A single Postgres table with two indexes outscored every published RAG pipeline on MuSiQue and ranked third on BRIGHT against systems with fine-tuned retrievers. While the numbers are great, they're really just to prove that more is less when it comes to these systems.

Every architectural addition in a RAG pipeline is a bet that the model will keep needing that help. A planner because today's model can't decompose multi-hop questions. A knowledge graph because it confuses entities. A reranker because the top-10 ordering is wrong. Each solves a real failure mode against a specific model. Each becomes maintenance debt the moment that model improves, and improvement now arrives every few months.

That bet gets worse every model generation. Query decomposition, entity disambiguation, knowing when to search again, reformulating after a missed search. Work that used to need scaffolding now moves back into the model. The weakness vanishes in a release.…