About making inexpensive models smarter by providing tools and context. (local models, gpt-5-mini, gpt-4.1-mini, gpt-4o-mini ...)

Ok, a few small updates to tools, prompt and so on:

  • The Entity Index tool often searched for “Inside” even when I asked for things in the garden or something like that. Most of the time it tried with ‘Outside’ after that, but this wastes tokens and time.
    So I added this to the prompt section for this tool (also updated in the post about this tool above):

    ALWAYS think first if you have to search for entities Inside / Outside the house (location parameter of the tool). Try to derive this information form the users question.
    Most of the time it will be inside. Use ‘Everywhere’ if you really don’t know if the user asked about an Inside or Outside location.

  • I finally switched completely to gpt-oss-120b due to way better latency / response time, after I found a fast provider and a LLM integration that works without flaws with it:
    Fast (low latency) cloud models mit OpenAI compatible endpoint: Which do you use and recommend?

  • We often used the commands “lower the volume”, “volume up” or “3 times louder” with Alexa. Voice Assist behaved quite differently every time we used these commands, as they weren’t clearly defined.
    So I added this to my prompt (also added to the "Get better than Alexa when it comes to music control post):

    When we don’t specify how much the volume should be changed, use HassSetVolumeRelative with volume_step set exactly to 10 or -10.
    THIS IS IMPORTANT: Always change the volume by setting the volume_step as a numeric value.
    Sometimes we say things like “2 times louder” oder “4 times louder”.
    This has to be interpreted as a multiplier to the default volume_step of 10.
    So “4 times louder” would be an increase of 4*10 = 40.

2 Likes