Microsoft Foundry

Microsoft Foundry Skill

This skill helps developers work with Microsoft Foundry resources, covering model discovery and deployment, RAG (Retrieval-Augmented Generation) applications, AI agent creation, evaluation workflows, and troubleshooting.

When to Use This Skill

Use this skill when the user wants to:

• Discover and deploy AI models from the Microsoft Foundry catalog
• Build RAG applications using knowledge indexes and vector search
• Create AI agents with tools like Azure AI Search, web search, or custom functions
• Evaluate agent performance using built-in evaluators
• Set up monitoring and continuous evaluation for production agents
• Troubleshoot issues with deployments, agents, or evaluations

Prerequisites

Azure Resources

• An Azure subscription with an active account
• Appropriate permissions to create Microsoft Foundry resources (e.g., Azure AI Owner role)
• Resource group for organizing Foundry resources

Tools

• Azure CLI installed and authenticated (az login)
• Azure Developer CLI (azd) for deployment workflows (optional but recommended)

Language-Specific Requirements

For SDK examples and implementation details in specific programming languages, refer to:

• Python: See [language/python.md](language/python.md) for Python SDK setup, authentication, and examples

Core Workflows

1. Getting Started - Model Discovery and Deployment

Use Case

A developer new to Microsoft Foundry wants to explore available models and deploy their first one.

Step 1: List Available Resources

First, help the user discover their Microsoft Foundry resources.

Using Azure CLI:

Bash

# List all Microsoft Foundry resources in subscription
az resource list \
  --resource-type "Microsoft.CognitiveServices/accounts" \
  --query "[?kind=='AIServices'].{Name:name, ResourceGroup:resourceGroup, Location:location}" \
  --output table

# List resources in a specific resource group
az resource list \
  --resource-group  \
  --resource-type "Microsoft.CognitiveServices/accounts" \
  --output table

Using MCP Tools:

Use the foundry_resource_get MCP tool to get detailed information about a specific Foundry resource, or to list all resources if no name is provided.

Step 2: Browse Model Catalog

Help users discover available models, including information about free playground support.

Key Points to Explain:

• Some models support free playground for prototyping without costs
• Models can be filtered by publisher (e.g., OpenAI, Meta, Microsoft)
• Models can be filtered by license type
• Model availability varies by region

Using MCP Tools:

Use the foundry_models_list MCP tool:

• List all models: foundry_models_list()
• List free playground models: foundry_models_list(search-for-free-playground=true)
• Filter by publisher: foundry_models_list(publisher="OpenAI")
• Filter by license: foundry_models_list(license="MIT")

Example Output Explanation: When listing models, explain to users:

• Models with free playground support can be used for prototyping at no cost
• Some models support GitHub token authentication for easy access
• Check model capabilities and pricing before production deployment

Step 3: Deploy a Model

Guide users through deploying a model to their Foundry resource.

Using Azure CLI:

Bash

# Deploy a model (e.g., gpt-4o)
az cognitiveservices account deployment create \
  --name  \
  --resource-group  \
  --deployment-name gpt-4o-deployment \
  --model-name gpt-4o \
  --model-version "2024-05-13" \
  --model-format OpenAI \
  --sku-capacity 10 \
  --sku-name Standard

# Verify deployment status
az cognitiveservices account deployment show \
  --name  \
  --resource-group  \
  --deployment-name gpt-4o-deployment

Using MCP Tools:

Use the foundry_models_deploy MCP tool with parameters:

• resource-group: Resource group name
• deployment: Deployment name
• model-name: Model to deploy (e.g., "gpt-4o")
• model-format: Format (e.g., "OpenAI")
• azure-ai-services: Foundry resource name
• model-version: Specific version
• sku-capacity: Capacity units
• scale-type: Scaling type

Deployment Verification: Explain that when deployment completes, provisioningState should be Succeeded. If it fails, common issues include:

• Insufficient quota
• Region capacity limitations
• Permission issues

Step 4: Get Resource Endpoint

Users need the project endpoint to connect their code to Foundry.

Using MCP Tools:

Use the foundry_resource_get MCP tool to retrieve resource details including the endpoint.

Expected Output: The endpoint will be in format: https://.services.ai.azure.com/api/projects/

Save this endpoint as it's needed for subsequent API and SDK calls.

2. Building RAG Applications with Knowledge Indexes

Use Case

A developer wants to build a Retrieval-Augmented Generation (RAG) application using their own documents.

