We improved a production retrieval system by changing one integer. No new models, no new infrastructure, no new code beyond a single number. Here is the number, and why the default is quietly wrong for most business AI.
What RRF does, and where the default comes from
When you run hybrid retrieval, with semantic search, keyword matching and domain constraints producing three result lists, you need to merge them into one ranking. The standard method is Reciprocal Rank Fusion (RRF). It is in every RAG tutorial. The formula sums 1/(k + rank) across each document, and the k parameter controls how much weight the top positions get. The default everyone uses is 60, the value from the original 2009 paper.
Why k=60 is harmful on a small corpus
k=60 was designed for web search: thousands of candidates across multiple engines. At that scale, the difference between rank 1 (0.0164) and rank 10 (0.0143) is gentle, and smooth blending is what you want. On a small corpus, a few hundred chunks and result lists of fifteen to twenty items, that “gentle” blending becomes “everything is the same.” Your carefully tuned retrieval might as well be random.
Drop k to 15 and rank 1 now scores 0.0625 while rank 10 scores 0.04. A real difference. Top-ranked results finally get rewarded for being top-ranked. We saw a measurable improvement in retrieval precision from that one change.
Check your k before you reach for something fancier
Before you go hunting for a more sophisticated fusion algorithm, check your k value. If it is 60 and your corpus is under a few thousand documents, you are leaving accuracy on the table. It is the same lesson as the reranker: the biggest gains in a production RAG pipeline often come from the least glamorous places, and only show up if you measure on your own data. Small thing, big difference.
What is the smallest parameter change that gave you the biggest measurable improvement?