guidance

You constrain LLM outputs to generate valid JSON and XML by applying regex and grammar-based constraints during generation. This enforces structured formats directly, ensuring predictable results with minimal post-processing.

Token healing is a constrained-generation technique that fixes boundary tokenization errors during LLM inference. It reduces invalid structured outputs by enforcing selection constraints and repairing token boundaries to maintain format integrity.

You build multi-step workflows for data extraction with constrained generation by orchestrating Pythonic control flows. This workflow chains data extraction, JSON generation, and tool-augmented reasoning steps while enforcing format constraints throughout.

Yes, constrained generation works with both local and API-backed models. The framework enforces grammar and regex constraints across different model deployments to ensure data integrity and structured outputs.

You can validate structured LLM outputs without leaking prompts by loading constraints via frontmatter and enforcing structured formats separately. The validation process checks regex and grammar compliance while keeping the original prompts secure.

The best way to enforce a predefined schema when generating code from an LLM is using grammar-based constraints. Grammars restrict the generation space to valid syntax, ensuring the output code adheres to required structural definitions without relying on post-processing.