Skip to main content

​
What is Reranking?

Reranking is a technique that refines the output of one or more vector queries by applying a secondary scoring strategy. Instead of relying solely on vector similarity scores, rerankers can:
  • Combine results from multiple vector fields
  • Apply semantic understanding using cross-encoder models
  • Merge rankings using fusion algorithms
Rerankers are used in the query() method of Collection via the reranker parameter.

​
When to Use Reranking

When querying multiple vector fields (e.g., title embeddings and content embeddings), reranking helps combine the results into a unified ranking: 
from zvec import Collection
from zvec.extension import RrfReRanker

collection = Collection("products")
reranker = RrfReRanker(topn=10)

results = collection.query(
    vectors={"title_vec": query_title_emb, "desc_vec": query_desc_emb},
    reranker=reranker
)

​
Semantic Relevance

For search scenarios requiring deep semantic understanding, model-based rerankers can provide more accurate relevance scores: 
from zvec.extension import QwenReRanker

reranker = QwenReRanker(
    query="machine learning algorithms",
    topn=5,
    rerank_field="content"
)

results = collection.query(
    vectors={"content_vec": query_embedding},
    reranker=reranker
)

​
Available Rerankers

Zvec provides several reranking strategies:

RrfReRanker

Reciprocal Rank Fusion for combining multi-vector results

WeightedReRanker

Weighted score combination with configurable field weights

QwenReRanker

Model-based semantic reranking using Qwen/DashScope API

​
ReRanker Base Class

​
RerankFunction
class
Abstract base class for all rerankers.

​
Constructor

RerankFunction(topn=10, rerank_field=None)
​
topn
int
default:"10"
Number of top documents to return after reranking.
​
rerank_field
str | None
default:"None"
Field name used as input for reranking (e.g., document title or body). Required for some rerankers like QwenReRanker.

​
Properties

​
topn
int
Number of top documents to return after reranking.
​
rerank_field
str | None
Field name used as reranking input.

​
Methods

​
rerank()

def rerank(self, query_results: dict[str, list[Doc]]) -> list[Doc]
Re-rank documents from one or more vector queries.
​
query_results
dict[str, list[Doc]]
Mapping from vector field name to list of retrieved documents (sorted by relevance).
​
returns
list[Doc]
Re-ranked list of documents (length ≤ topn), with updated score fields.
Subclasses must implement the rerank() method.

Build docs developers (and LLMs) love

Get started for freeTalk to us