Understanding RAG and Knowledge Indexes

Explain the Concept: RAG enhances AI responses by:

1. Retrieving relevant documents from a knowledge base
2. Augmenting the AI prompt with retrieved context
3. Generating responses grounded in factual information

Knowledge Index Benefits:

• Supports keyword, semantic, vector, and hybrid search
• Enables efficient retrieval of relevant content
• Stores metadata for better citations (document titles, URLs, file names)
• Integrates with Azure AI Search for production scenarios

Step 1: List Existing Knowledge Indexes

Using MCP Tools:

Use foundry_knowledge_index_list with your project endpoint to list knowledge indexes.

Step 2: Inspect Index Schema

Understanding the index structure helps optimize queries.

Using MCP Tools:

Use the foundry_knowledge_index_schema MCP tool with your project endpoint and index name to get detailed schema information.

Schema Information Includes:

• Field definitions and data types
• Searchable attributes
• Vectorization configuration
• Retrieval mode support (keyword, semantic, vector, hybrid)

Step 3: Create an Agent with Azure AI Search Tool

Implementation:

To create a RAG agent with Azure AI Search tool integration:

1. Initialize the AI Project Client with your project endpoint and credentials
2. Get the Azure AI Search connection from your project
3. Create the agent with:

• Agent name
• Model deployment
• Clear instructions (see best practices below)
• Azure AI Search tool configuration with:
• Connection ID
• Index name
• Query type (HYBRID recommended)

For SDK Implementation: See [language/python.md](language/python.md#rag-applications-with-python-sdk)

Key Best Practices:

• Always request citations in agent instructions
• Use hybrid search (AzureAISearchQueryType.HYBRID) for best results
• Instruct the agent to say "I don't know" when information isn't in the index
• Format citations consistently for easy parsing

Step 4: Test the RAG Agent

Testing Process:

1. Query the agent with a test question
2. Stream the response to get real-time output
3. Capture citations from the response annotations
4. Validate that citations are properly formatted and included

For SDK Implementation: See [language/python.md](language/python.md#testing-the-rag-agent)

Troubleshooting RAG Issues:

| Issue | Possible Cause | Resolution | |-------|---------------|------------| | No citations in response | Agent instructions don't request citations | Update instructions to explicitly request citation format | | "Index not found" error | Wrong index name or connection | Verify AI_SEARCH_INDEX_NAME matches index in Azure AI Search | | 401/403 authentication error | Missing RBAC permissions | Assign project managed identity Search Index Data Contributor role | | Poor retrieval quality | Query type not optimal | Try HYBRID query type for better results |

3. Creating Your First AI Agent

Use Case

A developer wants to create an AI agent with tools (web search, function calling, file search).

Step 1: List Existing Agents

Using MCP Tools:

Use foundry_agents_list with your project endpoint to list existing agents.

Step 2: Create a Basic Agent

Implementation:

Create an agent with:

• Model deployment name: The model to use
• Agent name: Unique identifier
• Instructions: Clear, specific guidance for the agent's behavior

For SDK Implementation: See [language/python.md](language/python.md#basic-agent)

Step 3: Create an Agent with Custom Function Tools

Agents can call custom functions to perform actions like querying databases, calling APIs, or performing calculations.

Implementation Steps:

1. Define custom functions with clear docstrings describing their purpose and parameters
2. Create a function toolset with your custom functions
3. Create the agent with the toolset and instructions on when to use the tools

For SDK Implementation: See [language/python.md](language/python.md#agent-with-custom-function-tools)

Step 4: Create an Agent with Web Search

Implementation:

Create an agent with web search capabilities by adding a Web Search tool:

• Optionally specify user location for localized results
• Provide instructions to always cite web sources

For SDK Implementation: See [language/python.md](language/python.md#agent-with-web-search)

Step 5: Interact with the Agent

Interaction Process:

1. Create a conversation thread for the agent interaction
2. Add user messages to the thread
3. Run the agent to process the messages and generate responses
4. Check run status for success or failure
5. Retrieve messages to see the agent's responses
6. Cleanup by deleting the agent when done

For SDK Implementation: See [language/python.md](language/python.md#interacting-with-agents)

Agent Best Practices:

1. Clear Instructions: Provide specific, actionable instructions
2. Tool Selection: Only include tools the agent needs
3. Error Handling: Always check run.status for failures
4. Cleanup: Delete agents/threads when done to manage costs
5. Rate Limits: Handle rate limit errors gracefully (status code 429)

4. Evaluating Agent Performance

Use Case

A developer has built an agent and wants to evaluate its quality, safety, and performance.

