Structured output is the foundation of reliable LLM applications. Three approaches ranked by reliability: 1) JSON Mode (OpenAI/Anthropic): set response_format={"type": "json_object"} - guarantees valid JSON output but not specific schema adherence. 2) Function Calling (OpenAI): define output schema as function parameters - LLM must output valid parameters matching the JSON schema. Most reliable approach. 3) Instructor library: wraps OpenAI/Anthropic APIs with Pydantic models, automatically retries on validation failure, provides clean Python API. from instructor import from_openai; client = from_openai(OpenAI()); response = client.chat.completions.create(model="gpt-4o", response_model=ProductReview, messages=[...]). Pydantic model defines the exact output schema with validators. Production patterns: 1) Always validate outputs even with JSON mode - LLMs can still produce semantically invalid data (negative prices, future birthdates). 2) Implement retry with error feedback: if validation fails, send error message back to LLM asking it to fix specific fields. 3) Partial outputs: for long extractions, use streaming with partial validation to return results as they are generated. 4) Schema versioning: when output schema changes, version your prompts and handle old format gracefully. Real-world example: entity extraction from unstructured text using instructor with nested Pydantic models for people, organizations, locations with confidence scores.