Agents and automation
Last updated
A morning brief from Home Assistant: turning sensor data into sentences with a local LLM
The quality of the brief depends less on the model than on which values you put into the sentence.
A local LLM1 does not read live sensor values by itself. An automation therefore builds a sentence from sensor and calendar values, hands it to conversation.process, and sends the reply as a notification. This guide covers which values to include and how to handle failure.
The model does not know the current state
A conversation model sees only what the sentence contains. It cannot know whether the room is 26 or 19 degrees, or whether a window is open. So a request for a summary has to carry the checked values inside the sentence.
Home Assistant's conversation.process hands that sentence to the chosen conversation agent and returns the reply. An automation can take that reply into a variable and use it as the notification text.
Choosing the values to include
Five to eight values are enough: indoor temperature and humidity, yesterday's power use, open doors or windows, the first event of the day, air quality — only what you would use to judge the day.
Attach units to numbers and drop sensors that have no value. A blank invites the model to guess and produce an odd sentence.
| Value | Where it comes from | Watch out |
|---|---|---|
| Indoor temperature and humidity | Temperature and humidity entities | Include the unit |
| Yesterday's power use | Cumulative value from a power sensor | Say whether it is a total or an instant |
| Open doors and windows | Door and window entities | Leave it out when absent |
| First event today | Calendar entity | Keep the time and title only |
Wiring it into an automation
Trigger on time, at the hour the household starts, such as 07:00. Add an action that builds the sentence, then conversation.process. When the reply comes back, send it as a notification.
Keep the sentence short. Putting the instruction and the values in one paragraph, such as 'use these values and tell me in three sentences what to prepare today', keeps the format steadier.
Instructions that keep the answer short
Set the sentence count and format in the agent's instructions: no greeting, three sentences, repeat the numbers as given, no lists. That suits a notification.
Longer instructions delay the first response. Keeping a separate agent for summaries, apart from the conversational one, is easier to manage.
Deciding what happens on failure
If the model server is down or the reply is empty, no notification arrives. Add a condition so an empty reply sends a fixed line such as 'the summary could not be built; check the sensor values'.
A week of notifications shows which values go missing often. Drop the sensors that frequently vanish, or find one that stands in for them.
Terminology notes
Large language model — A model trained on large text datasets to learn language patterns and process or generate text in context.
Back to the text
Frequently asked questions
- What should you check first in this article?
- A local LLM does not read live sensor values by itself. An automation therefore builds a sentence from sensor and calendar values, hands it to conversation.process, and sends the reply as a notification. This guide covers which values to include and how to handle failure.