Cross Encoder Reranker

This notebook shows how to implement reranker in a retriever with your own cross encoder from Hugging Face cross encoder models or Hugging Face models that implements cross encoder function (example: BAAI/bge-reranker-base). SagemakerEndpointCrossEncoder enables you to use these HuggingFace models loaded on Sagemaker.

This builds on top of ideas in the ContextualCompressionRetriever. Overall structure of this document came from Cohere Reranker documentation.

For more about why cross encoder can be used as reranking mechanism in conjunction with embeddings for better retrieval, refer to Hugging Face Cross-Encoders documentation.

#!pip install faiss sentence_transformers

# OR  (depending on Python version)

#!pip install faiss-cpu sentence_transformers

# Helper function for printing docs


def pretty_print_docs(docs):
    print(
        f"\n{'-' * 100}\n".join(
            [f"Document {i+1}:\n\n" + d.page_content for i, d in enumerate(docs)]
        )
    )

Set up the base vector store retriever

Let's start by initializing a simple vector store retriever and storing the 2023 State of the Union speech (in chunks). We can set up the retriever to retrieve a high number (20) of docs.

from langchain.document_loaders import TextLoader
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_text_splitters import RecursiveCharacterTextSplitter

documents = TextLoader("../../modules/state_of_the_union.txt").load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)
texts = text_splitter.split_documents(documents)
embeddingsModel = HuggingFaceEmbeddings(
    model_name="sentence-transformers/msmarco-distilbert-dot-v5"
)
retriever = FAISS.from_documents(texts, embeddingsModel).as_retriever(
    search_kwargs={"k": 20}
)

query = "What is the plan for the economy?"
docs = retriever.invoke(query)
pretty_print_docs(docs)

Doing reranking with CrossEncoderReranker

Now let's wrap our base retriever with a ContextualCompressionRetriever. CrossEncoderReranker uses HuggingFaceCrossEncoder to rerank the returned results.

from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import CrossEncoderReranker
from langchain_community.cross_encoders import HuggingFaceCrossEncoder

model = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-base")
compressor = CrossEncoderReranker(model=model, top_n=3)
compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor, base_retriever=retriever
)

compressed_docs = compression_retriever.invoke("What is the plan for the economy?")
pretty_print_docs(compressed_docs)

Document 1:

More infrastructure and innovation in America. 

More goods moving faster and cheaper in America. 

More jobs where you can earn a good living in America. 

And instead of relying on foreign supply chains, let’s make it in America. 

Economists call it “increasing the productive capacity of our economy.” 

I call it building a better America. 

My plan to fight inflation will lower your costs and lower the deficit.
----------------------------------------------------------------------------------------------------
Document 2:

Second – cut energy costs for families an average of $500 a year by combatting climate change.  

Let’s provide investments and tax credits to weatherize your homes and businesses to be energy efficient and you get a tax credit; double America’s clean energy production in solar, wind, and so much more;  lower the price of electric vehicles, saving you another $80 a month because you’ll never have to pay at the gas pump again.
----------------------------------------------------------------------------------------------------
Document 3:

Look at cars. 

Last year, there weren’t enough semiconductors to make all the cars that people wanted to buy. 

And guess what, prices of automobiles went up. 

So—we have a choice. 

One way to fight inflation is to drive down wages and make Americans poorer.  

I have a better plan to fight inflation. 

Lower your costs, not your wages. 

Make more cars and semiconductors in America. 

More infrastructure and innovation in America. 

More goods moving faster and cheaper in America.

Uploading Hugging Face model to SageMaker endpoint

Here is a sample inference.py for creating an endpoint that works with SagemakerEndpointCrossEncoder. For more details with step-by-step guidance, refer to this article.

It downloads Hugging Face model on the fly, so you do not need to keep the model artifacts such as pytorch_model.bin in your model.tar.gz.

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-base"
        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)