I have the OpenAI integration set up and working as I expect for the most part, but it seems like about half the time it simply can’t find the entity I am aiming for. I have an input button press entity I use for turning in the TV (with about a dozen relevant alias’ set up for it) and Assist can turn the TV on…but then utterly fails moments later to be able to turn it off no matter what I say. Am I setting it up for failure somehow?
#Assist with ChatGPT seems to find the right entity…sometimes.
rough grotto
split vigil
Same issue here with Ollama and Llama 3.1.
north jay
It's context size, most probably. Check the number of entities exposed to the Assist. Try to reduce that to bare minimum.
bitter loom
(I'm using Llama 3.1 fwiw)
If you're still troubleshooting I've found it helpful so far to keep pressing it for more detail whenever it says "theres an issue" until it gives me specifics on the error and details on what actions it took. This is easier if you're doing it over the UI rather than voice obviously
split vigil
It's context size, most probably. Check the number of entities exposed to the As...
Already is, it's just lights and some media players. Context size is 8192.
Not only that, it's doing insane things. It just picked up some TV (despite it not sounding anything like the wakeword) and this happened...
north jay
Well, that's LLM hallucinations, main reason why I don't use LLM directly. Apart from that, the false-positives are still pretty strong. Do you use Microwakeword? Kevin updated ok_nabu model few days ago, it should be much better if you use that new model.