Found 1 changed notebook. Review the changes at https://app.gitnotebooks.com/AnswerDotAI/claudette/pull/67

I think this does an either-or. If we pass both some tools and some tools_serialized it'll only get tools_serialized I think... Maybe something like this:

if tools or tools_serialized: kwargs['tools'] = [get_schema(o) for o in listify(tools)] + listify(tools_serialized)

Not sure what we want or if I'm right, just flagging before I forget!

In most cases if both are passed in there's probably a bug elsewhere so one solution is to encode that invariant (do something like only having a tools argument and faking type dispatch by doing isinstance(tools, List[str]) vs isinstance(tools, List[Dict]) and i except that isinstance doesn't play well with generics.).

Keeping this as is, I think your suggestion is a reasonable behavior for that case.

Thanks @erikgaas -- nice idea! How about, instead of having a separate param, we instead treat any str tools as pre-serialized?

Also, could you see whether we can use reclm from @comhar to have cached AI outputs, so that we don't have such big diffs in the future?

Thanks @erikgaas -- nice idea! How about, instead of having a separate param, we instead treat any str tools as pre-serialized?

Also, could you see whether we can use reclm from @comhar to have cached AI outputs, so that we don't have such big diffs in the future?

Ooo good idea.

@jph00 I tried reclm but there is an issue with streaming responses. I want to go into reclm and fix this. I'm hoping to just make it work for streaming, but I also think it would be good to have a disable_reclm function too.

As for your suggestion above, what do you think about serializing if the tool is callable else passing it through?
The other way of doing it would be if dict pass through else run schema.

@jph00 I tried reclm but there is an issue with streaming responses. I want to go into reclm and fix this. I'm hoping to just make it work for streaming, but I also think it would be good to have a disable_reclm function too.

As for your suggestion above, what do you think about serializing if the tool is callable else passing it through? The other way of doing it would be if dict pass through else run schema.

Adding streaming support to reclm is proving to be a little tricky. Rather than rushing a solution, let's hold off on integrating reclm into claudette for now.

With b0e0e66 + holding off on reclm for now, can we merge this?

I think it's ready to merge. Will add reclm in a followup if I'm the first to it.