Databricks. I’d rather just use a local server with spark at this rate.

I think you’re looking at the problem wrong by focusing just on the LLM aspect of it. If you’re deploying any type of application then it will depend on the demand you’re expecting from users and the use cases the application will be used. A failure rate for medical applications is probably a lot more important than a low budget game service.

It seems like a bad reason to use it. It would probably be better to just do a more “normal” type of search rather than trying to hit the problem over the head a more complex model. Occam’s Razor and all.