structured generation

Structured generation produces outputs in specific formats (JSON, XML, code) with guaranteed validity. **Problem**: LLMs sometimes produce invalid formats despite instructions - malformed JSON, syntax errors, schema violations. **Solution**: Constrain token selection to only valid continuations during decoding. **Approaches**: **Grammar-constrained**: Define format grammar, reject invalid tokens at each step. **Schema-guided**: JSON Schema or Pydantic models specify structure, generate compliant outputs. **Template-based**: Fill in designated slots in predefined structure. **Tools**: Outlines (fast grammar-guided generation), Instructor (Pydantic-based extraction), Marvin, Guidance, llama.cpp GBNF grammars. **JSON example**: Define schema → during generation, only allow valid JSON tokens → output guaranteed parseable. **Performance**: Minor latency overhead, major reliability improvement - eliminates format-related retries. **Best practices**: Define strict schemas, validate outputs anyway (defense in depth), handle edge cases in schemas. **Advanced**: TypeScript/Python type generation, nested object extraction, union types. Critical for production pipelines requiring reliable structured data extraction.