Function calling / tool use

[Boscop]

It would be really great if you could prioritize adding support for function calling / tool use :)
I started using this crate for a project where I need function calling, and then realized that this crate doesn't support it yet..
For now I'll emulate it by telling the model in the system prompt to output json in a certain format whenever it wants to "call a function", but it's a hacky workaround.

Please let me know, what's your planned timeline for adding function calling? :)

Thanks 🙏
I really appreciate your work on this crate 🙂

Btw, in general it would be very helpful to have an intended Roadmap in the Readme, or have a GitHub project that implies one :)

[jeremychone]

@Boscop Yes, make sense. Ok, this is next on the list now.

[Boscop]

Thanks so much 😊

[jeremychone]

@Boscop I did the first pass for OpenAI and Anthropic (see above for the commit). Not released yet, more work to be done, but just to let you know that this is in progress.

[jBernavaPrah]

Hi @jeremychone!

I would love to use your crate to simplify my project, but I will need at least the function calls/tools to be available. :)

What is the status and the timing to publish it and is there anyhow that I can help you? :)

Thanks a lot!

[jeremychone]

@jBernavaPrah A first pass at it has been implemented for openai adapter. You can check the test files.

[Boscop]

It would be great if function calling could be done in a nice typed way like https://github.com/frankfralick/openai-func-enums (but not just for OpenAI).

[jeremychone]

@Boscop Sorry for the late response. Here are a couple of points:

• Tool/function calling has been implemented for a while for OpenAI (and compatible APIs) and Anthropic.

• I am working on adding Gemini now.

• The reason why fully typed function calling is not part of the scope (and independent of supporting multiple AI models) is because I would like to keep it pretty open right now and allow some flexibility. In fact, it allows for a little bit more support, as it is less tightly bound to one way of thinking.

  • So, the Tool type takes schema as a JSON Value object and returns a JSON value.
  • One can use schemars if they want to have full types (definitely a very cool lib).
  • Eventually, perhaps, I might add some more "optional" types for convenience and model portability, but this is not part of the initial support.

Hope this answers the questions.

Have fun with tool support for OpenAI and Anthropic today; Gemini is coming soon.

I will also add some examples/...-tool.rs to better demonstrate how it works.

Happy coding!

[jeremychone]

@jBernavaPrah did you try the tool support?