Upcoming breaking change to function calling

[ahyatt]

I'm not aware of any package that currently uses llm's function calling, but #133 makes several breaking changes, notably changing the name "function calling" to "tool use", which is what the industry has settled on. It also changes somewhat how we create the tool to use, standardizing on the same basic format as GPTel.

@s-kostyaev FYI, please let me know if this is disruptive to you before I merge this in.

See #124 for background on this change.

[meain]

Glad to see a unified interface for defining tools! :)

[s-kostyaev]

Thank you for noticing me, Andrew. This LGTM.

Also can we introduce custom variable to add models in user configuration? For now we have only llm-models constant and users should wait for next release of llm library to experiment with new model if user need some capabilities. It will be useful for my experiments also. Related PR #135

[s-kostyaev]

I will see, thank you

[ahyatt]

Also I've been thinking of actually turning the JSON response into an elisp object before returning it. What do you think?

[s-kostyaev]

Maybe as an option? With json decoding I can prefer one elisp type or another. I think we can return json (as option) or we need to provide this flags to top level (like :object-type or :array-type)

[s-kostyaev]

We can now change default behavior, but I want it to be tweakable

[ahyatt]

After thinking about this some, I think that I'll leave it as text for now, so the consumer can detect if it is even json (sometimes it won't be!) and if so, convert it however they want it.

[ultronozm]

Thanks. I've been using tools via llm in my personal package https://github.com/ultronozm/ai-org-chat.el since last August. Updating to #133 went smoothly enough.

[ahyatt]

Interesting package! Note that likely #133 will get more breaking changes before it is merged in, so I guess when it is merged in then you can do a final pass at updating according to the final design.

[swapneils]

Curious if tool use is stable yet? I see the README still mentions it's in Beta.

[ahyatt]

It's more stable; I wouldn't expect large changes, but I wouldn't be surprised if some breaking changes happen; one just happened a few weeks ago on how false values are propagated to elisp functions. But it might be time to remove that warning.