It sounds like most of the heavy lifting is mimicking OpenAI's API. Are LLMs generic enough that they're otherwise interchangeable? So long as we're using a model that isn't wrapped in a bespoke API, we can just swap models?

That's pretty cool and I guess that's the benefit on natural language being the "raw" interface to all these models. For embeddings it seems a little trickier - the models tend to truncate differently afaict and the output vector can vary in the number of dimensions.