Score Fusion Techniques
What you'll learn: How to intelligently combine scores from vector similarity and keyword searches into a single, unified ranking.
The Problem with Two Scores
When you run a hybrid search (keyword + vector), you get two different scores for each document:
- A vector similarity score (e.g., cosine similarity: 0.85)
- A keyword relevance score (e.g., BM25: 12.3)
These scores live in completely different ranges and scales. How do you combine them into one final ranking?
Weighted Sum Approach
The simplest method: normalize both scores to the same range (usually 0-1), then add them with weights:
final_score = (weight_vector × vector_score) + (weight_keyword × keyword_score)
For example, if you trust semantic meaning more:
final_score = (0.7 × 0.85) + (0.3 × 0.92) = 0.871
The challenge? Choosing the right weights and ensuring fair normalization across different score distributions.
Reciprocal Rank Fusion (RRF)
A more robust alternative that doesn't require normalization. RRF combines rankings (not raw scores):
RRF_score = Σ(1 / (k + rank_in_list))
Where k is typically 60. Each search method produces a ranked list, and RRF rewards documents that appear high in multiple lists. A document ranked #1 in vector search and #3 in keyword search gets a strong combined score, regardless of the original score scales.
Key Takeaway: Score fusion techniques like weighted sums and RRF solve the challenge of combining different search types into unified rankings—weighted sums offer control but require normalization, while RRF is robust and rank-based, avoiding scale issues entirely.