Test lighton-ocr-1b model 🚀

A lightweight, local OCR pipeline designed to convert PDF documents into structured JSON data. This project leverages the LightOn-OCR-1B model running on Ollama for high-accuracy text extraction without data leaving your machine.

🌟 Key Features

• 100% Local Inference: Powered by Ollama, ensuring total data privacy.
• Optimized Model: Uses maternion/LightOnOCR-2, a vision-language model specialized in document text recognition.
• Modern Python Tooling: Powered by uv for lightning-fast dependency management and execution.

🛠️ Prerequisites

1. Ollama must be installed and running.
2. Pull the LightOnOCR model:

ollama pull maternion/LightOnOCR-2

3. uv installed on your system:

# Install uv via curl
curl -LsSf https://astral.sh/uv/install.sh | sh

🚀 Quick Start

Using uv, you don't need to manually manage virtual environments or install dependencies one by one.

1. Clone the repository:

git clone <your-repo-url>
cd test-lighton-ocr-1b

2. Run the script:

uv run main.py
uv run pdf.py

Note: uv will automatically handle dependencies (ollama, pypdfium2, requests, pillow) defined in the script or pyproject.toml.

📝 How it Works

The pipeline follows these steps:

1. Fetch: Downloads the target PDF directly from a URL (e.g., an arXiv research paper).
2. Render: Converts PDF pages into high-resolution images (200 DPI) using pypdfium2.
3. Process: Sends the image bytes to the LightOnOCR-2 vision model via Ollama's Chat API.
4. Structure: Aggregates the OCR results into a single, structured JSON object containing data for every page.

📊 Sample Output

[
    {
        "page": 1,
        "content": "Full text extracted from page 1..."
    },
    {
        "page": 2,
        "content": "Full text extracted from page 2..."
    }
]

⚠️ Important Notes

• Vision Resources: While LightOn-OCR-1B is highly efficient, ensure your system has sufficient VRAM/RAM for smooth image processing.
• Prompt Engineering: The current prompt requests JSON format. Depending on the model's output, you might occasionally need to strip Markdown code blocks (e.g., json ... ) from the response.

---