Reciprocal Rank Fusion
A hybrid retrieval result fusion algorithm that merges results from multiple recall channels by summing the reciprocals of their ranks. It effectively combines dense and sparse retrieval rankings without requiring parameter tuning.
Timeline (last 90 days)
The retrieval side uses four-way recall: vector, keyword, entity, and graph channels, merged via reciprocal rank fusion, ultimately outputting a context block constrained by a token budget, and it reports which channels failed.
Reciprocal rank fusion (RRF) takes 1/(k + rank) for each document in each result list (k is usually 60), sums them, and reranks by total score; it relies only on ranks and requires no normalization.
All Facts (2)
The retrieval side uses four-way recall: vector, keyword, entity, and graph channels, merged via reciprocal rank fusion, ultimately outputting a context block constrained by a token budget, and it reports which channels failed.
50%UnverifiedReciprocal rank fusion (RRF) takes 1/(k + rank) for each document in each result list (k is usually 60), sums them, and reranks by total score; it relies only on ranks and requires no normalization.
50%