Cross encoder reranker integration

Rerank retrieved documents with open-source cross-encoder models using LangChain Python.

A cross-encoder scores each (query, document) pair directly rather than comparing independent embeddings, which produces more accurate ordering at the cost of one extra inference per document. Applying a reranker on top of vector search (retrieve top-20 via embeddings, rerank down to top-5) is one of the highest-impact quality improvements for a RAG pipeline, and it runs locally on CPU for free when you use a small cross-encoder from Hugging Face.

This guide shows how to combine HuggingFaceCrossEncoder with LangChain’s CrossEncoderReranker and ContextualCompressionRetriever. The pattern works with any cross-encoder model on Hugging Face, including BAAI/bge-reranker-*, mixedbread-ai/mxbai-rerank-*, Alibaba-NLP/gte-multilingual-reranker-*, Qwen/Qwen3-Reranker-*, and the classic cross-encoder/ms-marco-* family.

Setup

pip install -qU langchain-huggingface langchain-community langchain-classic faiss-cpu

Build a base retriever

Start with a standard vector store retriever. Retrieve a relatively large k; the reranker will narrow it down.

The langchain-community package is no longer maintained. Examples that import from langchain_community may be outdated or broken. Use with caution.

from langchain_community.document_loaders import TextLoader
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter

documents = TextLoader("../../how_to/state_of_the_union.txt").load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)
texts = text_splitter.split_documents(documents)

embeddings = HuggingFaceEmbeddings(
    model_name="BAAI/bge-m3",
    encode_kwargs={"normalize_embeddings": True},
)
retriever = FAISS.from_documents(texts, embeddings).as_retriever(
    search_kwargs={"k": 20}
)

Rerank with a cross-encoder

CrossEncoderReranker wraps any cross-encoder and plugs into ContextualCompressionRetriever.

from langchain_classic.retrievers.contextual_compression import ContextualCompressionRetriever
from langchain_classic.retrievers.document_compressors import CrossEncoderReranker
from langchain_community.cross_encoders import HuggingFaceCrossEncoder

cross_encoder = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-v2-m3")
reranker = CrossEncoderReranker(model=cross_encoder, top_n=3)

compression_retriever = ContextualCompressionRetriever(
    base_compressor=reranker,
    base_retriever=retriever,
)

compressed_docs = compression_retriever.invoke("What is the plan for the economy?")
for i, doc in enumerate(compressed_docs, 1):
    print(f"Document {i}:\n{doc.page_content}\n")

Picking a cross-encoder

Model	Size	Notes
cross-encoder/ms-marco-MiniLM-L6-v2	22M	Fastest; English only, 2022-era baseline
BAAI/bge-reranker-v2-m3	568M	Multilingual, strong default for most workloads
mixedbread-ai/mxbai-rerank-large-v2	1.5B	Top-tier English quality, GPU recommended
Alibaba-NLP/gte-multilingual-reranker-base	306M	Multilingual, 8192-token context
Qwen/Qwen3-Reranker-0.6B	595M	Instruction-aware, multilingual

HuggingFaceCrossEncoder auto-selects the best available device (CUDA > MPS > CPU). To pin to a specific device, pass model_kwargs={"device": "cpu"} or similar.

Deploying to SageMaker

You can also host a cross-encoder on a SageMaker endpoint and use SagemakerEndpointCrossEncoder. Here is a sample inference.py that loads the model on the fly (no model.tar.gz artifacts required). See this walkthrough for step-by-step guidance.

import json
import logging
from typing import List

import torch
from sagemaker_inference import encoder
from transformers import AutoModelForSequenceClassification, AutoTokenizer

PAIRS = "pairs"
SCORES = "scores"


class CrossEncoder:
    def __init__(self) -> None:
        self.device = (
            torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
        )
        logging.info(f"Using device: {self.device}")
        model_name = "BAAI/bge-reranker-v2-m3"
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForSequenceClassification.from_pretrained(model_name)
        self.model = self.model.to(self.device)

    def __call__(self, pairs: List[List[str]]) -> List[float]:
        with torch.inference_mode():
            inputs = self.tokenizer(
                pairs,
                padding=True,
                truncation=True,
                return_tensors="pt",
                max_length=512,
            )
            inputs = inputs.to(self.device)
            scores = (
                self.model(**inputs, return_dict=True)
                .logits.view(
                    -1,
                )
                .float()
            )

        return scores.detach().cpu().tolist()


def model_fn(model_dir: str) -> CrossEncoder:
    try:
        return CrossEncoder()
    except Exception:
        logging.exception(f"Failed to load model from: {model_dir}")
        raise


def transform_fn(
    cross_encoder: CrossEncoder, input_data: bytes, content_type: str, accept: str
) -> bytes:
    payload = json.loads(input_data)
    model_output = cross_encoder(**payload)
    output = {SCORES: model_output}
    return encoder.encode(output, accept)

---

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.