Implement Generative
AI Solutions
This is the fastest-growing domain in AI-102. You'll learn how to build solutions using Microsoft Foundry, deploy and use Azure OpenAI models (GPT-4o, DALL-E, embeddings), implement RAG patterns, design prompts, and operationalise your AI solutions for production. Everything in Part 1 underpins this part — security, monitoring, and responsible AI all apply directly here.
Part 2 assumes you know: endpoint, API key, resource, managed identity, RBAC, environment variables, content filters, and prompt shields. All of these appear in Part 2 code and scenarios without re-explanation. If any feel unclear, revisit Part 1 first.
Generative AI is the newest and most actively updated domain in AI-102. Microsoft refreshed the exam objectives in December 2025 specifically to emphasise Azure AI Foundry as the central platform for building GenAI solutions, RAG patterns for grounding models in enterprise data, and prompt engineering techniques for improving model outputs. Questions here are heavily scenario-based — "a company needs to build X, which approach is best?" — requiring you to understand not just what each service does but when to use which deployment option, model, or technique.
What you'll learn in Part 2
Topics in this Part
Study in order — Topic 1 introduces concepts that Topics 2 and 3 build on.
The foundation topic for all generative AI work in Azure. You'll learn what Microsoft Foundry is and how to set up a hub and project, navigate the model catalog, implement Prompt Flow pipelines, build a RAG solution grounded in your own data, evaluate your models using built-in metrics, and integrate everything into a Python application using the Foundry SDK.
Deep dive into the Azure OpenAI service itself. Covers provisioning an Azure OpenAI resource, selecting and deploying models (GPT-4o, DALL-E, Whisper, embeddings), calling the Chat Completions API from Python, understanding the roles (system/user/assistant) and messages array structure, generating images with DALL-E, transcribing audio with Whisper, and working with multimodal models that understand both text and images.
Making your GenAI solution production-ready. Covers all generation parameters (temperature, top_p, penalties), quota and throughput management (TPM, PTU), prompt engineering techniques (zero-shot, few-shot, chain-of-thought), fine-tuning a model (when and how), monitoring token usage and performance, orchestrating multiple models, and deploying containers for edge and local use.