🛡️ Azure Content Safety Integration with Semantic Kernel

This sample demonstrates how to use Azure Content Safety to filter LLM interactions in a Semantic Kernel-powered application. It shows how to implement a PromptRenderFilter to block unsafe inputs or outputs based on:

• ✅ Built-in safety categories: Sexual, Self-Harm, Hate, Violence
• 🔧 Custom safety category support (e.g., personal-data , internal policy )

📌 Technologies Used:

• .NET 8 / C#
• Azure Content Safety API
• Microsoft Semantic Kernel
• PromptRenderFilter (for input/output filtering)

---

🧠 How It Works

When a user prompt is sent to the LLM (like GPT-4), the app:

1. Intercepts the prompt using a PromptRenderFilter
2. Calls Azure Content Safety API to check the prompt
3. Blocks or allows the request based on severity thresholds
4. Supports out-of-the-box categories and custom categories simultaneously

---

🛠️ Setting Up a Custom Category in Azure Content Safety

To enforce organization-specific safety filters (e.g., personal-data or sensitive project names), you can create and train a Custom Category in Azure Content Safety.

Follow these steps:

1. Go to Azure AI Foundry
   Navigate to Azure AI Studio or your Azure AI resource in the Azure Portal.

2. Open Guardrails + Controls
   From the left-hand navigation pane, click Guardrails + Controls.

3. Select the Custom Categories tab
   Scroll to the bottom and click the Create custom category button.

4. Add New Category Details

   • Click Add new category
   • Provide a Name (e.g., personal-data)
   • Add a Description
   • Enter a public URL pointing to your .jsonl training file stored in Azure Blob Storage , Sample data

     grant Data Reader and Contributor roles for both Azure AI Foundry and Azure Content Safety services

     🔐 Assign Roles for Azure AI Foundry and Content Safety

Click + Add → Add role assignment twice for each identity:

---

🧠 Role 1: Storage Blob Data Reader

Purpose: Allows reading your training data (the .jsonl file)

Assign to:

• Azure AI Foundry resource (type: managed identity)
• Azure AI Content Safety resource (type: managed identity)

---

🧩 Role 2: Storage Blob Data Contributor

Purpose: Allows writing temporary artifacts or metadata during training

Assign to:

• The same two resources listed above.

---

✅ Result:
These two roles together allow Azure AI Foundry and Azure Content Safety to securely read and process your training files from Azure Blob Storage.

5. Click Create and Train
   Azure will begin training your category. This may take a few minutes.