There's no particular requirement for a RAG application to use vector search or embeddings, and there's no requirement that semantic similarity is in play (i.e. "retriev[ing] documents...that do not contain the terms in the search query"). Fundamentally RAG is just doing some retrieval, and then doing some generation. While the traditional implementation definitely involves vector search and LLMs, there are plenty of other approaches; ultimately at anything beyond toy scale it sort of begins to converge with the long history of traditional search problems rather than being its own distinct thing.