#Best way to train the LLM?

1 messages · Page 1 of 1 (latest)

undone beacon
#

Hello! I'm using OpenAI within Home Assistant. It's pretty good, but every once in a while I'll say something and it simply won't understand what I want it to do. For example, when I say "The cat is outside" I would expect it to toggle the cat_is_outside boolean to true. Since I have children, the sentence that triggers that can't be just "The cat is outside" because they will then say something like "Tiger is out" and it will mean the same thing.

Is there a way that I can train the LLM / provide instructions so that it understands what these things mean?

daring surge
#

Add that to the prompt.

limpid oracle
# daring surge Add that to the prompt.

To add to this; By using OpenAI, you have a VERY large context available to you; A prompt has no real limit other than the context size.

Mine per instance is VERY detailed:

undone beacon
#

Wow I had no idea you could do that. I thought it was supposed to be as short as possible!

#

@limpid oracle when you're saying "Hass" in front of everything, what does that do in this context?

#

Also does your media player thing work? That would be a game changer, honestly. I was never able to get it to work with Spotify + Sonos (the old ones)

limpid oracle
undone beacon
#

@limpid oracle totally, that makes sense given what we're talking about 🙂 I was more thinking that I must have something set up wrong here. I didn't add anything to home assistant to define what those things should do.

#

After some googling, I see there's a repo that has a huge list of intents, but I'm not exactly sure if those are the ones that are built in or that's something I need to install.

limpid oracle
undone beacon
limpid oracle
#

That's the official
Open Home Foundation (Voice)
...

undone beacon
#

As simple as the question sounds, does that mean it's baked in? Or do I need to add those?

daring surge
#

They're built-in

candid loom
#

I would also suggest to add tings that it has to know (but that aren't easy to get by just the stuff that HA exposes) in the prompt.

Everything that it has to know about the home, the users and dependencies of entities.

But if it's just a second name for something, you can also add aliases to entities for assist in the view where you can also select for which voice assistants (alexa, assist) it should be exposed.

That might help with things like tiger / cat.
You can also add little descriptions / sentences as aliases when using an LLM, as it is not using the name as a defined string that you have to say to address the device.

So it's also a nice place to add additionals context to the device without placing it all in a large wall of text in the prompt (there will be enough text that isn't entity-specific when you start to use assist more and more ...

limpid oracle