generative-computing/mellea ? reverse-engineered prompt

Build me a Python library for writing reliable AI workflows, not just messy prompts.

I want a simple way to turn a typed Python function into an LLM call, where the docstring becomes the instruction and the return type is enforced as structured output. It should support Pydantic models, automatic validation, and retries when the model gives something invalid. I also want a session object with sensible defaults so people can start quickly and override settings when needed.

Please include a clean public API, support for a few model backends like OpenAI and local models, and a way to run multiple generations and pick the best answer. It would be great if existing code could be adapted easily too, and if the library could expose a tool interface for agent style use.

Use good tests and a few runnable examples so it’s easy to understand. If you need to check current docs or backend details online, go ahead.