Database / Qdrant Vector DB Interview questions
What is Reciprocal Rank Fusion (RRF) versus Distribution-Based Score Fusion (DBSF)?
Both are fusion methods for combining multiple ranked result lists into one, but they take different approaches to the core challenge that dense and sparse search scores aren't directly comparable — RRF sidesteps the problem by ignoring raw scores, while DBSF tackles it head-on by normalizing them.
| RRF | DBSF |
| Uses only each item's rank position in each list. | Normalizes each list's score distribution before combining. |
| Ignores the actual raw scores entirely. | Uses the underlying scores, after normalization. |
| Simple, robust, no tuning required. | Can better preserve the magnitude of strong matches. |
| Good general-purpose default. | Useful when score magnitude, not just rank, carries meaningful signal. |
Because dense (often cosine, bounded to [-1, 1]) and sparse (often BM25-based, unbounded) scores sit on fundamentally different scales that also shift per query, a naive combination using raw scores directly would be dominated by whichever method happens to produce larger numbers on a given query, regardless of actual relevance — RRF avoids this entirely by discarding scores and using only rank order, while DBSF addresses it by statistically normalizing each method's score distribution before combining them.
More Related questions...
