SmolDocling OCR App

The app provides advanced OCR capabilities for various document elements. Supported processing tasks include:

• General document conversion
• Table extraction: Outputs in OTSL format.
• Code extraction
• Formula conversion: Converts formulas to LaTeX.
• Chart data extraction
• Section header extraction

Results are provided in two structured formats: DocTags and Markdown.

To install the SmolDocling OCR App, clone the repository and install the required dependencies using either uv (recommended) or pip.

Prerequisites:

• Python 3.12+
• A Hugging Face account with an API token for model access.

# 1. Clone the repository
git clone https://github.com/AIAnytime/SmolDocling-OCR-App
cd smoldocling

# 2. Install dependencies (using uv)
uv pip install -r requirements.txt

# OR using pip
pip install -r requirements.txt

The application is built with Streamlit. To launch the interface, run the main.py file using the Streamlit CLI. Once running, access the app via your web browser (usually at http://localhost:8501).

streamlit run main.py

Once the Streamlit app is running in your browser, follow these steps to process documents:

1. Configure Mode: Use the sidebar to select between single or multiple image upload modes.
2. Select Task: Choose the specific processing task type (e.g., Table extraction, Code extraction, Formula conversion, etc.).
3. Upload: Upload your document image(s).
4. Process: Click the "Process Image(s)" button to begin the OCR.
5. Review/Download: View the results in DocTags and Markdown formats, and download them for use.

The application requires a Hugging Face token to access the underlying models. Create a .env file in the project root directory and add your token using the HF_TOKEN key.

HF_TOKEN=your_huggingface_token_here

The process_single_image and process_pdf functions accept a prompt_text parameter that determines the specific OCR task. The application supports the following task types:

• General OCR: "Convert this page to docling." or "Convert this PDF to docling."
• Table Extraction: "Convert this table to OTSL."
• Code Extraction: "Convert code to text."
• Formula Extraction: "Convert formula to latex."
• Chart Extraction: "Convert chart to OTSL."
• Header Extraction: "Extract all section header elements on the page."

Call check_dependencies to verify if the necessary libraries for the OCR engine and document processing are installed. It checks for the availability of transformers, huggingface_hub, and docling-core.

Returns:

• missing (list of str): A list of package names that are missing. If the list is empty, all dependencies are met.

missing = check_dependencies()
if missing:
    print(f"Please install: {', '.join(missing)}")

Use process_single_image to perform OCR on a single PIL Image object. This function authenticates with Hugging Face (using HF_TOKEN from environment variables), loads the ds4sd/SmolDocling-256M-preview model and processor, and generates both DocTags and Markdown content.

Parameters:

• image: A PIL.Image.Image object.
• prompt_text: A string defining the task (e.g., "Convert this page to docling."). Defaults to "Convert this page to docling.".

Returns:

• doctags (str): The extracted DocTags string.
• md_content (str): The converted Markdown content.
• processing_time (float): The time taken to process the image in seconds.

from PIL import Image

# Assuming image is a loaded PIL Image
doctags, md_content, processing_time = process_single_image(image, prompt_text="Convert this page to docling.")
print(f"Markdown: {md_content}")