into chunks that respect heading boundaries, making it ideal for feeding documents into LLMs.

Features

• Multi-format parsing -- PDF, DOCX, PPTX, HTML, EPUB
• Structure-aware chunking -- splits at heading boundaries, never mid-paragraph
• Image extraction -- saves images alongside markdown output
• Vision model integration (optional) -- OCR for scanned PDFs via LightOnOCR-2-1B, image captioning via Qwen3-VL-4B
• Post-processing -- header/footer removal, multi-page table merging, artifact cleanup (PDF)
• LLM-friendly output -- clean markdown to stdout or separate .md files per chunk with YAML frontmatter

Installation

# Core
uv pip install .

# With vision model support (OCR + image captioning)
uv pip install ".[vision]"

Requires Python 3.11+.

Usage

# Extract to stdout
textractor extract document.pdf

# Extract to directory with chunking
textractor extract document.pdf -o output/ -t 32000

# Force format detection
textractor extract file.txt -f html

# Enable vision features (OCR, image captioning)
textractor extract scanned.pdf --use-vision

# Verbose output
textractor extract document.docx -v

Options

Option	Description
-o, --output-dir DIR	Write output files to directory (default: stdout)
-t, --max-tokens N	Max tokens per chunk, 0 = no chunking (default: 0)
-f, --format FMT	Force input format instead of auto-detect
--use-vision	Enable OCR and image captioning (default: off)
-v, --verbose	Verbose progress on stderr
-V, --version	Show version

Output modes

• Single output (no -o or single chunk): markdown printed to stdout
• Chunked output (-o with -t): separate .md files per chunk + metadata.json

Supported formats

Format	Extensions	Library
PDF	.pdf	pymupdf4llm
DOCX	.docx	python-docx
PPTX	.pptx	python-pptx
HTML	.html, .htm	markdownify
EPUB	.epub	ebooklib

License

MIT