Understanding Agent Evaluators

Built-in Evaluators:

1. IntentResolutionEvaluator: Measures how well the agent identifies and understands user requests (score 1-5)
2. TaskAdherenceEvaluator: Evaluates whether responses adhere to assigned tasks and system instructions (score 1-5)
3. ToolCallAccuracyEvaluator: Assesses whether the agent makes correct function tool calls (score 1-5)

Evaluation Output: Each evaluator returns:

• {metric_name}: Numerical score (1-5, higher is better)
• {metric_name}_result: "pass" or "fail" based on threshold
• {metric_name}_threshold: Binarization threshold (default or user-set)
• {metric_name}_reason: Explanation of the score

Step 1: Single Agent Run Evaluation

Using MCP Tools:

Use the foundry_agents_query_and_evaluate MCP tool to query an agent and evaluate the response in one call. Provide:

• Agent ID
• Query text
• Project endpoint
• Azure OpenAI endpoint and deployment for evaluation
• Comma-separated list of evaluators to use

Example Output:

{
  "response": "The weather in Seattle is currently sunny and 22°C.",
  "evaluation": {
    "intent_resolution": 5.0,
    "intent_resolution_result": "pass",
    "intent_resolution_threshold": 3,
    "intent_resolution_reason": "The agent correctly identified the user's intent to get weather information and provided a relevant response.",
    "task_adherence": 4.0,
    "task_adherence_result": "pass",
    "tool_call_accuracy": 5.0,
    "tool_call_accuracy_result": "pass"
  }
}

Step 2: Evaluate Existing Response

If you already have the agent's response, you can evaluate it directly.

Using MCP Tools:

Use the foundry_agents_evaluate MCP tool to evaluate a specific query/response pair with a single evaluator.

For SDK Implementation: See [language/python.md](language/python.md#single-response-evaluation-using-mcp)

Step 3: Batch Evaluation

For evaluating multiple agent runs across multiple conversation threads:

1. Convert agent thread data to evaluation format
2. Prepare evaluation data from multiple thread IDs
3. Set up evaluators with appropriate configuration
4. Run batch evaluation and view results in the Foundry portal

For SDK Implementation: See [language/python.md](language/python.md#batch-evaluation)

Interpreting Evaluation Results

Score Ranges (1-5 scale):

• 5: Excellent - Agent perfectly understood and executed the task
• 4: Good - Minor issues, but overall successful
• 3: Acceptable - Threshold for passing (default)
• 2: Poor - Significant issues with understanding or execution
• 1: Failed - Agent completely misunderstood or failed the task

Common Evaluation Issues:

| Issue | Cause | Resolution | |-------|-------|------------| | Job stuck in "Running" | Insufficient model capacity | Increase model quota/capacity and rerun | | All metrics zero | Wrong evaluator or unsupported model | Verify evaluator compatibility with your model | | Groundedness unexpectedly low | Incomplete context/retrieval | Verify RAG retrieval includes sufficient context | | Evaluation missing | Not selected during setup | Rerun evaluation with required metrics |

5. Troubleshooting Common Issues

Deployment Issues

Problem: Deployment Stays Pending or Fails

Bash

# Check deployment status and details
az cognitiveservices account deployment show \
  --name  \
  --resource-group  \
  --deployment-name  \
  --output json

# Check account quota
az cognitiveservices account show \
  --name  \
  --resource-group  \
  --query "properties.quotaLimit"

Common Causes:

• Insufficient quota in the region
• Region at capacity for the model
• Permission issues

Resolution:

1. Check quota limits in Azure Portal
2. Request quota increase if needed
3. Try deploying to a different region
4. Verify you have appropriate RBAC permissions

Agent Response Issues

Problem: Agent Doesn't Return Citations (RAG)

Diagnostics:

1. Check agent instructions explicitly request citations
2. Verify the tool choice is set to "required" or "auto"
3. Confirm the Azure AI Search connection is configured correctly

Resolution:

Update the agent's instructions to explicitly request citations in the format [message_idx:search_idx†source] and to only use the knowledge base, never the agent's own knowledge.

For SDK Implementation: See [language/python.md](language/python.md#update-agent-instructions)

Problem: "Index Not Found" Error

Using MCP Tools:

Use the foundry_knowledge_index_list MCP tool to verify the index exists and get the correct name.

Resolution:

1. Verify AI_SEARCH_INDEX_NAME environment variable matches actual index name
2. Check the connection points to correct Azure AI Search resource
3. Ensure index has been created and populated

Problem: 401/403 Authentication Errors

Common Cause: Missing RBAC permissions

Resolution:

Bash

# Assign Search Index Data Contributor role to managed identity
az role assignment create \
  --assignee  \
  --role "Search Index Data Contributor" \
  --scope /subscriptions//resourceGroups//providers/Microsoft.Search/searchServices/

# Verify role assignment
az role assignment list \
  --assignee  \
  --output table

Evaluation Issues

Problem: Evaluation Dashboard Shows No Data

Common Causes:

• No recent agent traffic
• Time range excludes the data
• Ingestion delay

**Res

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Tyler-R-Kendrick
• Source: Tyler-R-Kendrick/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.