Skip to main content

​
Overview

WeightedReRanker combines scores from multiple vector fields using a weighted sum. Each vector field’s relevance score is normalized based on its metric type, then scaled by a user-provided weight. Final scores are summed across fields.
This reranker is specifically designed for multi-vector scenarios where query results from multiple vector fields need to be combined with configurable weights.

​
Constructor

from zvec import MetricType
from zvec.extension import WeightedReRanker

reranker = WeightedReRanker(
    topn=10,
    rerank_field=None,
    metric=MetricType.L2,
    weights={"title_vec": 2.0, "content_vec": 1.0}
)

​
Parameters

​
topn
int
default:"10"
Number of top documents to return after reranking.
​
rerank_field
str | None
default:"None"
Ignored. This parameter exists for API consistency but has no effect.
​
metric
MetricType
default:"MetricType.L2"
Distance metric used for score normalization. Supported metrics: L2, IP (inner product), COSINE. Scores are normalized to [0, 1] range.
​
weights
dict[str, float] | None
default:"None"
Weight per vector field. Keys are vector field names, values are weight multipliers. Fields not listed use weight 1.0.

​
Properties

​
weights
dict[str, float]
Weight mapping for vector fields.
​
metric
MetricType
Distance metric used for score normalization.
​
topn
int
Number of top documents to return.

​
Methods

​
rerank()

def rerank(self, query_results: dict[str, list[Doc]]) -> list[Doc]
Combine scores from multiple vector fields using weighted sum.
​
query_results
dict[str, list[Doc]]
Results per vector field. Keys are vector field names, values are lists of documents.
​
returns
list[Doc]
Re-ranked documents with combined scores in the score field.

​
Score Normalization

Raw distance/similarity scores are normalized to [0, 1] based on the metric type:

​
L2 (Euclidean Distance)

normalized_score = 1.0 - 2 * arctan(distance) / π
Smaller distances map to higher scores.

​
IP (Inner Product)

normalized_score = 0.5 + arctan(score) / π
Higher inner products map to higher scores.

​
COSINE (Cosine Similarity)

normalized_score = 1.0 - distance / 2.0
Smaller cosine distances map to higher scores.

​
Usage Examples

from zvec import Collection, MetricType
from zvec.extension import WeightedReRanker

# Create collection
collection = Collection("articles")

# Title matches are 3x more important than content
reranker = WeightedReRanker(
    topn=10,
    metric=MetricType.COSINE,
    weights={
        "title_vec": 3.0,
        "content_vec": 1.0
    }
)

results = collection.query(
    vectors={
        "title_vec": title_embedding,
        "content_vec": content_embedding
    },
    topn=20,
    reranker=reranker
)

for doc in results:
    print(f"ID: {doc.id}, Weighted Score: {doc.score:.4f}")

​
Dynamic Weight Adjustment

# Adjust weights based on query type
def get_reranker(query_type: str) -> WeightedReRanker:
    if query_type == "navigational":
        # Favor title matches for navigational queries
        weights = {"title_vec": 5.0, "content_vec": 1.0, "tags_vec": 2.0}
    elif query_type == "informational":
        # Favor content for informational queries
        weights = {"title_vec": 1.0, "content_vec": 4.0, "tags_vec": 1.5}
    else:
        # Balanced for other queries
        weights = {"title_vec": 1.5, "content_vec": 2.0, "tags_vec": 1.0}
    
    return WeightedReRanker(
        topn=10,
        metric=MetricType.COSINE,
        weights=weights
    )

reranker = get_reranker("navigational")
results = collection.query(
    vectors=query_vectors,
    reranker=reranker
)

​
Equal Weights (Uniform Fusion)

# When no weights specified, all fields have weight 1.0
reranker = WeightedReRanker(
    topn=10,
    metric=MetricType.L2
)

results = collection.query(
    vectors={
        "vec1": embedding1,
        "vec2": embedding2,
        "vec3": embedding3
    },
    reranker=reranker
)

​
How It Works

  1. Normalize Scores: Convert raw distance/similarity scores to [0, 1] range
  2. Apply Weights: Multiply normalized scores by field weights
  3. Sum Scores: For each document, sum weighted scores across all fields
  4. Rank: Return top n documents by combined score

​
Example

Given query results with COSINE metric: 
query_results = {
    "title_vec": [Doc(id="A", score=0.1), Doc(id="B", score=0.3)],
    "content_vec": [Doc(id="A", score=0.2), Doc(id="C", score=0.15)]
}

weights = {"title_vec": 2.0, "content_vec": 1.0}
Normalized and weighted scores:
  • doc_A: (1 - 0.1/2) * 2.0 + (1 - 0.2/2) * 1.0 = 1.9 + 0.9 = 2.8
  • doc_B: (1 - 0.3/2) * 2.0 = 1.7
  • doc_C: (1 - 0.15/2) * 1.0 = 0.925
Final ranking: doc_A, doc_B, doc_C

​
When to Use Weighted Reranking

Use WeightedReRanker when:
  • Different vector fields have different importance
  • You have domain knowledge about field weights
  • You need fine-grained control over score combination
  • Working with multi-vector hybrid search
Consider alternatives when:
  • All fields have equal importance (use RrfReRanker)
  • You need semantic understanding (use QwenReRanker)
  • Weights are difficult to determine (try RRF first)

​
Choosing Weights

Best Practices:
  • Start with equal weights (1.0) and adjust based on results
  • Use 2-5x differences for important vs. less important fields
  • Test with representative queries and measure relevance
  • Consider query intent when setting weights dynamically

Build docs developers (and LLMs) love

Get started for freeTalk to us