Why fuse
Keyword retrieval excels at exact matches; vector retrieval excels at meaning. Hybrid retrieval runs both and fuses their results so you get the strengths of each.
Two fusion styles
- Score fusion combines the numeric scores from each retriever, after normalizing them onto a comparable scale.
- Rank fusion ignores raw scores and combines based on each document position in each list. A common method gives a document a contribution that shrinks with its rank, then sums contributions across lists.
Why rank fusion is popular
Keyword and vector scores live on different scales and are hard to compare directly. Rank fusion sidesteps this by using positions, which makes it robust without careful score calibration.
Tuning
You can weight one retriever more than the other when you know your corpus leans toward exact or conceptual matching. The fused list then goes to a reranker for final ordering.
Diagram
Key idea
Hybrid retrieval runs keyword and vector search together and fuses them, with rank fusion avoiding the trouble of comparing scores on different scales.