Microsoft Foundry Series (Part 4) — Foundry Services: Agents, Responses API, Tools, Memory and Real-World Use Cases

 1┌─────────────────────────────────────────────┐
 2│              Prompt Agent                   │
 3│                                             │
 4│  Instructions (prompt) ──► Model (GPT-5)    │
 5│         │                      │            │
 6│         └──── Tools ──────────┘             │
 7│         (file search, web search, etc.)     │
 8│                                             │
 9│  Runtime: Fully managed by Foundry          │
10└─────────────────────────────────────────────┘

 1import os
 2from openai import OpenAI
 3
 4client = OpenAI(
 5    api_key=os.getenv("AZURE_OPENAI_API_KEY"),
 6    base_url="https://YOUR-RESOURCE.openai.azure.com/openai/v1/",
 7)
 8
 9response = client.responses.create(
10    model="gpt-4.1-mini",
11    input="Generate a short product description for a wireless mouse."
12)
13
14print(response.output_text)

 1# First turn
 2first = client.responses.create(
 3    model="gpt-4.1-mini",
 4    input="I need help writing product descriptions for an e-commerce store."
 5)
 6
 7# Second turn — automatically carries forward context
 8second = client.responses.create(
 9    model="gpt-4.1-mini",
10    previous_response_id=first.id,
11    input="The first product is a noise-cancelling headphone. Price: $149."
12)
13
14print(second.output_text)

1stream = client.responses.create(
2    model="gpt-4.1-mini",
3    input="Write a detailed product description for a mechanical keyboard.",
4    stream=True,
5)
6
7for event in stream:
8    if event.type == "response.output_text.delta":
9        print(event.delta, end="")

 1import json
 2
 3response = client.responses.create(
 4    model="gpt-4.1-mini",
 5    tools=[
 6        {
 7            "type": "function",
 8            "name": "get_product_inventory",
 9            "description": "Check inventory level for a product by SKU",
10            "parameters": {
11                "type": "object",
12                "properties": {
13                    "sku": {"type": "string", "description": "Product SKU"}
14                },
15                "required": ["sku"],
16            },
17        }
18    ],
19    input="What's the inventory level for SKU WM-2024-BLK?",
20)
21
22# Process function calls
23for item in response.output:
24    if item.type == "function_call":
25        args = json.loads(item.arguments)
26        # Call your actual inventory API
27        inventory = {"sku": args["sku"], "quantity": 142, "warehouse": "EU-West"}
28
29        # Return the result to the model
30        final = client.responses.create(
31            model="gpt-4.1-mini",
32            previous_response_id=response.id,
33            input=[{
34                "type": "function_call_output",
35                "call_id": item.call_id,
36                "output": json.dumps(inventory),
37            }],
38        )
39        print(final.output_text)

1response = client.responses.create(
2    model="gpt-4.1-mini",
3    tools=[{"type": "code_interpreter", "container": {"type": "auto"}}],
4    instructions="You are a data analyst. Write and run Python code to answer questions.",
5    input="Calculate the compound annual growth rate if revenue grew from $1M to $2.5M over 4 years."
6)
7
8print(response.output_text)

1response = client.responses.create(
2    model="gpt-4.1-mini",
3    tools=[{"type": "web_search_preview"}],
4    input="What are the top trending wireless mouse models in 2026?"
5)
6
7print(response.output_text)

 1response = client.responses.create(
 2    model="gpt-4.1-mini",
 3    tools=[
 4        {
 5            "type": "mcp",
 6            "server_label": "github",
 7            "server_url": "https://gitmcp.io/erudinsky/microsoft-foundry-series",
 8            "require_approval": "never"
 9        }
10    ],
11    input="What files are in this repository?"
12)
13
14print(response.output_text)

 1response = client.responses.create(
 2    model="gpt-4.1-mini",
 3    tools=[
 4        {
 5            "type": "mcp",
 6            "server_label": "internal-api",
 7            "server_url": "https://api.contoso.com/mcp",
 8            "headers": {"Authorization": f"Bearer {mcp_token}"},
 9            "require_approval": "never"
10        }
11    ],
12    input="List all active products."
13)

 1response = client.responses.create(
 2    model="gpt-4.1-mini",
 3    tools=[{"type": "memory"}],
 4    input="Remember that our brand voice is professional but friendly, and we always mention free shipping."
 5)
 6
 7# In a completely new conversation later...
 8response2 = client.responses.create(
 9    model="gpt-4.1-mini",
10    tools=[{"type": "memory"}],
11    input="Write a product description for a yoga mat."
12)
13
14# The agent recalls the brand voice preference from memory
15print(response2.output_text)

