guidance

Regex and grammar constraints enforce valid JSON and XML generation by restricting token selection to only syntactically correct sequences. This approach guarantees structured outputs without relying on post-generation validation or retry logic.

Token healing is a technique that automatically fixes boundary token artifacts during constrained generation. It adjusts the prompt's final token to ensure smooth regex and grammar enforcement without breaking the intended text format.

You can build multi-step LLM workflows using Pythonic context managers and functions to orchestrate grammar constraints and tool-assisted prompts. This allows you to compose reusable generation patterns across diverse backends and local models for complex tasks.

Yes, regex and CFG grammar constraints are supported with multiple backends, including local models. This allows you to enforce structured, valid results directly on local hardware without relying on external API providers.

The best way to prevent invalid code syntax is applying context-free grammar constraints during inference. This restricts the LLM to only generate valid code structures, eliminating syntax errors before they occur.