Turn off Kitchen Lights.
Turned off the light
Turn them back on.
It looks like the 'HassTurnOn' service can't be used to turn on multiple devices at once. Can you please specify which device or area you would like to turn back on?
If there is only one device of a specific type, I can try to find it and control it for you. If there are multiple devices, I'll ask you to specify the area they're in so I can target them correctly.```
How do I fix this problem?#After adding Llama 3.2 to HA as the conversation agent it isnt able to keep context. Example:
boreal atlas
limber jackal
Increase context window and make sure you have required minimum of entities exposed to Assist.
Which model do you use? It doesn't make sense using less than 14B. At least 7B is capable only for generic answers, but is too dumb to use tools.
jade lotus
Agree, 14b even struggles enough to be a pain.
small cove
Yeah, for very basic things like device control, with enough context and smarts it works OK. I've mainly seen Qwen2.5 7b/14b do the best here, but they still mess up from time to time enough to be unreliable/annoying as compared to something like gpt-4o-mini which I haven't had mess up at all, or at least not enough to annoy me. When you toss in more advanced things like calendar query, weather, or requesting music, it just starts falling apart pretty fast.
Something I have started looking at is the project mentioned a few newsletters ago: https://github.com/goruck/home-generative-agent?tab=readme-ov-file
It seems to take the Agentic AI approach of using different models/LLMs for various tasks, with a large LLM as the coordinator (in this case they are able to run a local 32b variant of Qwen which probably is much more accurate than the 7b/14b models as the coordinator, but also support using something like GPT-4o/4o-mini) and combine that with agents like llama-3.2-Vision:11b for camera analysis and Qwen2.5:3b for context summarization, along with mxbai-embed-large for embeddings (used in this addon for RAG semantic search to ask stuff about the house).
small cove
Hm though after trying to set up the above I ran into a lot of problems. Seems cool but still needs some work it seems 😦
jade lotus
Maybe we need an LLM trained specifically for home assistant.