🌉 ModelOp Bridge: Azure Copilot Governance Connector

Turn on the lights in your "Shadow AI" basement.

Welcome! You’ve likely arrived here because your organization just deployed Microsoft 365 Copilot or Azure OpenAI. It’s exciting! "Citizen Developers" in HR, Finance, and Legal are building bots faster than you can track them.

But here’s the nagging question: > Do you actually know what those bots are saying?

This tool is your Flashlight. It is a specialized ETL (Extract, Transform, Load) bridge designed to:

1. CONNECT to your Microsoft Azure Tenant to pull real chat history.
2. SIMULATE realistic traffic (if you don't have real users yet).
3. RED TEAM your AI by injecting adversarial attacks and PII leaks.
4. DELIVER clean, standardized data to ModelOp Center for governance.

📊 How It Works: The Logic Engine

When you feed data into ModelOp Center (whether real or simulated), it immediately passes through our OOTB Standardized Monitors. The llm_data_monitoring_diagram.png diagram shows exactly which "lights" turn on based on the data you generate with this tool.

📂 Repository Structure

What's inside the box?

.
├── azure_moc_connector.py     # 🧠 The Worker: Connects to Azure, Simulates AI, and Red Teams.
├── generate_demo_data.py      # 🎭 The Orchestrator: Automates "Phase 1 Lite" scenarios w/ resume logic.
├── config.yaml                # ⚙️ Configure simulation rates, Azure creds, and Red Team settings.
├── mock_expansion_data.json   # 🎨 Example JSON used by the AI to learn your corporate "Voice".
├── requirements.txt           # 📦 Python libraries (Pinned for Python 3.12).
├── llm_data_monitoring_diagram.png # 📊 Source image for the monitoring logic flow.
├── .cursor                    # 📍 Tracks progress for resuming interrupted runs.
└── generated_chats/           # 📤 Where your raw JSON logs appear.
    ├── baseline_data.json     #    -> The "Control" (Clean Data)
    └── comparator_data.json   #    -> The "Variable" (Risk Data)
└── phase_1_lite_demo/         # 💎 Pre-packaged "Story Mode" datasets.
    └── 02_Archived/           # 🗄️ Timestamped backups of previous runs.

☕ The "Coffee Break" Factor & Resilience

Running Local AI (Ollama) is heavy work.

• Speed: Expect ~1-3 minutes per conversation depending on your GPU.
• Resiliency: We know things crash. If your computer sleeps or the script is interrupted (Ctrl+C), don't panic.
  • The tool now saves Granular Checkpoints (.checkpoint_*.jsonl) for every single record generated.
  • When you run the script again, it will detect the crash and ask if you want to RESUME exactly where you left off.

⚠️ Important: Python Version Compatibility

• Target Version: This project is standardized on Python 3.12.
• Why? To support the latest Azure Function libraries and Spacy 3.8+ wheels.
• Legacy Note: Older versions of this tool used Python 3.8. Please upgrade your environment if you are returning to this repo.

🏁 Part 1: New User Setup Guide (Windows/VS Code)

Follow these steps exactly if you are setting this up for the first time.

1. Install Python 3.12 & Git

Open your PowerShell or Command Prompt and paste this command to install the correct versions.

winget install -e --id Python.Python.3.12
winget install -e --id Git.Git

Restart your computer after this finishes to ensure PATH updates.

2. Clone the Repo

1. Open VS Code.
2. Go to File > New Window.
3. On the "Welcome" screen, click "Clone Git Repository...".
4. Paste: https://github.com/s-neub/azure_moc_connector
5. Select a folder (e.g., Documents) to save it.
6. Crucial: When asked "Would you like to open the cloned repository?", click Open.

3. Create Virtual Environment

VS Code can build the environment for you using the requirements.txt file.

1. Open requirements.txt in the editor.
2. Open the Command Palette (Ctrl+Shift+P).
3. Type and select: Python: Create Environment...
4. Select Venv.
5. Select Python 3.12.x (The version you just installed).
6. Check the list: Ensure requirements.txt is selected (checked).
7. Click OK.

4. Final Setup Command

Once the environment is ready, open a New Terminal (Ctrl + Shift + ~). You should see (.venv) at the start of the line. Run this final command to download the grammar model:

python -m spacy download en_core_web_sm

🛠️ Part 2: Get the "Brains" (Ollama)

Skip this if you are strictly connecting to Real Azure (Phase 3), but we recommend it for the Red Teaming features!

1. Download It: Go to ollama.com and install it.

2. Get the Model: Open your terminal and type:

    ollama pull qwen2.5
   

⚙️ Part 3: Choose Your Adventure (Configuration)

🟢 Phase 1: The "Mock" - "Kick the Tires" (Fastest)

Goal: I want to see the dashboard light up NOW without waiting for data generation.

1. Run the Storyteller: Execute python generate_demo_data.py.
2. Open the Gift Shop: Navigate to the phase_1_lite_demo/ folder.
3. Upload: Drag and drop the pre-cooked JSON files into ModelOp.

🟡 Phase 2: Custom Simulation (Best for Demos)

Goal: Test specific scenarios (e.g., "What if my bot is rude?") and prove ModelOp catches bad guys.

1. Configure config.yaml:
   • Mode: Set use_real_azure: false.
   • Rates: Adjust rates (e.g., increase toxicity to 0.5).
2. Run the script: python azure_moc_connector.py.

🔴 Phase 3: The "Real World" (The Holy Grail)

Goal: Connect to your actual Microsoft 365 Copilot to audit real user interactions.

1. Configure config.yaml:
   • Connect: Set use_real_azure: true and fill in your azure credentials.
2. The Pipeline: Run python azure_moc_connector.py.
3. Upload: Upload the resulting audit log to ModelOp.

🔮 Roadmap & Future Work

We are actively evolving this tool from a "Local Utility" to an "Enterprise Service".

• Azure Function Deployment (Serverless):
  • Current State: Runs locally on a laptop.
  • Future State: The code is already Python 3.12 compatible. The next release will include function.json bindings to deploy this as a Timer Triggered Azure Function, allowing it to harvest logs nightly without user intervention.
• Direct ModelOp API Integration:
  • Current State: Outputs JSON files for manual drag-and-drop.
  • Future State: Implementation of the ModelOpClient to POST results directly to the ModelOp Center API immediately after generation/extraction, removing the manual upload step.
• Enhanced Red Teaming:
  • Current State: Ollama-based text rewriting.
  • Future State: Integration with Microsoft PyRIT (Python Risk Identification Toolkit) for industry-standard adversarial attack patterns.

❓ Troubleshooting Guide

1. "PermissionError: [WinError 5] Access is denied"

• The Context: OneDrive is trying to sync the generated JSON files at the same time the script is trying to modify them.
• The Fix: The script now includes "Retry Logic" to handle this. If you see warnings about locked files, just wait. The script will pause and retry automatically.

2. "Microsoft Visual C++ 14.0 is required"

• The Context: pip is trying to compile spacy from source.
• The Fix: Ensure you are using Python 3.12. Older versions (3.14 alpha) or End-of-Life versions (3.8) may not have pre-built wheels available.

3. "Ollama not running"

• The Context: The script relies on the local Ollama API.
• The Fix: Ensure the Ollama icon is visible in your system tray (bottom right on Windows). If not, open the Ollama app.