[nl]

You will need very powerful hardware to deploy the deep learning models on-prem for incremental learning.

This isn't true.

I've build (neural-network) vector based search extensions for search. You don't train the model - you use a pretrained model (that understands English in your domain) and then use it as an encoder.

Sometimes there is once-off pretraining process for domain adaptation, but honestly this isn't a big deal. Even on a CPU based machine you could do this overnight or over a weekend, and since it is once off that time doesn't really matter.