Document Processing System

An automated Azure-based document processing pipeline that leverages Azure Document Intelligence and OpenAI LLM for intelligent document classification and analysis.

📋 Overview

This project implements a serverless document processing system that:

1. Ingests PDF documents from Azure Blob Storage or http post
2. Extracts content using Azure Document Intelligence (prebuilt-layout model)
3. Classifies documents using Azure OpenAI LLM (GPT-4)
4. Publishes results to Microsoft Farbic Lakehouse using Azure Event Hub

The system is built as an Azure Function that automatically triggers on new document uploads or http requets, enabling a fully automated, scalable processing workflow.

🏗️ Architecture

Components

• Azure Functions - Serverless compute for document processing orchestration
• Azure Blob Storage - Document ingestion and storage
• Azure Document Intelligence - OCR and document layout analysis
• Azure OpenAI - LLM-based document classification
• Azure Event Hub - Event streaming and results publishing

Workflow

Document Upload → Blob Trigger/http trigger → Document Intelligence → LLM Classification → Event Hub
     (Input)         (Function)       (Content Extraction)   (Document Type)     (Output)

📂 Project Structure

DocumentProcessingSystem/
├── function_app.py                 # Azure Function entry point with blob trigger
├── run_DocumentIntelligence.py     # Document Intelligence API integration
├── run_LLMClasscification.py       # Azure OpenAI classification logic
├── run_FabricEventHub.py           # Event Hub publisher
├── code_testing.ipynb              # Testing and development notebook
├── requirements.txt                # Python dependencies
├── host.json                       # Azure Functions configuration
├── local.settings.json             # Local environment settings
└── README.md                       # This file

🚀 Getting Started

Prerequisites

• Python 3.9+
• Azure Functions Core Tools
• Azure CLI
• An Azure subscription with:
  • Storage Account (Blob Storage)
  • Document Intelligence resource
  • Azure OpenAI resource
  • Event Hub namespace

Installation

1. Clone the repository

   git clone <repository-url>
   cd DocumentProcessingSystem

2. Create a Python virtual environment

   python -m venv .venv
   .venv\Scripts\activate  # On Windows

3. Install dependencies

   pip install -r requirements.txt

4. Configure environment variables

   Create or update local.settings.json with your Azure credentials:

   {
     "IsEncrypted": false,
     "Values": {
       "AzureWebJobsStorage": "DefaultEndpointsProtocol=https;...",
       "Eventhub_endpoint": "your-eventhub-namespace.servicebus.windows.net",
       "Eventhub_name": "your-eventhub-name",
       "docintelligenceendpoint": "https://your-region.api.cognitive.microsoft.com/",
       "docintelligencekey": "your-document-intelligence-key",
       "openai_endpoint": "https://your-resource.openai.azure.com/",
       "openai_key": "your-openai-key",
       "rgdocumentprocessinb772_STORAGE": "your-blob-storage-connection-string"
     }
   }

Local Development

1. Start the Azure Functions runtime

   func start

2. Upload a test document

   Upload a PDF to the document-processing-dropzone/Input/ blob container

3. Monitor execution

   Check the function logs in the terminal for processing status

📝 Configuration

Document Intelligence Settings

• Model: prebuilt-layout (for general document layout analysis)
• API Version: 2024-11-30
• Poll Interval: 2 seconds (configurable in function_app.py)
• Max Wait: 60 seconds

Document Classification Types

Supported document classifications (defined in run_LLMClasscification.py):

• Medical Aid / Medical Scheme Certificate
• Employee Tax Certificate
• Retirement Annuity Certificate
• Investment Income Certificate
• Medical Expenses
• Travel Log Book
• Other

LLM Configuration

• Model: gpt-4.1
• API Version: 2024-12-01-preview
• Provider: Azure OpenAI

📦 Dependencies

Package	Purpose
azure-functions	Azure Functions SDK
azure-storage-blob	Blob Storage integration
azure-eventhub	Event Hub integration (For Fabric Lakehouse storage)
azure-identity	Azure authentication
requests	HTTP client for Document Intelligence API
openai	Azure OpenAI SDK
cryptography	Encryption utilities

🔄 Processing Flow

1.1 Blob Trigger

• Monitors the document-processing-dropzone/Input/ container
• Automatically triggers on PDF upload

1.2 HTTP Trigger

• Monitors the http endpoint for the function app
• Post requests with APIKey to the endpoint will trigger the process e.g. http://localhost:7071/api/func_document_processing?code===
• Pdf file must form part of the binary body

2. Document Intelligence

• Converts PDF to base64 encoding
• Posts to Document Intelligence API for layout analysis
• Polls for completion (up to 60 seconds)
• Extracts structured content and metadata

3. LLM Classification

• Processes extracted text content
• Uses Azure OpenAI to classify document type
• Generates confidence scores and reasoning

4. Event Hub Publishing

• Packages results with document metadata
• Publishes to Event Hub for downstream processing
• Enables real-time data consumption and analytics

🧪 Testing

Use code_testing.ipynb for:

• Unit testing individual components
• Testing API endpoints
• Debugging extraction and classification logic
• Manual workflow validation

⚙️ Deployment

Deploy to Azure

1. Create Azure Functions resource

   az functionapp create --resource-group <rg-name> \
     --consumption-plan-location <region> \
     --runtime python --runtime-version 3.11 \
     --functions-version 4 \
     --name <function-app-name>

2. Deploy the function

   func azure functionapp publish <function-app-name>

3. Configure application settings

   az functionapp config appsettings set \
     --name <function-app-name> \
     --resource-group <rg-name> \
     --settings <setting-key>=<setting-value>

📊 Monitoring & Logging

• Azure Functions integrated logging
• Document Intelligence API response tracking
• Event Hub message publishing verification
• Application Insights integration (optional)

🛠️ Known Limitations & Future Enhancements

Current Limitations

• ✋ Embeddings for large documents not yet implemented
• ✋ No chunking strategy for documents exceeding 2 MB item limits

Planned Enhancements

• Implement document chunking for large files
• Add vector embeddings for semantic search
• Migrate to Document Intelligence Python SDK
• Implement dead-letter handling for failed documents
• Add comprehensive error tracking and alerting

🔐 Security Considerations

• Use Azure Key Vault for sensitive credentials (recommended)
• Enable managed identities for Azure service authentication
• Restrict blob container access with appropriate RBAC
• Validate input documents before processing
• Monitor and audit Event Hub consumers

📝 Environment Variables Reference

Stored in local.settings.json

Variable	Description
Eventhub_endpoint	Event Hub namespace endpoint
Eventhub_name	Event Hub instance name
docintelligenceendpoint	Document Intelligence API endpoint
docintelligencekey	Document Intelligence API key
openai_endpoint	Azure OpenAI API endpoint
openai_key	Azure OpenAI API key
AzureWebJobsStorage	Blob Storage connection string
rgdocumentprocessinb772_STORAGE	Blob Storage connection for function trigger

📞 Support

For issues or questions, please open an issue in the repository or contact the Andrew Schleiss.