#How can I make assist call scripts with an LLM?

1 messages · Page 1 of 1 (latest)

marble egret
#

I have an exposed script but it seems the LLM can't see it, it doesn't help that I can not find the prompt homeassistant sends to the LLM anywhere in homeassistant, I also for some reason can only use Google: Gemini Flash 2.0 Experimental, I've tried many different free LLMs on openrouter including llama and deepseek v3 (not r1 and r1), and I get the error "Sorry, I had a problem talking to OpenAI". The integration I am using is "OpenAI Compatible Conversation" which is the same as OpenAI conversation but allows changing the base URL

zealous cove
#

Maybe openrouter doesn't support tooling? Just a wild guess.
I use Ollama and it works good.

marble egret
#

I think most models don't but the Gemini flash experimental does, even then it still can't see scripts

#

But why is tool use required

#

Can I make it run without tool use and just have it detect words in the response (since isn't that what tool use is anyway) I think most LLMs are smart enough to format their messages correctly

#

However the not calling scripts issue isn't due to not supporting tool use as the one LLM that does work (and can turn lights on and off etc...) can not see the script

marble egret
#

I am now using a local LLM (home-v3:3b) and it says this but nothing happens

image.png
# 
"text": "executing command: \"script.clean__bedroom\"\nhomeassistant\n{\"service\": \"script.run\", \"target_device\": \"script.clean__bedroom\"}\n",