Using it for a RAG is smart indeed, especially with a multimodal encoder (vision-rag), as the implementation would be straightforward from what you already have.