 1import json
 2import os
 3from openai import OpenAI
 4
 5client = OpenAI(
 6    api_key=os.getenv("AZURE_OPENAI_API_KEY"),
 7    base_url=f"https://{os.getenv('FOUNDRY_RESOURCE')}.openai.azure.com/openai/v1/",
 8)
 9
10TOOLS = [
11    {
12        "type": "function",
13        "name": "get_product_details",
14        "description": "Retrieve product details from the catalog database",
15        "parameters": {
16            "type": "object",
17            "properties": {
18                "sku": {"type": "string", "description": "Product SKU identifier"}
19            },
20            "required": ["sku"],
21        },
22    },
23    {"type": "web_search_preview"},
24    {"type": "memory"},
25]
26
27INSTRUCTIONS = """You are a product description writer for an e-commerce store.
28
29When asked to write a description:
301. Use get_product_details to fetch product info from the catalog
312. Use web search to check competitor positioning and trending keywords
323. Check memory for brand voice guidelines and past preferences
334. Write a compelling, SEO-friendly product description
34
35Format: Title, subtitle, 3-4 bullet points, and a short paragraph."""
36
37
38def handle_function_call(item):
39    """Simulate a product catalog lookup."""
40    args = json.loads(item.arguments)
41    # In production, this calls your actual database
42    catalog = {
43        "KB-MEC-2026": {
44            "name": "ProType Mechanical Keyboard",
45            "price": 129.99,
46            "features": ["Cherry MX Brown switches", "RGB backlighting",
47                         "USB-C", "Hot-swappable keys"],
48            "category": "Peripherals",
49            "in_stock": True,
50            "stock_quantity": 284,
51        }
52    }
53    product = catalog.get(args["sku"], {"error": "Product not found"})
54    return json.dumps(product)
55
56
57def generate_description(sku: str) -> str:
58    """Run the agentic loop to generate a product description."""
59    response = client.responses.create(
60        model="gpt-4.1-mini",
61        tools=TOOLS,
62        instructions=INSTRUCTIONS,
63        input=f"Write a product description for SKU: {sku}",
64    )
65
66    # Handle tool calls in a loop
67    while any(item.type == "function_call" for item in response.output):
68        tool_outputs = []
69        for item in response.output:
70            if item.type == "function_call":
71                result = handle_function_call(item)
72                tool_outputs.append({
73                    "type": "function_call_output",
74                    "call_id": item.call_id,
75                    "output": result,
76                })
77
78        response = client.responses.create(
79            model="gpt-4.1-mini",
80            tools=TOOLS,
81            instructions=INSTRUCTIONS,
82            previous_response_id=response.id,
83            input=tool_outputs,
84        )
85
86    return response.output_text
87
88
89if __name__ == "__main__":
90    description = generate_description("KB-MEC-2026")
91    print(description)

 1# After a long conversation, compact the context
 2compacted = client.responses.compact(
 3    model="gpt-4.1-mini",
 4    previous_response_id=response.id,
 5)
 6
 7# Continue with the compacted context
 8follow_up = client.responses.create(
 9    model="gpt-4.1-mini",
10    input=[*compacted.output, {"role": "user", "content": "Now write it in French."}],
11)

1response = client.responses.create(
2    model="gpt-4.1-mini",
3    input=conversation,
4    store=False,
5    context_management=[{"type": "compaction", "compact_threshold": 200000}],
6)

 1@description('Models for agent workloads')
 2param models array = [
 3  {
 4    name: 'gpt-4-1-mini'
 5    modelName: 'gpt-4.1-mini'
 6    modelVersion: '2025-04-14'
 7    capacity: 10
 8  }
 9  {
10    name: 'gpt-4-1'
11    modelName: 'gpt-4.1'
12    modelVersion: '2025-04-14'
13    capacity: 5
14  }
15]
16
17resource deployments 'Microsoft.CognitiveServices/accounts/deployments@2025-04-01-preview' = [
18  for model in models: {
19    parent: foundry
20    name: model.name
21    sku: {
22      name: 'GlobalStandard'
23      capacity: model.capacity
24    }
25    properties: {
26      model: {
27        format: 'OpenAI'
28        name: model.modelName
29        version: model.modelVersion
30      }
31    }
32  }
33]

1az group delete --name rg-foundry-demo --yes --no-wait