Are floors exposed to Assist?

[stone magnet]

I have several floors with several areas defined. However, it seems Assist (ChatGPT) has no idea about my floors. I'm confident this has worked for me before.

Are floors exposed to Assist? Is there a setting for it?

Examples:

On requests like "turn off all lights in 1st floor", it leaves out some rooms (e.g. the bedroom). Same if I ask which rooms belong to the 1st floor. When I ask where the bedroom is, it says 2nd floor. These answers are consistent, but wrong.

I also have an "outside" floor, which consists of the garage, pool house and some more. "On what floor is the pool house?" - "The pool house is in the pool house".

[hallow fern]

Floors are definitely exposed to Assist, as well as areas.

But ChatGPT isn't Assist

You're confusing "Assist" with "conversation agent", apparently.

[stone magnet]

I might have mixed up these two, but I don't see how it really matters regarding my trouble. As the home assistant GUI says, the conversation agent is the "brain" of Assist. So surely if I set up Assist to use chatGPT as conversation agent, chatGPT should be provided with all the info that's exposed to Assist?

[stone magnet]

Screenshots to show "Schlafzimmer" (bedroom) is assigned to "Erdgeschoss" (ground floor), while chatGPT ANF google generative ai say it's in "uppermost floor" / 1st floor.

Obergeschoss (1st floor) only has the "OG" area.

[regal rune]

There's no intent for getting the floor of an area and i doubt you'll ever ask that in a non-test environment. As such, depending on what model you use and how many entities you have exposed (i.e. what size the context is), a LLM just might be dumb enough not to be able to correlate the information.

Floor are exposed to Assist. A LLM-based conversation agent should be able to understand things like "turn off the lights on the ground floor". If you find anything that doesn't work, please provide an example of such a phrase. Asking what floor an area is on is not supported, though.

[stone magnet]

I'm sorry, I seem to be completely wrong in my expectations of what LLM-based conversation agents should be capable of, then?

Of course I wouldn't ask what floor an area is on. The wrong answers the LLM gives, however, are consistent with its failures in acting on my actual input.

For example, I have controllable lights in the living room (Wohnzimmer), hallway (Flur) and bedroom (Schlafzimmer). All in the ground floor (Erdgeschoss), especially the sleeping room as depicted in a screenshot above.

"Turn off all lights on the ground floor"

• "lights in hallway and living room have been turned off. Couldn't find the light in the kitchen because it's currently unavailable"

(There's no light in the kitchen connected to HA)

"Why did you not turn off the light in the bedroom?"

• "because you asked for the lights on the ground floor"

[hallow fern]

It looks like one of three:

1. Regular LLM hallucination.
2. Not enough of context window to hold your entities/areas/floors information there.
3. Some trouble in HA mechanism of sending info to LLM during request.

[stone magnet]

I'll have to accept that, then. I'm lacking time and skills to dig deeper into this.
This is not at all intended as a complaint or something like that. I'm impressed by all that's possible with HA.
Thanks for your time!