I did most of it myself. I asked mistral small parts on how to do very specific things. Most of the time the result was inaccurate though as I think that model wasn’t trained with recent changes to home assistants code-base. Though, I also didn’t use codestral… maybe the result would have been better with the codestral model.
Currently I am struggling with how I could get function calling flexible enough for end-users to easily implement it. Function calling itself seems to be pretty easy to implement, but I would like for it to be as simple as possible for end-users to use it… but I am pretty sure that would still require people to at least perform two steps:
Create a file where functions are described
Implement the function in Python
I could add some generic functions ready for re-use, but the true power of function-calling will only happen when feeding the ai with your custom use-cases and sensor info.
I’m sure I could get function calling itself done pretty easily in a basic version. Can’t wait to try it. Due to the current holidays I don’t have as much time to develop it though
Just invested some time and realized that function calls are possible for sure, but I can’t figure out a generic way to handle em to be comfortable for home assistant users without them having to modify the code.
I’ll prepare the code-base to be able to support function calls, but the actual implementation will have to be done by the user.
Bummer, but maybe we can figure out a more flexible way over time.
I also noticed just now that mistral ai got their own ready-to-use python lib, still, I don’t see a reason to switch right now.
awsome integration.
Can you use the integration to generate AI images for download, similar to action openai_conversation.generate_image from the OpenAI integration?
Hello there, I’m quite interested into having Mistral as an available assistant. I would like to use it along the Home Assistant Voice. I like Mistral is quite fast and it’s European. Is there any way I can help you with the development @BlaXun?
I have compiled my edits into a separate fork, since I prefer to use the existing openAI system package rather than installing another one
I have also added an image recognition service openai_compatible_conversation.mistral_vision and a search agent openai_compatible_conversation.web_search for Mistral.
same, works great. Also LLM Vision works well, one minor change needed there as the Mistral AI implementation of the OpenAI API protocol is slightly off