Guidance

Guidance is a domain-specific programming paradigm designed to solve the inherent unpredictability of Large Language Models (LLMs). By treating LLM interactions not as simple text-in/text-out prompts, but as a stateful execution of a template, Guidance allows developers to interleave generation, prompting, and control logic seamlessly. Its technical architecture relies on a specialized interpreter that can force the model to follow specific grammars, such as regular expressions or JSON schemas, at the token level. This prevents the model from generating invalid syntax or hallucinating structural elements, significantly reducing the need for post-generation validation or retry loops. In the 2026 landscape, Guidance serves as a critical infrastructure layer for 'Deterministic AI Agents,' bridging the gap between stochastic model outputs and strict software engineering requirements. It supports multiple backends including OpenAI, Hugging Face, and Llama.cpp, and utilizes advanced features like 'token healing' to eliminate common tokenization artifacts that degrade model performance at the start of generated strings.