I think you’re under the impression that the difference between those things is far greater than it actually is. Large Language Models work by developing statistical maps of associations. That’s discrimination. They’re a direct evolution of categorization models. The ability to associate a hash of a JPEG with “cat” is the same as the ability to associate “How are you?” with “Great, how about you?” It’s all associative mapping. LLMs are just the current leading edge of that technology. If you want to, for example, generate a list of tags that describe a document, an LLM is the best tool we currently have for doing that.
To put it another way, what you term “discriminative AI” is “generative AI.” It’s generating a category or list of categories in response to an input. That’s not functionally different than generating a sentence in response to a sentence, it’s just an order of magnitude less complex. You can argue terminology but the technology exists on an evolutionary curve, with no real hard boundaries.




Absolutely there’s a difference. LLMs, when it comes to this specific task, are better. That’s why they’re being used here. It’s a job they are uniquely well suited to. They do indeed come with high hardware requirements, which is why you’re not forced to use them, and why they provide the option to offload the work to a cloud service.
Personally I would absolutely not want to ever feed my documents into an off-device model, but the point of self-hosted software is that it does what you tell it to and they absolutely should include letting you make bad decisions.