RAG: Measurements Matter More Than Code. The Story of a Bot Where 'Correct' Improvements Worsened Results
A developer built a RAG bot for moderating crypto chats but found that standard improvements—BM25 and a reranker—only degraded metrics. The only change that worked was calibrating the cutoff threshold. The article emphasizes the importance of careful measurements on one's own data.
Habr — хаб ИИ27.07 · 09:03