Guidance

Guidance deserves a place in the database because it represents a very specific and still important design philosophy in LLM tooling: instead of treating prompts as plain text and post-processing the output later, it lets builders constrain generation while the model is producing tokens. The official docs emphasize constrained generation, regex and grammar control, token healing, and the ability to interleave generation with program logic. That gives Guidance a distinct place in the builder stack even as native structured-output APIs have improved.

The pricing story is open-source. Guidance is MIT-licensed and installable from PyPI, so the framework itself is free to use. The real economics come from the model backend you choose. If you run Guidance against OpenAI or another hosted provider, you still pay the provider. If you run it locally through llama.cpp or Transformers, your cost shifts toward infrastructure and hardware. That is the right editorial framing: Guidance is free software that can reduce waste and improve control, but it does not remove model-compute costs.