Rerankers

Re-score retrieved documents to improve ranking quality.

Available Rerankers

API-based

Reranker	Type	Provider
Cohere	API	Cohere
Jina	API	Jina AI
VoyageAI	API	Voyage AI
MixedbreadAI	API	Mixedbread AI
RankGPT	LLM	Any LLM
UPR	LLM	Any LLM
FlagEmbedding LLM	LLM-based	BAAI

Local Models

Reranker	Type	Library
SentenceTransformer	CrossEncoder	sentence-transformers
FlagEmbedding	CrossEncoder	FlagEmbedding
FlashRank	ONNX	flashrank
ColBERT	Token-level	torch, transformers
MonoT5	Seq2Seq	torch, transformers
TART	Instruction-T5	torch, transformers
KoReranker	Korean	torch, transformers
OpenVINO	HW-optimized	optimum-intel

Installation

API rerankers require optional dependencies:

pip install autorag-research[reranker]

GPU/local rerankers require the gpu extra:

pip install autorag-research[gpu]

OpenVINO reranker requires:

pip install autorag-research[openvino]

Base Class

All rerankers extend BaseReranker:

from autorag_research.rerankers import BaseReranker, RerankResult


class MyReranker(BaseReranker):
    def rerank(self, query: str, documents: list[str], top_k: int | None = None) -> list[RerankResult]:
        # Return sorted results by relevance
        pass

    async def arerank(self, query: str, documents: list[str], top_k: int | None = None) -> list[RerankResult]:
        # Async version
        pass

Methods

Method	Description
rerank(query, docs, top_k)	Single query reranking
arerank(query, docs, top_k)	Async single query
rerank_documents(queries, docs_list, top_k)	Batch reranking
arerank_documents(queries, docs_list, top_k)	Async batch

Usage with Injection

from autorag_research.injection import load_reranker, with_reranker

# Load from config
reranker = load_reranker("cohere")

# Or use decorator
@with_reranker()
def my_func(reranker):
    return reranker.rerank("query", ["doc1", "doc2"])