guidance

Constraining LLM outputs with regex and grammar rules requires a Guidance runtime that applies token healing and context-free grammar constraints, enforcing valid structured outputs across JSON, XML, and code-like formats during generation.

Constrained generation is a decoding mechanism that applies regex constraints and context-free grammars to language models, preventing malformed JSON, XML, or CSV responses and eliminating the need for output parsing and validation.

Constrained generation supports multiple LLM backends, including both local and API-based models, applying grammar constraints and robust validation uniformly across the Guidance runtime to ensure consistent structured outputs.

You can use regex constraints for data extraction in multi-step workflows by applying grammar rules that enforce structured results, enabling robust validation across sequential generation steps in software engineering contexts.

The best way to ensure valid JSON without post-processing is to apply context-free grammar constraints during generation, forcing the model to produce syntactically correct structured outputs directly at the token level.

LLMs output malformed structured data because prompt instructions lack hard constraints; applying regex and grammar rules during decoding enforces valid token sequences, preventing formatting errors in JSON, XML, and CSV outputs.