3.2. Use Azure OpenAI to Generate Content

💡 First Principle: Generative AI is like having a brilliant colleague who can write, draw, and understand context—but needs clear instructions and guardrails. The model doesn't "know" things in the human sense; it predicts what tokens should come next based on patterns. This is why prompt engineering matters: vague prompts get vague outputs; specific prompts get useful results.

What breaks without proper configuration:

• Without appropriate temperature settings, responses are either too random (creative writing when you need facts) or too deterministic (repetitive outputs)
• Without token limits, responses can be cut off mid-sentence or consume excessive quota
• Without system messages, the model lacks behavioral constraints and may drift off-topic

Consider a customer support scenario: you want helpful, professional responses that stay on-topic. The system message sets those constraints before the conversation even begins—it's the "briefing" your AI assistant receives. The exam tests whether you understand how these parameters shape model behavior.

This section covers three capabilities: chat completions for text generation, DALL-E for images, and embeddings for semantic search.