Naive RAG pipelines fail in production; seven architectures address specific retrieval failures

A DEV Community article argues that the standard retrieve-and-stuff approach to retrieval-augmented generation works in demos but breaks on real questions, often pulling wrong or incomplete chunks. The author identifies recurring failure modes, including vocabulary mismatch, lost context from torn-out chunks, mismatched question and answer phrasing, and questions whose answers span whole documents. The piece presents seven distinct architectures, each designed to address one or two of these failures rather than being a single universal upgrade. It recommends hybrid search combining vector and keyword retrieval plus reranking as a baseline improvement before adopting more complex designs. Query rewriting techniques such as multi-query and HyDE are also covered as ways to bridge the gap between short questions and long answer passages.