langchain-reranker

Add init_reranker to the LangChain ecosystem, following the same conventions as init_chat_model and init_embeddings.

Installation

pip install langchain-reranker

Provider integration packages are optional and installed on demand, e.g. pip install langchain-cohere.

Usage

from langchain_reranker.rerankers import init_reranker

reranker = init_reranker("cohere:rerank-v3.5")

compressed = reranker.compress_documents(documents, query="what is langchain?")

init_reranker returns a BaseDocumentCompressor, so it plugs directly into existing retrieval pipelines such as a contextual compression retriever.

reranker = init_reranker(model="jina-reranker-v2-base-multilingual", provider="jina")

Custom rerankers

Implement rerank and inherit compress_documents for free.

from langchain_reranker.rerankers import BaseReranker, RerankResult


class MyReranker(BaseReranker):
    model: str = "my-reranker"

    def rerank(self, query, documents, top_n=None):
        scores = my_api(query, [doc.page_content for doc in documents])
        ranked = sorted(zip(documents, scores), key=lambda pair: -pair[1])
        return [
            RerankResult(index=i, relevance_score=float(score), document=doc)
            for i, (doc, score) in enumerate(ranked)
        ]

Local cross encoders

reranker = init_reranker("huggingface:BAAI/bge-reranker-v2-m3")