Use cheaper LLM model for translation need detection

[ChristophWurst]

Is your feature request related to a problem? Please describe.

As a Nextcloud admin, I want more efficient detection of the language to keep costs low. Right now the detection of the language can eat lots of tokens quickly.

Describe the solution you'd like

Try to use a cheap model for this task, if possible. Lower quality is absolutely fine for this job.

Describe alternatives you've considered

No response

Additional context

No response

[AnAnalogGuy]

Why using a LLM at all? Even a light-weighted LLM is way oversized, brings complexity and load where it's not needed.

Language detection in most case can be done by (a) using existing meta data or (b) a simple algo. I.e. you could count letters and compare their distribution (each language has its own fingerprint) or you count n-grams. Works good for everything else than very short texts. You could also use stop lists from search engines which are language depended. And count the hits. Unless you're affiliated with chip makers, please do us a favour and go for the simple solution.

[AnAnalogGuy]

Btw, "As a Nextcloud admin, I want efficient LLM usage because it's expensive. Right now the detection of the language can eat lots of tokens quickly." isn't a well formulated user story because you formulate the soltion as a requirement. Should rather be something like: "As a Nextcloud admin, I want more efficient detection of the language to keep costs low."

[ChristophWurst]

All LLM features are optional. If you don't want them, don't turn them on.

The more simple detection sounds good. Do you have references where that is used?

I've updated the user story. Thanks!

[ChristophWurst]

https://www.npmjs.com/package/franc-min could be an option. Or one day https://developer.mozilla.org/en-US/docs/Web/API/Translator_and_Language_Detector_APIs/Using.