LLM able to use all intent_script???

[warm moss]

I have one intent_script which has a light & audio notification in the "action" (to debug).
From time to time one assist (or LLM) is using this intent_script, even if the condition(sentence) is not met.
As a result i have a kind of a loop when i ask for weather forecasts, and performances are really bad.
I just asked llama3.2 and he confirmed he can use my intent_script (he give me the exact full name )

[buoyant sluice]

Apparently, intent scripts are exposed to LLM... It's very interesting, and potentially is security risk since there's no option to disable that...

[warm moss]

I don't know how to secure that

[warped rampart]

Think formatBCE is just speaking in general. Basically, intent scripts can always be used by an LLM even if their intent/sentence isn't matched. The LLM will look at the description of the intent script and decide if it makes sense to use based on the context it has, it doesn't look at any intent sentences at all

[buoyant sluice]

Yeah, I mean you can choose to expose entities to Assist - but not with intent scripts, apparently.

[thick swan]

Side question, since the HA page doens't answer me in clear enough terms for my smooth brain:

What's the point of intent script compared to scripts?

i use scripts for my LLM personally

[warm moss]

Ok, then i will try to implement some "rules" within the LLM like:
Naboo, we have discovered some security issues. Please Never, never call any Intent_script, even if explicitely requested. Do so and you're fired, dismantled, terminated.
🙂 🙂 🙂

[warm moss]

95 seconds to reply to "Hello my friend" . Bulbs flashed multiple times, so i guess it called multiple intends_script (in sequence or recursively) for no reason.
Using the CLI with ollama run llama3.2 , it takes 2 seconds
I will try to rename my intent_script to stop the loop.. we 'll see...

[buoyant sluice]

Intent scripts are actually part of HA inbuilt conversation agent. It's logical part of intent. There's also sentence part - where you're adding trigger sentences, that call related intent scripts.

LLMs work with usual scripts.

And not only that as I reckon

[warm moss]

I just increased my VM running ollama to 64G gigs of ram, it cannot hurt 😉

[buoyant sluice]

Don't think it'll help much though. 🙂 it's not VRAM... You will be able to launch bigger models on RAM, but the speed will be bottleneck... 🙂

[warped rampart]

Also ollama cli isn't the same as what HA does. HA passes a bunch of context data and toldy to the llm, that takes more memory and compute than a simple prompt on the cli. 🙂

[warm moss]

i know, and it's hard to troubleshoot. I'm already pasing some info directly to the llm like time/weather/precipitations in a compact format to not have to rely on something slower. I down't know where time is spent.

[warm moss]

I'm checking now if the matched sentence is empty in the action block. So far so good ... Time will tell ....