guidance

To guarantee LLM output matches a specific JSON schema, constrained generation enforces regex patterns and context-free grammars during decoding, ensuring valid structured data and reducing post-processing errors.

Yes, constrained generation supports local models via Transformers and llama.cpp, alongside OpenAI and Anthropic backends, allowing fine-grained control over LLM output across various execution environments.

You control LLM generation by applying Pythonic syntax, regular expressions, and context-free grammars to constrain the decoding process, forcing the model to produce outputs that strictly adhere to defined formats.

The best way to build multi-step conversational agents is using Pythonic control flow to manage complex reasoning chains, enforcing structured data output at each step for reliable agentic workflows.

LLMs generate invalid JSON when unconstrained, but applying context-free grammars during generation guarantees format compliance, ensuring outputs strictly match defined validation rules before being accepted.