Guidance programs improve the control of language models for better performance. They allow for interleaving generation, prompting, and logical control for a continuous flow. Using powerful language models like GPT-4, guidance programs make structure easier and cheaper. Guidance has features like simple syntax, rich output structure, streaming in Jupyter/VSCode Notebooks, and smart seed-based generation caching. It also supports role-based chat models and easy integration with Hugging Face models. To use Guidance, install it with pip install guidance. Live streaming in Jupyter Notebooks allows for faster prompt development cycles. Guidance provides well-defined linear execution order that produces clear and parsable results. Chat dialog is also supported using API-based chat models and open chat models. Overall, Guidance improves language model performance and control in a cost-effective way.
