Reciprocal Rank Fusion
Reciprocal Rank Fusion (RRF) is a method for combining several ranked result lists into one by scoring each result on its position in each list. Nicholas Hopper used it in Memector to fuse up to five search channels, and measured where it fails.
Why rank fusion and not score fusion
Memector searches meme templates through several channels: CLIP text-to-image similarity, a MiniLM text embedding, Postgres keyword search, and two sketch channels. Their scores are not comparable. CLIP cosine similarities bunch between about 0.2 and 0.3, the text embedding spreads across 0 to 1, and the keyword rank is unbounded. Nicholas Hopper chose RRF because it reads only rank order, so no normalisation needs retuning for each corpus.
He swept the RRF constant over five values and left it at 60, because the effect was small and not monotonic.
RRF punishes absence, not mediocrity
A correct result ranked first on two channels and 85th on a third. The third channel’s cutoff was 80, so the result scored zero there and lost to a result that was about fifth everywhere. Hopper’s note: a document outside a channel’s cutoff contributes nothing, not a small value.
Raising the candidate depth per channel from 80 to 150 moved recall@5 from 98% to 100%.
Results
On 50 vague queries that do not contain the template’s name, Memector measured recall@1 of 94% and recall@5 of 100%, with median latency around 100 ms.