guidance

To ensure LLM outputs match a JSON or XML schema, you apply constrained generation using regex and grammars. This enforces structured formats during decoding, guaranteeing valid syntax and preventing malformed output.

Constrained generation controls LLM outputs by applying regex patterns and grammars during token sampling. It works with local models by enforcing token healing and grammar-based validation to improve correctness and reliability.

You can use regex to validate specific fields like emails in LLM responses by applying grammar constraints. This forces the model to generate text matching the regex pattern, ensuring valid email formats and structured output.

You orchestrate multi-step prompts with structured outputs by using Pythonic guidance functions. This allows you to build multi-step workflows that apply constraints and grammar-based generation across complex task sequences.

Constrained generation works for generating code with validation by applying grammars to enforce syntax rules. Token healing and grammar-based validation improve correctness, making it suitable for reliable code generation and audit trails.

Token healing is a technique that fixes boundary token issues during constrained generation. You need it for structured output when applying regex and grammars to prevent tokenization errors and ensure seamless format compliance.