I have Home Assistant Assist running locally through Ollama with qwen3.5:9b-q8_0, and the basic conversational side is now working well enough that I have started looking beyond device commands.
The question I want the lab to answer is:
Anything I should know?
Home Assistant already has the state, history, statistics, presence information and service telemetry. What I am experimenting with now is an architecture where HA first qualifies potentially relevant conditions, supplies only the necessary context to the local model, and lets the model interpret whether something deserves attention, an explanation, a question or no interruption at all.
So rather than feeding the LLM the whole house, the flow is roughly:
HA state/history → condition qualification → relevant context → local LLM → explain / ask / bounded action → feedback
Some examples I want it to handle are:
“Anything unusual happen overnight?”
“Why is power consumption higher today?”
“What changed in the lab since yesterday?”
And proactively, something like detecting that I have arrived home after dark and asking whether I want the lights on rather than blindly executing an automation.
The part that surprised me while working through this was where the same architecture led when I removed the smart home entirely and applied the underlying control idea to a completely different operating environment.
That became the second half of the experiment.
Can a Smart Home Teach a Banking Database What Matters?
I would be particularly interested in how others running local LLMs with HA are handling history, relevance filtering and proactive interaction. I am less interested in giving the model more entities than in giving it better evidence.
