Describe Features of Generative AI Workloads on Azure
How generative AI works
Large language models (LLMs) are trained on huge text corpora to predict the next token. At inference you send a prompt and receive a completion. Prompts have roles: a system message sets behaviour and guardrails, user messages are the request, assistant messages are prior model turns. Parameters matter: temperature controls randomness - higher is more creative and varied, lower is more deterministic.
Azure offerings
Azure OpenAI hosts Microsoft-managed models (GPT text and code models, image generation, embeddings) with Azure security, networking and content filters. Azure AI Foundry / model catalog adds third-party and open models. Embeddings convert text to vectors that power semantic search. A copilot is an application pattern that embeds a model into a workflow, not a model type.
Grounding: RAG versus fine-tuning
Retrieval-augmented generation (RAG) retrieves relevant documents at query time and puts them in the prompt, so answers are grounded in your current data - the model's weights never change. Fine-tuning trains a base model further on your examples to shape style, tone or format. Fine-tuning is a poor way to inject facts that change often, and it cannot guarantee freshness.
Traps and limitations
Hallucinations (confident but wrong output), bias, data cut-off, latency and cost are inherent. Prompt injection - malicious instructions hidden in retrieved content - is a generative AI-specific risk that content filters alone do not fully solve. Classic exam trap: choose RAG for "our product documentation changed yesterday"; choose fine-tuning for "always answer in our brand voice and output JSON".
Rule of thumb
Facts and freshness → RAG. Behaviour and format → fine-tuning.