Why Ranking Fixes Failed: The Real Problem Was Document Selection

A developer spent three months tweaking ranking algorithms—BM25, hybrid search, cross-encoders, and embedding models—but none improved answer quality. The root cause was that the key document never entered the candidate set, so ranking could only reorder irrelevant results. The author reframes retrieval as three layers: selection, assembly, and ranking, noting that optimizing the last step is futile if earlier ones fail.