extended_openai_conversation
extended_openai_conversation copied to clipboard
jekalmin
0.1.x (Use tools instead of functions)
- Merged PR(https://github.com/jekalmin/extended_openai_conversation/pull/25) into 0.1.x for releasing 0.1.0-beta1
- Need to merge into main branch when API is stable
@rkistner
Maybe I will add Use Tools option in the next release (probably 1.0.0), so that I don't have to keep track of multiple codes.
Hi! if using tools instead of functions, does this prompt need to be changed or the functions need to removed? thanks
currently running the beta and would like to test. thanks
Thanks for your interest!
I released this in 1.0.2-beta1. You don't have to change prompt or functions.
Please try it and give a feedback.
Thank you!
I wanted to share that I've been experimenting with Beta1 using LocalAI.
With LocalAI, I have the Mixtral 8x7b (v 2.7) 6Q GGUF model setup, which is supposedly one of the best models out right now.
I pointed your integration to this model and then toggled the "use tools" button and pressed submit.
I was pleasantly surprised that by doing the above, it was able to read sensor information in my home assistant quite well with no errors.
However, when I asked it to turn on a light, it seems to have gone through the actions and acknowledged that it's on, but the fact is it didn't actually turn on the light.
So I think using tools is moving in the right direction. However, there might be some tweaks that are required for it to actually do service calls. Do you know if that's something that can be done in the prompt?
Thank you!
I also use the assist to help trobleshoot, it seems to get the correct entity and correct service, so not sure why its not working as its hitting the correct service and entitiy:
As always, thanks @Anto79-ops for your cooperation. Is there a log about how LLM called a function?
I can check if there's a way to look at the logs of LocalAI... while I send the command to turn off the light. Is that something that would be useful or do you need the home assistant integration logs?
fantastic. You'll be very suprised how well it works, Here is the model im using, Q6_K version
https://huggingface.co/TheBloke/dolphin-2.7-mixtral-8x7b-GGUF
its not a small model, so it may 20 to 40 seconds to reply if you don't have a decent CPU/GPU computer.
Let me know how it goes!
LocalAI does not support function calling right now, you need to instruct your model to generate functions and parse the output.
This integration relies on the response from the openai api having the is_function_call value set, localAI models are not trained to perform this. I am investigating integration with: https://github.com/MeetKai/functionary which combined with their special VLLM server seems to be promising in its responses - but it's weak at general responses, so you really need multiple models.. tricky tricky.
@ex10ded Thanks for your comments. Have you been able to get the functionary v2 gguf model to work with LocalAI? It seems to require a special chat template if you use gguf (not vLLM):
https://github.com/mudler/LocalAI/discussions/1641
@ex10ded Thanks for your comments. Have you been able to get the functionary v2 gguf model to work with LocalAI? It seems to require a special chat template if you use gguf (not vLLM):
No, I hit the same issue as you - the template does not seem to work when using anything other than their special vLLM server (not even standard vllm) - they seem to do a lot of pre-processing of the tools sent even before applying a template too.
I'm on 1.0.3 and for some reason it works when use tools off but when I turn it on it seems to REALLY want to use execute_services for getting states.. This is when using default model and prompt. Did anyone else notice this difference? It should be the same thing, right?
I got localAI working with this here addon, but the LLM seems to start looping on function call.
Anyone encountered this?
