Reciprocal Rank Fusion vs Reranking
People think reciprocal rank fusion is a kind of reranking, and it is not. Fusion merges two or more ranked lists using only each result's place in each list, so it is fast and reads no text. A reranker is a model that reads the question with each result and scores it again. Fusion joins the lists. The reranker then puts the short list in a better order.
Reciprocal Rank Fusion (RRF)
A way to join ranked lists. Each place earns points.
Reciprocal rank fusion, or RRF, is a way to merge two or more ranked lists into one. Each list gives a result 1 divided by 60 plus its place. So first place earns 1/61, and tenth place earns 1/70. The points from every list are added, and the top total comes first. It looks only at places, not scores, so lists with different kinds of score can be merged. It was published in 2009.
Reranking
A second, slow look at the top hits. It puts them in a better order.
Reranking is a second pass that puts search results in a better order. A fast search first pulls, say, the top 50 candidates. Then a reranker model reads the question together with each candidate and gives it a new score. That is slower per text but much more accurate, so it is only run on the short list. The best few after reranking are the ones passed on to the user or the model.
In one line
- Reciprocal Rank Fusion (RRF)
- A way to join ranked lists. Each place earns points.
- Reranking
- A second, slow look at the top hits. It puts them in a better order.