Developing AI-Powered Apps with C# and Azure AI

What is Artificial Intelligence?

In this chapter you will get a short overview about what AI is exactly, and what we can do with it.

• Definitions of Artificial Intelligence
• Machine Learning Basics
• Domains of Artificial Intelligence
• AI vs ML Engineering
• History, Current State and Future
• Overview of Microsoft AI Frameworks and SDKs

Introduction to Microsoft Foundry

Microsoft Foundry is a comprehensive platform that streamlines AI solution development and deployment. In this chapter, discover how to use Foundry for building and testing AI solutions, projects for grouping and deploying AI apps, and tools for managing resources, all while ensuring responsible AI practices are followed.

• What is Microsoft Foundry?
• Collaborative AI Development
• Deploying AI Applications with Projects
• Managing Resources and Projects
• LAB: Creating a Microsoft Foundry Project

Pre-Built AI Services with Foundry Tools

Microsoft Foundry brings together a comprehensive suite of out-of-the-box and customizable AI tools, APIs, and pre-trained models that detect sentiment, recognize speakers, understand documents and images, and much more. In this chapter you will explore these ready-to-use tools and learn how to combine them into intelligent solutions without training a single model yourself.

• Azure AI Services Overview
• Azure Content Understanding
• Azure AI Speech
• LAB: Creating an AI Powered Insurance Claim Processor

Search by Meaning with Vector Search

Vector search is a powerful technique that allows you to retrieve semantically related data from large datasets such as company documents or databases. This chapter will teach you how vector search works and how it enables you to find relevant information without depending on exact keyword based search terms or language of the information in the dataset.

• Capture Semantic Meaning with Embeddings
• Vector Search
• Vector Search Design Considerations
• LAB: Using Vector Embeddings and Semantic Search

Semantic Search Made Easy with Azure AI Search

Azure AI Search is Microsoft's enterprise-grade retrieval engine for indexing and searching your own content at scale. In this chapter you will learn how to ingest data from your own sources, build powerful search indexes, and combine keyword, vector, and semantic techniques to retrieve exactly the information your applications need. This retrieval foundation is what later powers Retrieval Augmented Generation and agentic retrieval.

• What is Azure AI Search?
• Indexing Your Own Data: Blob Storage, Azure SQL, OneLake...
• Building and Managing Search Indexes
• Keyword, Vector, and Hybrid Search
• Improving Relevance with Semantic Reranking
• Extracting Insights with AI Enrichment and Skillsets
• LAB: Building a Searchable Knowledge Index with Azure AI Search

Large Language Model Fundamentals

This module introduces Azure OpenAI and the GPT family of Large Language Models (LLMs). You'll learn about available LLM models, how to configure and use them in the Azure Portal, and the Transformer architecture behind models like GPT-5. GPT models offer Function Calling, enabling connections to external tools, services, or code, allowing the creation of AI-powered Agents. Additionally, you'll discover how Azure OpenAI provides a secure way to use LLMs without exposing your company's private data.

• Introducing OpenAI and Large Language Models
• The Transformer Model
• Understanding Tokens
• Azure OpenAI Chat Completions API
• Role Management: System, User and Assistant
• Extending LLM capabilities with Function Calling
• LAB: Building a Belgian Railway Copilot

Modern .NET AI Development with Microsoft Extensions AI

Microsoft Extensions AI is a set of core libraries providing a unified layer of abstractions for interacting with AI services. It enables developers to integrate Large Language Models (like OpenAI, Azure OpenAI, or local models via Ollama) into .NET applications using consistent, standardized interfaces. This package focuses on composable building blocks, allowing you to easily chain middleware for logging, caching, and telemetry while interacting with your custom code.

• An Introduction to Microsoft Extensions AI
• Unified Integration with AI Interfaces
• Invoking Code with Tool Calling
• Dependency Injection for AI Services
• LAB: Create a Natural Language to SQL Translation Copilot

Processing Different Modalities with LLMs

Text is no longer the only way to interact with Large Language Models. In this chapter, you will explore the multimodal capabilities of modern models like GPT-5, which allow you to reason across audio, vision, and text. You will learn how to integrate computer vision directly into your chat flow to analyze images and incorporate voice capabilities using Text-to-Speech and Speech-to-Text models.

• Introduction to Multimodal Large Language Models
• Vision with GPT Models: Chatting with your Images
• Visual Prompting Techniques
• Processing Audio: Speech-to-Text and Text-to-Speech
• LAB: Building a Voice-Enabled Visual Analysis App

Prompt Engineering and Design Patterns

In this chapter, you'll explore advanced techniques allowing you to control the model's output, transforming generic responses into precise, valuable results. Additionally the chapter covers emerging design patterns in the field of Gen AI app development that help you increase quality of model responses and reduce costs, including the orchestration patterns used to coordinate multiple LLM calls and collaborating agents.

