pdfsmith 0.2.0

PDF to Markdown conversion with multiple backend support

pdfsmith

PDF to Markdown conversion with multiple backend support

A unified interface to 19+ PDF parsing libraries including frontier LLMs. Pick the right tool for the job, or let pdfsmith choose for you.

Why pdfsmith?

• One API, many backends - Switch between parsers without changing your code
• Auto-selection - Automatically uses the best available parser
• Lightweight core - Install only the backends you need
• Battle-tested - Wrappers refined through extensive benchmarking

Installation

# Core package (no backends)
pip install pdfsmith

# With lightweight backends
pip install pdfsmith[light]

# Recommended stack (good balance of quality and speed)
pip install pdfsmith[recommended]

# All open-source backends
pip install pdfsmith[all]

# Frontier LLMs (GPT, Claude, Gemini)
pip install pdfsmith[frontier]

# Commercial cloud APIs
pip install pdfsmith[commercial]

# Specific backend
pip install pdfsmith[docling]

Quick Start

from pdfsmith import parse

# Auto-select best available backend
markdown = parse("document.pdf")

# Use a specific backend
markdown = parse("document.pdf", backend="docling")

# Check available backends
from pdfsmith import available_backends
for backend in available_backends():
    print(f"{backend.name}: {backend.description}")

CLI Usage

# Parse PDF to stdout
pdfsmith parse document.pdf

# Parse to file
pdfsmith parse document.pdf -o output.md

# Use specific backend
pdfsmith parse document.pdf -b docling

# List available backends
pdfsmith backends

Available Backends

Open Source

Backend	Weight	Best For
docling	heavy	Highest quality, complex documents
marker	heavy	Academic papers, LaTeX content
pymupdf4llm	medium	Good balance of speed and quality
kreuzberg	medium	Fast extraction with OCR
unstructured	medium	Versatile document processing
pdfplumber	light	Tables and structured data
pymupdf	light	Fast general-purpose extraction
pypdf	light	Lightweight, pure Python
pdfminer	light	Mature, handles encodings well
pypdfium2	light	Chrome's PDF engine
extractous	medium	Rust-based extraction

Commercial Cloud APIs

Backend	Provider	Cost	Best For
aws_textract	AWS	$1.50/1k pages	High-accuracy OCR
azure_document_intelligence	Azure	$1.50/1k pages	Enterprise documents
google_document_ai	Google Cloud	$1.50/1k pages	Multi-language support
databricks	Databricks	~$3/1k pages	SQL-based workflows
llamaparse	LlamaIndex	$0.003/page	Cost-effective API

Frontier LLMs

Backend	Model	Cost	Best For
anthropic	Claude Sonnet 4.5	~$0.04/page	High accuracy
openai	GPT-4o	~$0.02/page	General purpose
gemini	Gemini 2.0 Flash	~$0.001/page	Budget LLM option

Note: Frontier LLM backends require API keys set via environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY).

Choosing a Backend

• Best quality: anthropic or openai - Frontier LLM accuracy (highest cost)
• Best value: llamaparse - Near-LLM quality at 10x lower cost
• Structure preservation: docling - Deep learning, GPU recommended
• Academic papers: marker - Optimized for LaTeX/equations
• Tables: pdfplumber - Excellent table detection
• Speed: pymupdf or kreuzberg - Fast extraction
• Minimal dependencies: pypdf - Pure Python, no binaries
• Budget LLM: gemini with gemini-2.0-flash - Very low cost LLM option

System Dependencies

Some backends require system packages for OCR functionality:

Tesseract OCR (for kreuzberg and unstructured with OCR):

# Ubuntu/Debian
sudo apt-get install tesseract-ocr

# macOS
brew install tesseract

# Windows
# Download from https://github.com/UB-Mannheim/tesseract/wiki

Without tesseract, these backends will still work for text-based PDFs but cannot extract text from scanned/image PDFs.

Async Support

from pdfsmith import parse_async

# Async parsing (uses backend's native async if available)
markdown = await parse_async("document.pdf")

Benchmarks

pdfsmith's backend wrappers were developed and refined through the pdf-bench benchmarking project, which evaluates parser performance across diverse document types.

License

MIT

Contributing

Contributions welcome! Please read our contributing guidelines before submitting PRs.