guidance

You can constrain LLM outputs by applying regex and grammar rules to enforce predictable structures like JSON, XML, and code. This constrained generation ensures outputs reliably match your required data formats and schemas.

You validate LLM outputs against a JSON schema by applying grammar constraints and automatic validation during generation. This enforces schema compliance directly, ensuring the generated text matches your format requirements before completion.

Yes, Guidance works with local transformers models for constrained generation. It supports multi-step workflows and Pythonic control flow to orchestrate complex tasks across local transformers or API backends.

The best way to enforce regex constraints is using grammar compilation and token healing. This approach applies regex rules directly during the generation process to deliver reliable, low-latency, and verifiable text outputs.

LLM generated JSON objects contain syntax errors due to unconstrained token prediction. You prevent them by using constrained generation with grammar rules, which restricts the LLM to only produce valid, verifiable JSON structures.

Yes, you can use constrained generation for multi-step workflows across API backends. Guidance allows easy backend switching while applying grammar compilation and Pythonic control flow to orchestrate complex tasks reliably.