• What is Prompt Engineering?
• Few-Shot Prompting
• Structured Query Generation
• LLM Orchestration Patterns: Prompt Chaining, Routing and Parallelization
• Multi-Agent Orchestration Patterns: Group Chat, Handoff and Magentic
• Saving Costs with Semantic Caching

Building Agentic AI Systems with Microsoft Agent Framework

This chapter introduces the fundamentals of building agentic AI systems, with a focus on the Microsoft Agent Framework. You will learn what agents are, when they are the right architectural choice, and how the Agent Framework helps structure agent behavior using models, tools, instructions, and memory. The chapter also explores the internal reasoning loop of an agent that drives it to plan, act, and respond as it works toward a goal.

• Understanding the Agent Reasoning Loop
• Core Principles of Agent Design: Models, Tools and Instructions
• Building Agents with the Microsoft Agent Framework
• Managing Agent Interactions with Sessions
• Customizing AI Agents with Middleware
• Making your Agents Observable
• LAB: Building a SQL Agent

Context Engineering and Retrieval Augmented Generation

An agent can only reason over the information it is given, so deciding what goes into the context window is one of the most important skills in AI Engineering. This chapter introduces context engineering and the Retrieval Augmented Generation (RAG) pattern, which enhances an agent's knowledge with your own company-specific data. Working in the Microsoft Agent Framework, you will learn how to ground agents in retrieved content, shape contextual system prompts by injecting user information and memories, and keep long conversations under control with context compaction. You will also see how doing this incorrectly invalidates the KV-cache and drives token costs up, and how to avoid it.

• What is Context Engineering?
• The Retrieval Augmented Generation (RAG) Pattern
• Grounding Agents with Retrieved Data
• Contextual System Prompts: Injecting User Info and Memories
• Managing the Context Window
• Context Compaction and Summarization
• Protecting the KV-Cache to Avoid Runaway Token Costs
• Reducing Hallucinations with Grounding

Extending Agents with Skills

As agents take on more responsibilities, stuffing every instruction and capability into a single prompt quickly becomes unmanageable. Agent Skills solve this by packaging instructions, scripts, and resources into portable, reusable units that agents discover and load only when needed. In this chapter you will learn how skills bring scalable context engineering to the Microsoft Agent Framework, giving your C# agents domain expertise without overwhelming the context window.

• What are Agent Skills?
• Skills vs Tools vs Instructions
• Anatomy of a Skill
• Progressive Disclosure: Loading Context On Demand
• Registering Skills with the AgentSkillsProvider
• Skills as Scalable Context Engineering
• LAB: Building a Reusable Skill for an Agent

Designing Effective Agents

Building an agent is easy; building one that is reliable, predictable, and maintainable is the real challenge. Drawing on industry guidance for building effective agents, this chapter focuses on the engineering discipline behind production-ready agents using C# and the Microsoft Agent Framework. You will learn when an agent is the right tool, how to keep agentic systems as simple as possible, and how to design for reliability, oversight, and debuggability.

• When to Use an Agent (and When Not To)
• Building Scalable Tools for your Agent
• Authoring System Instructions for your Agent
• Keeping your Agent on Track with Structured Note-Taking
• Letting your Agent Delegate Work with Subagents
• LAB: Building a Reliable Task-Completing Agent

Keeping AI Safe with Content Safety and Guardrails

How do you prevent an LLM or agent from responding inappropriately or being manipulated by malicious input? This chapter explores responsible AI in practice: detecting and blocking harmful content, defending against jailbreaks and prompt injection, and implementing guardrails that keep your agents safe and reliable. You will learn how the Azure AI Content Safety Service and human oversight help you ship trustworthy AI applications.

• Responsible AI Principles
• Preventing Inappropriate LLM Responses
• Leveraging the Azure AI Content Safety Service
• Detecting Harmful Content, Jailbreaks and Prompt Injection
• Implementing Guardrails for Safe and Reliable Agents
• Human Oversight and Safety Controls
• LAB: Protecting an AI App with Azure AI Content Safety

Evaluating Models and Agents with Microsoft Foundry

How can you ensure an LLM provides relevant and coherent answers to users' questions using the correct information? Discover the answer by exploring evaluation metrics in Microsoft Foundry. Additionally, learn how to simplify quality assessments in your .NET intelligent apps using the Microsoft.Extensions.AI.Evaluation libraries, so you can measure and continuously improve the performance of your AI solutions.

