For large or dynamic knowledge bases RAG is what we have tended to go with. It allows for quite a bit of scalability with little overhead.
CPT is good for building foundational knowledge around a subject, your internal industry information for example. Yes, you can apply RAG on top of CPT, in fact this is probably the most robust implementation. It shouldn't unlearn your CPT learned information, it will supplement with the new documents. There is risk of conflict if there is conflict in the documents but otherwise you should be good!
Hope that helps ๐
"All models are wrong, some are useful"