Implement Generative AI and Agentic Solutions
Deployment and control surface
You deploy a model (for example GPT-family) into an Azure OpenAI resource and consume it by deployment name, not model name. Deployment types trade cost against latency predictability: Global Standard is cheapest with variable latency, Standard is regional, Provisioned reserves throughput for steady high load. Know the sampling knobs: temperature and top_p should be tuned one at a time; deterministic extraction tasks want near-zero temperature, creative tasks want higher. max_tokens caps output, not input context.
Grounding with retrieval
For answers that must be traceable, use RAG: chunk source content, create embeddings, store them in Azure AI Search, retrieve with hybrid (keyword + vector) search and the semantic ranker, then instruct the model to answer only from the retrieved passages and cite them. RAG reduces hallucination but never eliminates it - add a groundedness check and a fallback such as "I don't know." Content filters score harm categories by severity and can block prompts or completions; you tune thresholds, you do not simply switch safety off.
Agents and tools
An agent combines instructions, a model, and tools (function calling, code interpreter, retrieval, API actions). The loop is: user message → run → tool call → tool result → final response, persisted in a thread. Define tools with precise names, descriptions, and JSON schemas; vague schemas cause wrong or missing tool calls. Keep the tool surface small.
Traps
Fine-tuning versus RAG (format/style versus fresh facts), confusing deployment name with model name, and assuming a higher temperature improves accuracy.
Rule of thumb: facts and citations → RAG; tone, format, or a narrow classification style → fine-tuning; multi-step action taking → agent with explicit tools.