guidance

To ensure valid JSON output from an LLM, use constrained generation with regex constraints and grammars to enforce exact formats, preventing malformed responses without writing retry logic.

Constrained generation for LLM workflows is the process of controlling model output using grammars, selection controls, and token healing to guarantee valid structured results like JSON or XML.

Yes, you can use constrained generation with both local models and API models, applying the same constraint-driven prompting approach across different backends to maintain output consistency.

You can build multi-step LLM workflows with decision branches by composing prompts and defining reusable guidance functions to create agents, extraction, and classification pipelines.

Token healing in structured output generation is a control mechanism that repairs boundary tokens to ensure the final text strictly adheres to the specified regex or grammar constraints.

You can extract entities from text without retry logic by applying constrained generation to enforce the exact output format, eliminating the need to fix malformed responses after generation.