Epic: Compositional OpenAI components for completion, streaming, tools calls, and structured outputs

[EvanBoyle]

Related issues:

• Improve support for structured outputs #163
• Support Tools #127
• Support for Structured Output from LLM #54

We need to support structured outputs, streaming, tools, and everything exposed via the completions API. But we want to follow a few principles:

1. Transparency of the underlying API: don't hide or simplify the underlying OpenAI API types - get access to the raw ChatCompletion input and outputs so that you don't outgrow the abstraction, miss features, etc.
2. Separation of concerns through composition: separate layers like tools, structured outputs, and providers so that they can be used independently or together.
3. Progressive disclosure of complexity: simple use cases like sync or streaming completions are simple. complex features like tools are opt in.
4. Escape hatches all the way down: If you outgrow any part of the abstraction, you should be able to easily replace it, and continue using the useful bits it is composed of independently.
5. Right level of abstraction: Handle tedious parts like translating tools, but don't be too smart or make too many assumptions.
6. Lean core OpenAI compatability: stay away from non-standard APIs like response format and auto-running tools that might not play well with other providers.

In summary, the abstraction should give structure and eliminate boilerplate, but stay out of the way and let you evolve patterns over time without topping out.

TODO:

• Compositional component that supports:
  • Streaming
  • Tools
  • Structured outputs
  • Land on names
• Detect and fork streams to observe outputs for checkpointing #220
• Customize the way that tools are rendered in checkpoints (custom toJSON to show the openAPI schema, remove all of the zod native functions) #221
• Add dedicated gsx.workflow top level to execute in new contexts
• Consider pattern for gsx.workflow("name", <Component>).run();
• Update gsx.execute to throw error when called outside of a workflow context / component #222
• Add lightweight generateText method to replace the current ChatCompletion
• Update examples to use generateText
• Add examples for composition component
  • Streaming
  • Structured output
  • Tools
• Update OpenAI package docs
• Tests