Reciprocal rank fusion k=60 underperforms on BEIR
The author argues that using the default reciprocal rank fusion setting of k=60 often underperforms compared to single-retriever baselines or tuned configurations. Testing on two scientific BEIR datasets showed that smaller k values or weighted combinations frequently retained higher nDCG@10 scores. The article provides a Python implementation of RRF and compares built-in fusion options across popular search and database engines. Open article →