engineering · 1 min read
Reciprocal Rank Fusion vs Weighted Sum: Picking a Hybrid Fuser
When fusing dense and sparse results, RRF and weighted sum behave differently. Here is which one to pick - and why Bhogar AI defaults to RRF.
BABhogar AI TeamProduct & Engineering
You decided to do hybrid retrieval. Now you have to fuse the dense and sparse results. The two dominant options - Reciprocal Rank Fusion (RRF) and weighted sum - produce noticeably different rankings.
Why it matters
Weighted sum requires careful score normalisation; small bugs there create silent quality regressions. RRF works on ranks, is parameter-light, and behaves well across heterogeneous retrievers.
How Bhogar AI approaches it
Bhogar AI defaults to RRF for hybrid retrieval. Weighted sum is available for advanced use cases where you have well-calibrated retrievers and a strong reason to prefer score blending.
- RRF default with sensible k parameter
- Weighted sum optional with score normalisation built-in
- Per-KB fuser selection
- Telemetry comparing both on the same queries
- Compatible with reranking on top of fusion
What you get
For 9 out of 10 customer KBs, RRF matches or beats weighted sum on top-3 recall - without parameter tuning.