guidance

You can guarantee valid JSON generation by constraining LLM outputs using context-free grammars and regex patterns. This enforces structured continuations during generation, ensuring downstream systems reliably consume JSON, XML, or code-like structures without parsing errors.

Token healing is a mechanism that automatically fixes awkward token-boundary artifacts during generation. It improves correctness for structured continuations by ensuring proper spacing and identifiers, preventing malformed model outputs in constrained workflows.

You can force LLM outputs to match a specific regex pattern by applying constrained generation. This restricts fields to required formats such as emails, dates, IDs, and numeric patterns, preventing invalid text from breaking structured data extraction pipelines.

Yes, you can restrict LLM classification outputs using selection constraints. This limits generation to a fixed set of categories for classification and routing, ensuring the model only returns valid selectable options for multi-step workflows.

Guidance works with transformers to build multi-step workflow pipelines using Pythonic control flow and reusable functions. This dependency integration allows you to compose deterministic, multi-round pipelines that maintain guaranteed structural validity across varied inputs.

The best way to prevent malformed model outputs in data extraction is applying grammar-based structured generation. Using context-free grammars guarantees structural validity for nested JSON, XML, or code, ensuring formats remain reliable across varied inputs.