PDF to Markdown Converter

A Python tool that converts PDF files to Markdown format using various AI models. The tool supports multiple AI providers and models.

Features

• Support for multiple AI providers:
  • Anthropic (Claude)
  • Google (Gemini)
  • OpenAI
  • Mistral
  • Ollama
  • Unstructured.io
• Configurable model selection for each provider
• Structured output with proper markdown formatting
• Environment variable configuration for API keys

Installation

1. Clone the repository:

git clone https://github.com/cotrane/pdf2md.git
cd pdf2md

2. Create and activate a virtual environment:

uv venv
source .venv/bin/activate  # On Unix/macOS
# or
.venv\Scripts\activate  # On Windows

3. Install dependencies:

uv pip install .

Configuration

Environment Variables

Create a .env file similar to the .env.tmpl file and add your API keys.

The supported API keys are:

• GOOGLE_API_KEY
• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• MISTRAL_API_KEY
• UNSTRUCTURED_API_KEY

In order to test AWS Textract you require an AWS account and credentials which can then be added to the .env file using

• AWS_ACCESS_KEY_ID
• AWS_SECRET_ACCESS_KEY
• AWS_REGION

Usage

Basic Usage

uv run src/run.py --input input.pdf --parser anthropic

The output file will be created in folder output and be called <input_file_name>_<parser>_<model>.md.

Available Parsers

• anthropic: Uses Claude 3.5 Sonnet
• googleai: Uses Google's Gemini Pro
• openai: Uses GPT-4 Turbo
• mistral: Uses Mistral Large
• ollama: Uses Ollama models
• textract: Uses Textract for text extraction
• unstructuredio: Uses Unstructured.io API

Model Selection

Each parser supports different models. Use the --model option to specify a model:

uv run src/run.py --input input.pdf --parser anthropic --model claude-3-sonnet-20240229

Available models per parser:

• Anthropic:
  • claude-3-7-sonnet-20250219 (default)
  • claude-3-5-sonnet-20241022
• Google AI:
  • gemini-1.5-flash
  • gemini-2.0-flash (default)
  • gemini-2.0-flash-thinking-exp-01-21
  • gemini-2.5-pro-exp-03-25
• OpenAI:
  • gpt-4o (default)
  • gpt-4o-mini
  • gpt-4.5-preview
  • o1
• Mistral: mistral-ocr-latest
• Ollama: Any model available in your Ollama installation
• Textract: No model selection
• Unstructured.io:
  • gpt-4o (default)
  • claude-3-5-sonnet-20241022
  • gemini-2.0-flash-001
  • hi-res
  • fast

Evaluating Results

The tool includes a utility to evaluate the similarity between markdown files. This script creates a similarity heatmap between the various output files and serves as a rough measure of how accurate the created markdown file is once one is checked manually.

uv run src/evaluate.py -f <filename>

The evaluation provides several metrics:

• Cosine Similarity: TF-IDF based similarity score (0-1)
• Word Overlap Ratio: Ratio of common words to total unique words (0-1)
• Levenshtein Ratio: Normalized similarity score based on Levenshtein distance (0-1)
• Word Statistics: Counts of words, common words, and unique words in each file

In order to evaluate only the OCR part of the output file, we can remove all markdown notation by running the script as follows:

uv run src/evaluate.py -f <filename> -m

Development

Code Style

The project uses:

• Pylint for code linting (configured in pyproject.toml)
• Black for code formatting
• MyPy for type checking

Development Dependencies

Install development dependencies:

uv pip install ".[dev]"

Evaluation Dependencies

Install evaluation dependencies:

uv pip install ".[eval]"

License

This project is licensed under the MIT License - see the LICENSE file for details.