A single search can miss relevant documents because one phrasing captures only part of a user’s intent. Reciprocal rank fusion (RRF) combines ranked result lists from multiple searches into one list. In a RAG pipeline, those searches might use different query variants; the fused list is assembled before documents are passed to the answer generator. section 8.0 section 8.4
RRF ignores each search’s original scores. For every document, it adds 1 / (k + rank) for each list in which the document appears, then sorts documents by the total. A document found consistently across several lists can outrank one that tops only a single list. This avoids averaging scores that may have different scales, such as cosine similarity and BM25 scores. section 8.3
Listing 8.2 Reciprocal Rank Fusion algorithm.
from collections import defaultdict
from typing import List, Tuple, Dict
def reciprocal_rank_fusion(
ranked_lists: List[List[str]], k: int = 60
) -> List:
fused_scores: Dict[str, float] = defaultdict(float) #A
for ranked_list in ranked_lists: #B
for index, doc_id in enumerate(ranked_list):
rank = index + 1 #C
score = 1.0 / (k + rank) #C
fused_scores[doc_id] += score #D
sorted_docs = sorted( #E
fused_scores.items(),
key=lambda item: item[1],
reverse=True #F
)
return sorted_docs
The Python implementation accumulates reciprocal-rank contributions by document ID and sorts the resulting totals; k defaults to 60. section 8.3
A product-search result that one query misses
A search for a warm winter jacket returns close matches to those words but few kinds of winter clothing. Separate searches for parkas, fleeces, and ski jackets produce different candidates. Fusing their lists brings a down parka and a fleece-lined ski jacket into the top results alongside a warm jacket. section 8.5
Choose RRF for its simple rank-based merge, but keep its limits in view:
- It does not require score normalization, unlike score-summing methods whose inputs may be uncalibrated. section 8.3
- It discards match strength: a first-place result with a weak raw score contributes as much as one with a strong raw score. Carry raw scores separately if a later stage needs confidence information. section 8.3
- Its results are sensitive to k. In multi-query retrieval, extra variants also add cost, latency, and opportunities for noise; three to five queries with k = 60 are a starting point, not a guarantee. section 8.3 section 8.7
RRF also fits hybrid retrieval: a HyDE vector search and a lexical search can run in parallel, then merge their rankings. That retains semantic reach while giving exact terms—such as part numbers or drug names—a route into the final list. section 7.6.2