• Ensuring Coherent and Relevant LLM Responses
• Utilizing Correct Information in AI Answers
• Using Microsoft Foundry's built-in Evaluators
• Implementing Code-Based Evaluations with Microsoft.Extensions.AI.Evaluation
• Measuring and Improving AI Performance
• LAB: Evaluating the Performance of a RAG App

Unified Knowledge for Agents with Foundry IQ

Foundry IQ is a managed knowledge layer that turns scattered enterprise content into reusable, permission-aware knowledge bases for your agents. Instead of wiring each agent to each data source, you connect agents to a single knowledge base and let Foundry IQ's agentic retrieval engine plan queries, search multiple sources in parallel, and return grounded answers with citations. This chapter shows how Foundry IQ builds on Azure AI Search to deliver superior context to your agents.

• What is Foundry IQ?
• Knowledge Bases and Knowledge Sources
• Multi-Source Retrieval: Blob Storage, SharePoint, OneLake and the Web
• Agentic Retrieval: Query Planning, Decomposition and Reranking
• Permission-Aware Retrieval with ACLs and Microsoft Entra Identity
• Connecting Knowledge Bases to Agents
• Foundry IQ, Fabric IQ and Work IQ
• LAB: Building an Agentic Retrieval Solution with Foundry IQ

Self-Hosting Language Models

This chapter empowers you to bring powerful AI capabilities to end-user environments like mobile devices, personal computers and browsers, enhancing scalability, costs and performance. Additionally you will learn how to deploy and host your own open-source Language Models in the form of an API that you have full control over.

• Open-Source Language Models Overview
• Deploying AI Models on Mobile and Edge Devices with ONNX Runtime
• Hosting and Deploying Language Models on-prem and in the cloud with Ollama

Deploying Models and Agents with Microsoft Foundry

The cost and quality of your AI-powered app depend largely on your choice of AI model and how you deploy it. Learn about the available model catalog, featuring state-of-the-art Azure OpenAI models as well as models from Anthropic, Hugging Face, Meta, Google, Microsoft, Mistral, and many more. You will learn how to pick the right deployment type for a model - from pay-per-token standard to reserved provisioned throughput and dedicated managed compute - and how the global, data zone, and regional data-processing options affect cost, latency, and data residency. Finally, you will learn how to take your agents to production with Foundry's hosted agents and keep both models and agents monitored once they are live.

• Model Catalog Overview
• Model Benchmarks
• Standard vs Provisioned Throughput vs Managed Compute Deployments
• Data Processing Options: Global, Data Zone and Regional
• Balancing Cost, Latency and Data Residency
• Deploying Agents to Production with Foundry Hosted Agents
• Monitoring Deployed Models and Agents
• LAB: Deploying a Model and a Hosted Agent on Microsoft Foundry

Model Context Protocol (MCP): Standardizing LLM Integrations

This chapter explores the Model Context Protocol (MCP), an open standard revolutionizing how applications provide context to LLMs. MCP acts as a 'USB-C for AI', standardizing connections between LLMs and various data sources or tools. Crucially, MCP empowers companies to define, once and for all, precisely how their proprietary data and tools are utilized by AI systems.

• What is the Model Context Protocol (MCP)?
• Understanding MCP's Core Components: Hosts, Clients and Servers
• Defining MCP Prompts, Tools and Resources
• Building MCP Servers with C#
• LAB: Building MCP Servers and Clients with C#

Fine-tuning AI Models with Microsoft Foundry

This chapter explores the advantages of fine-tuning pre-trained LLMs for higher accuracy and customized behavior compared to Retrieval Augmented Generation (RAG). While RAG offers dynamic updates and cost-effectiveness, fine-tuning provides superior precision for specialized tasks, making it ideal for achieving domain-specific results. You will discover the different fine-tuning methods Microsoft Foundry offers - from supervised fine-tuning to preference-based and reinforcement techniques, as well as distillation - and learn how to choose the right one for your scenario.

• Introduction to Fine-Tuning LLMs
• How to decide between Fine-Tuning and RAG?
• Using Task-Specific Data for Enhanced Performance
• Supervised Fine-Tuning (SFT) for Domain Specialization
• Aligning to Preferences with Direct Preference Optimization (DPO)
• Optimizing Complex Behaviors with Reinforcement Fine-Tuning (RFT)
• Distilling Large Models into Smaller, Cheaper Ones
• Reducing Hallucinations with Fine-Tuning