OCR Orchestrator Service

Overview

The OCR Orchestrator Service is a versatile, FastAPI-based solution designed to serve various OCR (Optical Character Recognition) usecases. It provides a unified interface for processing different types of document images, including cheques, payslips, bank statements, passports, and utility bills.

Key Features

• Modular architecture supporting multiple OCR processors
• Configurable tasks for different document types and extraction needs
• Support for both online (real-time) and offline processing
• Integration with various AI models and external APIs
• Extensible design for easy addition of new processors and tasks

Setup

The project uses PDM to manage dependencies. To install the project:

pdm install --no-self --no-lock

Quickstart

Run the script to start the Uvicorn server. Gradio UI will be available at http://localhost:8000/ocrorchestrator/ui. FastAPI Swagger will be available at http://localhost:8000/docs

localrun=true bash entrypoint.sh

Folder Structure

src/ocrorchestrator/
├── __init__.py
├── config/
│   └── app_config.py
├── datamodels/
│   └── api_io.py
├── managers/
│   └── processor.py
├── processors/
│   ├── __init__.py
│   ├── api.py
│   ├── base.py
│   ├── factory.py
│   ├── gradio.py
│   ├── llm.py
│   └── pytorch.py
├── repos/
│   ├── __init__.py
│   ├── base.py
│   ├── factory.py
│   ├── gcs.py
│   └── local.py
├── routers.py
├── deps.py
├── main.py
├── ui.py
└── utils/
    ├── constants.py
    ├── img.py
    ├── logging.py
    ├── mixins.py
    └── misc.py

Technical Details

Core Components

1. FastAPI Service: The main entry point, handling HTTP requests and responses.
2. Processor Manager: Manages the lifecycle of processors and routes requests to the appropriate processor.
3. Processor Factory: Creates processor instances based on configuration.
4. App Config: Loads and manages the application configuration from YAML files.
5. Repos: Handles data storage and retrieval (supports both GCS and local storage).

Processors

The service uses a modular processor architecture. Current processor types include:

1. LLMProcessor: Utilizes large language models (e.g., GPT) for text extraction and analysis.
2. DocumentValidationProcessor: Uses PyTorch models for document validation tasks.
3. ApiProcessor: Integrates with external OCR APIs.
4. GradioProcessor: Leverages Gradio-based models for specific tasks.

Adding New Processors

To add a new processor:

1. Create a new file in the processors/ directory (e.g., new_processor.py).
2. Define your processor class, inheriting from BaseProcessor:

from .base import BaseProcessor

class NewProcessor(BaseProcessor):
    def __init__(self, task_config, general_config, repo):
        super().__init__(task_config, general_config, repo)
        # Initialize any specific attributes

    def _setup(self):
        # Implement any setup logic (e.g., loading models)

    def _process(self, req: OCRRequest) -> Dict[str, Any]:
        # Implement the main processing logic
        # Return the extracted information as a dictionary

3. Update processors/__init__.py to import your new processor.
4. Update processors/factory.py to include the new processor in the creation logic.

Using Mixins

The project uses mixins to share common functionality across processors. Key mixins include:

• TorchClassifierMixin: Provides methods for loading and using PyTorch classifiers.
• VertexAILangchainMixin: Offers methods for interacting with VertexAI and Langchain.

To use a mixin in a new processor:

from ..utils.mixins import TorchClassifierMixin
from .base import BaseProcessor

class NewClassifierProcessor(BaseProcessor, TorchClassifierMixin):
    def __init__(self, task_config, general_config, repo):
        super().__init__(task_config, general_config, repo)
        # Mixin-specific initialization

    def _setup(self):
        # Use mixin methods for setup
        self.load_model(self.task_config.model, self.task_config.checkpoint)

    def _process(self, req: OCRRequest) -> Dict[str, Any]:
        # Use mixin methods in processing
        result = self.predict(req.image)
        return {"classification": result}

Configuration

The service uses a YAML configuration file to define processors and their parameters. To add a new processor to the configuration:

categories:
  new_category:
    new_task:
      processor: NewProcessor
      model: "path_to_model"
      prompt_template: "path_to_prompt.txt"
      fields:
        - field1
        - field2
      params:
        - param1: value1
        - param2: value2

How to Use the Service

1. Setup

Ensure you have the necessary dependencies installed and the service is running. The IT department can provide you with the service endpoint.

2. Making a Request

To use the OCR service, send a POST request to the /predict endpoint with the following JSON structure:

{
  "image": "base64_encoded_image_string",
  "category": "document_type",
  "task": "specific_task",
  "fields": ["field1", "field2", "field3"]
}

• image: The document image encoded as a base64 string
• category: The type of document (e.g., "cheque", "payslip", "passport")
• task: The specific task to perform (e.g., "extract_details", "validate")
• fields: List of fields to extract (if applicable)

3. Interpreting the Response

The service will respond with a JSON object containing:

{
  "status": "OK",
  "status_code": 200,
  "message": {
    "field1": "extracted_value1",
    "field2": "extracted_value2",
    "field3": "extracted_value3"
  }
}

Supported Integrations/Processors

1. LLM Processor: Uses large language models for text extraction and analysis
2. Document Validation Processor: Employs PyTorch models for document validation
3. API Processor: Integrates with external OCR APIs
4. Gradio Processor: Utilizes Gradio-based models for certain tasks