Typstify 🚀

Typstify is an AI-powered agentic application that converts PDF documents into high-quality, editable Typst templates.

Using LangGraph to manage an iterative refinement loop and Advanced LLMs (like Gemini, GPT-4o, or Claude) for vision and code generation, Typstify doesn't just "guess" the layout—it verifies it by compiling the code and visually comparing the result against the original.

Typstify operates on a Generate -> Compile -> Compare lifecycle:

1. Generate Node: Analyzes the original PDF images and generates Typst code. If previous attempts failed, it receives the error logs or visual feedback to perform surgical fixes.
2. Compile Node: Uses the local typst CLI to compile the generated code into a PDF.
3. Compare Node:
   • Calculates a Visual Similarity Score (SSIM) between the original and the new PDF.
   • If the match is below 95%, it sends both versions back to the LLM to identify layout, font, or spacing discrepancies.
4. Routing: The graph automatically loops back to the generation step with specific feedback until it reaches a near-perfect match or hits the iteration limit.

typstify/
├── app/                # Core logic package
│   ├── nodes/          # Individual graph nodes (generate, compile, compare, etc.)
│   ├── graph.py        # LangGraph workflow orchestration
│   ├── state.py        # Agent state definition
│   └── utils.py        # Helper functions (PDF processing, SSIM calculation)
├── inputs/             # Place your source PDF files here
├── output/             # Final .typ templates and verified PDFs
├── main.py             # App entry point (LLM injection happens here)
└── requirements.txt    # Python dependencies

The app must have access to the typst binary to verify the code.

• Ubuntu/Debian: sudo snap install typst
• macOS: brew install typst
• Manual: Download from Typst Releases.

pdf2image requires Poppler:

• Linux: sudo apt-get install poppler-utils
• macOS: brew install poppler

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

1. Set up your Environment: Create a .env file in the root directory and add the API key for your preferred provider:

   # Example for Gemini (Default)
   GOOGLE_API_KEY=your_gemini_api_key
   
   # Or for OpenAI
   # OPENAI_API_KEY=your_openai_api_key

2. Add your PDF: Place the PDF you want to convert into the inputs/ folder.

3. Run the Agent:

   python main.py inputs/your_document.pdf

4. Get your Results: Check the output/ folder for the final editable source and the verified PDF.

Typstify uses the LangChain/LangGraph LLM interface, meaning you can use any model that supports vision and is compatible with LangChain (Gemini, OpenAI, Anthropic, etc.).

By default, it is configured for Gemini in main.py, but you can easily swap it:

# In main.py: Switch from Gemini to OpenAI
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o", temperature=0)
app = create_app(llm)

Note: Ensure your chosen model supports Vision (multi-modal input), as it needs to "see" the PDF pages for accurate layout replication.

• Implement Prompt Caching.