Introduced on day 48
Reciprocal rank fusion
A way to merge several ranked lists by summing 1/(c + rank) for each document across lists, which ignores the retrievers' score scales.
In detail
Because fusion uses ranks, lexical scores and vector similarities need no calibration to each other. It discards score magnitude, so a very confident hit and a marginal one at the same rank count equally, and it should be checked per query slice before replacing the best single retriever.