guidance

To guarantee valid JSON output from an LLM without retry loops, apply constrained generation using regex and grammar constraints. This enforces token-level syntax rules so generated fields always match required formats and eliminates brittle post-validation.

Token healing is a constrained generation mechanism that repairs token boundaries during LLM generation. It ensures structured outputs like JSON, XML, and code remain syntactically valid by preventing partial token mismatches from breaking the required format.

Constrained generation works with both local models and API backends. You can apply grammar constraints, selection constraints, and backend configuration across Anthropic, OpenAI, Transformers, and llama.cpp to enforce valid structured outputs consistently.

To extract structured entities from documents using LLMs, build Pythonic multi-step workflows with selection constraints. This forces the model to generate validated API payloads and structured data formats directly from the extracted text without parsing errors.

You can build stateful or stateless ReAct-style agents that call tools with validated inputs. This is achieved by applying constrained generation, token healing, and selection constraints within multi-step workflows to ensure inputs match required formats.

LLMs generate malformed JSON due to uncontrolled token generation. Prevent this by applying CFG grammars and token-level regex constraints during generation, which enforce well-formed structured outputs and eliminate downstream parsing errors.