Building AI-Powered Applications with Python on Azure

Azure OpenAI vs OpenAI direct, deployment model (resource/deployment/model), Python SDK with AzureOpenAI client, chat completions (messages array/system/user/assistant), system prompts for behavior, parameters (temperature/max_tokens/top_p)

JSON mode (response_format json_object), structured outputs with Pydantic/JSON schema, function calling (tool schemas/tools parameter/parsing tool_calls), multi-turn with tools (call/execute/return/continue), validation and retry on schema violations

What embeddings are (text to vectors/semantic similarity/cosine), Azure OpenAI text-embedding-3-small, Azure AI Search (index with vector fields/upload documents), vector and hybrid search, document chunking strategies

RAG architecture (query/retrieve/augment/generate), building the pipeline (embed query/search index/assemble context), prompt engineering for RAG (grounding/citations/not found handling), Azure AI Search integration with SearchClient, quality (hallucination/attribution/context limits)

Agent architecture (observe/think/act loop), tool definitions as Python functions with JSON schemas, agent loop (while tool_calls: execute/append/continue), Azure-integrated tools (Blob Storage/Cosmos DB/Functions), guardrails (max iterations/budget/content safety)

Evaluation (ground truth/LLM-as-judge/precision-recall for RAG), cost management (tiktoken counting/caching/model selection), rate limiting (TPM/RPM/retry with backoff), streaming (stream=True/FastAPI StreamingResponse), production architecture (async clients/health checks/graceful degradation)