The 5.2 notes say support for OpenAI and OpenAI-compatible providers for the new “LLM chat” is coming “soon”. Is anyone already working on this, or is it planned for a specific release?
I’m keen on it for an EU/GDPR setup and/or a generic provider with a configurable base URL. That would then also cover OpenAI, OpenRouter and Ollama in one. My personal go-to would be to point the HLI at Mistral’s OpenAI-compatible API.
If no one has picked it up yet, I’d be glad to help or contribute.
Nice to see this landing in OH 5.2, thanks a lot for all the good work.
I have one of the first ESP32-S3-BOX and have been “toying” with it to see what it can do, mostly with the HeyWillow framework.
The openHAB support in there is very simplistic as it was made around the OH4.0 release time.
As I get STT from Willow, I believe I should be able to better make use of it with OH5 but I’m not sure where to restart this endeavor. Well, first thing is to get a 5.2 instance up and running, but after that, I’m a bit stumped.
I plan to do it for the 5.3 release, with the goal to complete it not just a few weeks before the release but to post it to the marketplace for 5.2.
I haven’t started working on it yet and won’t for the next few weeks … so …
… feel free to contribute it yourself. I think it should be quite simple, from the current capabilities in the ChatGPT binding, all DTOs and API interaction should already be implemented.
If that’s really the case, the major work would be to update the ChatGPTHLIService to implement the specific interpret method that passes tools, conversation, system prompt etc.
You can take a look at GeminiHLIService as a reference implementation. Most of the LLM chat feature is implemented in core, add-ons only need to glue the core APIs and the LLM provider APIs together.
As I’ve discovered that my unversity offers a selection of open-source models hosted locally and even allows API usage of these models, I got a inference to test open-source models on. This gave me enough motivation to adapt the ChatGPT binding right now to work with the LLM-based HLI infrastructure:
