guidance

To constrain LLM outputs, you apply regex and context-free grammars to the generation process, guaranteeing valid JSON/XML/code. This structured generation framework enforces formatting rules during decoding to reduce erroneous outputs.

Grammar-based constrained generation uses context-free grammars to restrict LLM token selection. You need it when generating structured output like formatted code or data objects, ensuring deterministic results and preventing formatting inconsistencies.

Yes, it supports multi-step workflows alongside local and API model backends. This allows you to chain constrained generation tasks and validate structured outputs across flexible model integrations within your existing environment.

You apply regex constraints during the generation process to enforce specific patterns for structured output validation. This ensures LLM responses match predefined rules, while token healing fixes tokenization artifacts to maintain pattern integrity.

The best way to fix tokenization errors is using token healing, which adjusts boundary tokens during constrained generation. This corrects mismatches between regex grammar rules and the LLM's token vocabulary, ensuring valid structured output.

You need the transformers and guidance libraries installed to use grammar-based generation. These dependencies provide the necessary environment for applying context-free grammars and regex constraints to local models.