Generic reranking cuts scientific RAG precision 38.7%

SciRet found that an MS MARCO-trained cross-encoder cut scientific RAG precision at five by 38.7%, from 0.600 to 0.368, over its 5,000-paper corpus. The new controlled study also measured a 32.7% drop at 1,000 papers, while BM25 plus BGE-M3 rank fusion reached perfect recall at ten in two of three corpus sizes. The evidence is deliberately narrow—15 queries, titles and abstracts only, with pseudo-relevance labels derived from the hybrid retriever—but builders should still benchmark domain transfer before adding a reranker, the same task-level discipline urged by the archive’s product-search benchmark analysis.