ollama-ocr

To run the interactive web interface with drag-and-drop support and real-time processing, follow these steps:

1. Clone the repository and enter the directory.
2. Install the required dependencies.
3. Navigate to the src/ollama_ocr directory.
4. Run the Streamlit application.

git clone https://github.com/imanoop7/Ollama-OCR.git
cd Ollama-OCR
pip install -r requirements.txt
cd src/ollama_ocr
streamlit run app.py

Install the ollama-ocr package using pip to use it as a Python library.

pip install ollama-ocr

Before using the package, ensure Ollama is installed on your system and pull the desired vision models. Supported models include llama3.2-vision:11b, granite3.2-vision, moondream, and minicpm-v.

ollama pull llama3.2-vision:11b
ollama pull granite3.2-vision
ollama pull moondream
ollama pull minicpm-v

To use ollama-ocr in a Google Colab environment, install the package along with colab-xterm to manage the Ollama service via a terminal interface.

!pip install colab-xterm ollama-ocr

To use the library, import OCRProcessor from ollama_ocr and initialize it by specifying a model_name. This model name should correspond to a vision-capable model available in your Ollama instance (e.g., granite3.2-vision or llama3.2-vision:11b).

from ollama_ocr import OCRProcessor

# Create an instance
ocr = OCRProcessor(model_name='granite3.2-vision')

To use the OCR capabilities within an AutoGen agent workflow, install both ollama-ocr and autogen using pip.

! pip -q install ollama-ocr
! pip -q install autogen

To use the package, instantiate the OCRProcessor class. You can specify the model_name and a custom base_url if you are using a custom Ollama API endpoint.

from ollama_ocr import OCRProcessor

# Initialize with a specific model and custom API URL
ocr = OCRProcessor(model_name='llama3.2-vision:11b', base_url="http://host.docker.internal:11434/api/generate")

When calling process_image or process_batch, you can specify the format_type to control the structure of the extracted data. Supported formats include:

• markdown: Markdown string with headers and lists.
• text: Plain text string.
• json: JSON object.
• structured: Structured object.
• key_value: Dictionary of labeled information.
• table: Extracted tabular data.

You can wrap the OCRProcessor.process_image method inside a Python function to serve as a tool for an AutoGen AssistantAgent.

When initializing OCRProcessor, specify a vision-capable model (e.g., granite3.2-vision). The process_image method accepts image_path, format_type (e.g., "text"), and language (e.g., "eng").

from ollama_ocr import OCRProcessor

def doc_parser(file_path:str)->str:
    ocr = OCRProcessor(model_name='granite3.2-vision')
    result = ocr.process_image(
        image_path=file_path,
        format_type="text",
        language="eng",
    )
    return result

Use the process_image method to extract text from images or PDF files.

Parameters:

• image_path (str): Path to the image or PDF file.
• format_type (str): The desired output format. Supported options: markdown, text, json, structured, key_value.
• custom_prompt (str, optional): A specific instruction for the model to guide extraction.
• language (str, optional): The language of the text to be extracted.

from ollama_ocr import OCRProcessor

ocr = OCRProcessor(model_name='granite3.2-vision')

result = ocr.process_image(
    image_path="/content/pdf.pdf",
    format_type="markdown",
    custom_prompt="Extract all text, focusing on dates and names.",
    language="English"
)
print(result)