Skip to main content

​
Overview

RrfReRanker combines results from multiple vector queries using Reciprocal Rank Fusion (RRF). This algorithm assigns higher weight to documents that appear early in multiple result lists, without requiring relevance scores. The RRF score for a document at rank r is calculated as: 
score = 1 / (k + r + 1)
where k is the rank constant.
This reranker is specifically designed for multi-vector scenarios where query results from multiple vector fields need to be combined.

​
Constructor

from zvec.extension import RrfReRanker

reranker = RrfReRanker(
    topn=10,
    rerank_field=None,
    rank_constant=60
)

​
Parameters

​
topn
int
default:"10"
Number of top documents to return after reranking.
​
rerank_field
str | None
default:"None"
Ignored by RRF. This parameter exists for API consistency but has no effect.
​
rank_constant
int
default:"60"
Smoothing constant k in the RRF formula. Larger values reduce the impact of early ranks, making the fusion more balanced across different result lists.

​
Properties

​
rank_constant
int
The rank constant used in the RRF score calculation.
​
topn
int
Number of top documents to return.

​
Methods

​
rerank()

def rerank(self, query_results: dict[str, list[Doc]]) -> list[Doc]
Apply Reciprocal Rank Fusion to combine multiple query results.
​
query_results
dict[str, list[Doc]]
Results from one or more vector queries. Keys are vector field names, values are lists of retrieved documents.
​
returns
list[Doc]
Re-ranked documents with RRF scores in the score field.

​
Usage Example

​
Multi-Vector Search with RRF

from zvec import Collection
from zvec.extension import RrfReRanker

# Create collection with multiple vector fields
collection = Collection("products")

# Create RRF reranker
reranker = RrfReRanker(
    topn=10,
    rank_constant=60  # Default smoothing constant
)

# Query multiple vector fields
results = collection.query(
    vectors={
        "title_vec": title_embedding,
        "description_vec": desc_embedding,
        "tags_vec": tags_embedding
    },
    topn=20,  # Retrieve more results before reranking
    reranker=reranker
)

for doc in results:
    print(f"ID: {doc.id}, RRF Score: {doc.score}")

​
Adjusting Rank Constant

The rank_constant parameter controls how aggressively early ranks are weighted: 
# Lower rank_constant = more weight on top-ranked documents
reranker_aggressive = RrfReRanker(topn=10, rank_constant=10)

# Higher rank_constant = more balanced fusion
reranker_balanced = RrfReRanker(topn=10, rank_constant=100)

​
How It Works

  1. Collect Results: Gather documents from all vector field queries
  2. Calculate RRF Scores: For each document, sum RRF scores across all queries where it appears
  3. Deduplicate: Documents appearing in multiple result lists are processed once
  4. Rank: Return top n documents by RRF score

​
Example

Given two query results: 
query_results = {
    "title_vec": [doc_A, doc_B, doc_C],  # ranks 0, 1, 2
    "desc_vec": [doc_B, doc_D, doc_A]    # ranks 0, 1, 2
}
With rank_constant=60:
  • doc_A: 1/(60+2+1) + 1/(60+0+1) = 0.0159 + 0.0164 = 0.0323
  • doc_B: 1/(60+1+1) + 1/(60+0+1) = 0.0161 + 0.0164 = 0.0325 ← highest
  • doc_C: 1/(60+2+1) = 0.0159
  • doc_D: 1/(60+1+1) = 0.0161
Final ranking: doc_B, doc_A, doc_D, doc_C

​
When to Use RRF

Use RRF when:
  • Combining results from multiple vector fields
  • You don’t have access to raw similarity scores
  • You want a simple, parameter-free fusion method
  • Query fields have similar importance
Consider alternatives when:
  • Different fields should have different importance (use WeightedReRanker)
  • You need semantic understanding (use QwenReRanker)
  • Working with single-vector search (reranking not needed)

Build docs developers (and LLMs) love

Get started for freeTalk to us