NathanCu
(Nathan Curtis)
July 3, 2025, 1:11pm
122
Where do I add all this Context you’re always yammering on about, Nathan?
Here: Is it possible to define aliases for ToDo items? - #5 by NathanCu
But all of the entties support this. Note every word you put in here eats toks for context. be descriptive but be aware of the result. This will directly impact the context size and either be concise or use a model with a gigantic context window and be ready to burn toks.
NathanCu
(Nathan Curtis)
July 3, 2025, 3:59pm
123
Workflow Ala HA voice…
Agree at least with feentrant scripts and event triggers you can In theory both have a script call out and ask for clarification if said info isn’t in (condition X) and we can now ask a question and receive answers with clarified slots. You could then restart the script with the correct data which would skip the question etc…
Workflow Ala HA voice.
This is now technically possible in 2025.7.x. Someone please write a good example?
NathanCu
(Nathan Curtis)
July 3, 2025, 6:23pm
124
In today’s episode of
wierd stuff we haven’t confirmed but sure seems to be the case from the data…
Are we sure the device id is passed in the entdump into context?
Im pretty sure it’s the net result of the ent list for each:
[image]
or clearer:
{{states['light.back_hall_light']}}
Results:
<template TemplateState(<state light.back_hall_light=off; supported_color_modes=[<ColorMode.BRIGHTNESS: 'brightness'>], color_mode=None, brightness=None, icon=mdi:dome-light, friendly_name=Back Hall Light , supported_features=32 @ 2025-07-02T15:43:05.787523-05:00>)>
Means LLM Context: (with alias dat…
Essentially what this means, HA’s default context does include name, alias, and entity_id but does NOT include device_id.
Given context crafting rules that means
until I create context for her to understand device ID IT DOES NOT EXIST in the Assist context space…
…therefore targeting by device ID without building your own resolution process for an LLM won’t work in most cases.
I believe it’s deliberately not possible to control via device_ID through Assist. In fact, it seems that they don’t want us doing it that way at all, but haven’t yet removed the ability to do so.
emory
(Emory)
July 3, 2025, 11:43pm
126
it’s a very lively summer and we had quite an adventure with an oven throwing fireballs and then thank g-d for carbon monoxide detectors and homepod sound detection. i will be catching up soon, i’ve just been in a funnel. <4
1 Like
NathanCu
(Nathan Curtis)
July 4, 2025, 12:58am
127
Sounds like events you didn’t want. Hope all is ok.
NathanCu
(Nathan Curtis)
July 18, 2025, 3:38pm
128
Notes on Searching with OpenAI
You’re correct and was about to say the same model matersy… I believe GPT4.1-mini and maybe nano support it as well.
In fact I am having HUGE success with both 4.1-mini and 4.1-nano (and price/perf better than GPT4.0)
Op. - don’t run a full reasoning model like o.1, o.3, o.3 pro etc. unless you REALLY like burning cash. I’ve not found anything that o.whatever can do as far as HA is concerned where you need it. Mini or nano are your best bets.
Edit 2: your llm will search for answers if it do…
If you turn on web search in your llm provider, make sure you tell it how to behave…
(I found a bug in the files crud handler had to fix before posting. Probably tomorrow. This will do till then.)