Skip to content

AI_FOUNDRY_INFRASTRUCTURE_SETUP.md

Latest commit

 

History

History
214 lines (172 loc) · 6.71 KB

AI_FOUNDRY_INFRASTRUCTURE_SETUP.md

File metadata and controls

214 lines (172 loc) · 6.71 KB

AI Foundry Infrastructure Setup - Complete

Summary

Successfully implemented the Azure AI Foundry infrastructure following the pattern from foundry-samples/45-basic-agent-bing.

Infrastructure Components Deployed

1. Bing Search for Grounding (modules/bing/grounding-bing-search.bicep)

  • Resource: Microsoft.Bing/accounts@2020-06-10
  • SKU: G1
  • Kind: Bing.Grounding
  • Purpose: Provides web search capabilities for the AI agent

2. AI Foundry Account (modules/ai/account.bicep)

  • Resource: Microsoft.CognitiveServices/accounts@2025-04-01-preview
  • Kind: AIServices
  • Name: aifoundry-{resourceToken}
  • Features:
    • System-assigned managed identity
    • Supports project management (allowProjectManagement: true)
    • Model deployment: gpt-4o-realtime-preview
    • Public network access enabled

3. AI Foundry Project (modules/ai/project.bicep)

  • Resource: Microsoft.CognitiveServices/accounts/projects@2025-04-01-preview
  • Name: project-{resourceToken}
  • Parent: AI Foundry Account
  • Features:
    • System-assigned managed identity
    • Project description and display name

4. Bing Search Connection (modules/ai/bing-connection.bicep)

  • Resource: Microsoft.CognitiveServices/accounts/connections@2025-04-01-preview
  • Category: ApiKey
  • Auth Type: ApiKey (using Bing Search keys)
  • Purpose: Connects AI Foundry account to Bing Search for grounding

5. RBAC Configuration (modules/ai/rbac.bicep)

  • Role: Azure AI Developer (64702f94-c441-49e6-a78b-ef80e0188fee)
  • Assignees:
    • Backend Container App Managed Identity
    • User Principal (for testing/development)

Environment Variables Added to Backend

The following environment variables are now available in the backend Container App:

AZURE_AI_FOUNDRY_ENDPOINT=<AI Foundry Account Endpoint>
AZURE_AI_FOUNDRY_PROJECT_ID=<Project Resource ID>
AZURE_AI_FOUNDRY_BING_CONNECTION_ID=<Bing Connection Resource ID>

Deployment Validation

✅ Bicep Compilation: Successfully validated with az bicep build

  • No errors
  • Expected warnings for Bing resource type (no schema available)

Next Steps for Phase 1 Implementation

1. Create MCP Route in Backend (see previous recommendation)

# src/backend/routes/mcp_agents.py
from fastapi import APIRouter, HTTPException
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential

mcp_router = APIRouter()

class AIFoundryAgentService:
    def __init__(self):
        self.client = AIProjectClient(
            credential=DefaultAzureCredential(),
            endpoint=os.getenv("AZURE_AI_FOUNDRY_ENDPOINT")
        )
        self.agent_id = None
        
    async def initialize(self):
        # Create agent with Bing Search tool
        agent = await self.client.agents.create_agent(
            model=os.getenv("AZURE_OPENAI_GPT_REALTIME_DEPLOYMENT"),
            name="Web Search Agent",
            instructions="...",
            tools=[{"type": "bing_grounding"}]
        )
        self.agent_id = agent.id
        
    async def search_web(self, query: str) -> str:
        # Create ephemeral thread
        thread = await self.client.agents.create_thread()
        try:
            # Execute search via AI Foundry agent
            # ... (implementation as previously described)
        finally:
            await self.client.agents.delete_thread(thread.id)

2. Register MCP Route

# src/backend/main.py
from routes.mcp_agents import mcp_router

app.include_router(mcp_router, prefix="/api/mcp", tags=["mcp"])

3. Update Orchestrator

# src/backend/services/assistant_service.py
async def handle_tool_call(self, tool_name: str, parameters: dict, call_id: str):
    if tool_name == "search_web_ai_foundry":
        # Call local MCP endpoint
        response = await self.http_client.post(
            f"{self.backend_base_url}/api/mcp/ai-foundry",
            json={
                "jsonrpc": "2.0",
                "id": call_id,
                "method": "tools/call",
                "params": {
                    "name": tool_name,
                    "arguments": parameters
                }
            }
        )
        # ... handle response
    else:
        # Existing tool handling
        pass

4. Deploy and Test

# Deploy infrastructure
azd provision

# Deploy backend with new route
azd deploy backend

# Test MCP endpoint
curl -X POST https://<backend-url>/api/mcp/ai-foundry \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/list"
  }'

Architecture Diagram

Browser WebSocket
    ↓
FastAPI Backend (Container App)
    ├─ /api/mcp/ai-foundry (NEW MCP endpoint)
    │     ↓
    │  AI Foundry Agent Service
    │     ↓
    │  Azure AI Foundry Project
    │     ├─ AI Foundry Account
    │     └─ Bing Search Connection
    │           └─ Bing Grounding Service
    │
    └─ Existing routes (/api/*, /api/realtime/*)

Key Benefits

  1. ✅ Minimal Infrastructure Changes: Added to existing backend, no new Container App needed
  2. ✅ Zero Additional Cost: Uses existing backend Container App
  3. ✅ Low Latency: In-process HTTP call (~1-5ms vs 20-50ms network hop)
  4. ✅ Simple Deployment: Single azd up deploys everything
  5. ✅ Easier Debugging: All logs in one place
  6. ✅ Gradual Migration: Can extract to separate service later if needed

Files Modified/Created

Created:

  • infra/modules/ai/account.bicep
  • infra/modules/ai/project.bicep
  • infra/modules/ai/bing-connection.bicep
  • infra/modules/ai/rbac.bicep

Modified:

  • infra/main.bicep (added AI Foundry resources, RBAC, env vars, outputs)

Existing (Used):

  • infra/modules/bing/grounding-bing-search.bicep (already present)

Deployment Command

# Provision all infrastructure (including AI Foundry)
azd provision

# The following resources will be created:
# - Bing Search for Grounding
# - AI Foundry Account
# - AI Foundry Project
# - Bing Search Connection
# - RBAC role assignments

Documentation References