Cross Encoder Reranker

Overview

The Cross Encoder Reranker is a technique designed to enhance the performance of Retrieval-Augmented Generation (RAG) systems. This guide explains how to implement a reranker using Hugging Face's Cross Encoders to refine the ranking of retrieved documents, promoting those most relevant to a query.

Table of Contents

References

Hugging Face Cross Encoders

Environment Setup

Set up the environment. You may refer to Environment Setup for more details.

[Note]

  • langchain-opentutorial is a package that provides a set of easy-to-use environment setup, useful functions and utilities for tutorials.

  • You can checkout the langchain-opentutorial for more details.

# Set environment variables
from langchain_opentutorial import set_env

set_env(
    {
        "OPENAI_API_KEY": "",
    }
)

You can alternatively set OPENAI_API_KEY in .env file and load it.

[Note] This is not necessary if you've already set OPENAI_API_KEY in previous steps.

# Configuration file to manage API keys as environment variables
from dotenv import load_dotenv

# Load API key information
load_dotenv(override=True)
%%capture --no-stderr
%pip install langchain-opentutorial

Key Features and Mechanism

Purpose

  • Re-rank retrieved documents to refine their ranking, prioritizing the most relevant results for the query.

Structure

  • Accepts both the query and document as a single input pair, enabling joint processing.

Mechanism

  • Single Input Pair : Processes the query and document as a combined input to output a relevance score directly.

  • Self-Attention Mechanism : Uses self-attention to jointly analyze the query and document , effectively capturing their semantic relationship.

Advantages

  • Higher Accuracy : Provides more precise similarity scores.

  • Deep Contextual Analysis : Explores semantic nuances between query and document .

Limitations

  • High Computational Costs : Processing can be time-intensive.

  • Scalability Issues : Not suitable for large-scale document collections without optimization.

Practical Applications

  • A Bi-Encoder quickly retrieves candidate documents by computing lightweight similarity scores.

  • A Cross Encoder refines these results by deeply analyzing the semantic relationship between the query and the retrieved documents .

Implementation

  • Use Hugging Face cross encoders or BAAI/bge-reranker models.

  • Easily integrate with frameworks like LangChain through the CrossEncoderReranker component.

Key Advantages of Reranker

  • Precise Similarity Scoring : Delivers highly accurate measurements of relevance between the query and documents.

  • Semantic Depth : Analyzes deeper semantic relationships, uncovering nuances in query - document interactions.

  • Refined Search Quality : Improves the relevance and quality of the retrieved documents .

  • RAG System Boost : Enhances the performance of Retrieval-Augmented Generation (RAG) systems by refining input relevance.

  • Seamless Integration : Easily adaptable to various workflows and compatible with multiple frameworks.

  • Model Versatility : Offers flexibility with a wide range of pre-trained models for tailored use cases.

Document Count Settings for Reranker

  • Reranking is generally performed on the top 5–10 documents retrieved during the initial search.

  • The ideal number of documents for reranking should be determined through experimentation and evaluation, as it depends on the dataset characteristics and computational resources available.

Trade-offs When Using a Reranker

  • Accuracy vs Processing Time : Striking a balance between achieving higher accuracy and minimizing processing time.

  • Performance Improvement vs Computational Cost : Weighing the benefits of improved performance against the additional computational resources required.

  • Search Speed vs Relevance Accuracy : Managing the trade-off between faster retrieval and maintaining high relevance in results.

  • System Requirements : Ensuring the system meets the necessary hardware and software requirements to support reranking.

  • Dataset Characteristics : Considering the scale, diversity, and specific attributes of the dataset to optimize reranker performance.

Explaining the Implementation of Cross Encoder Reranker with a Simple Example

# Helper function to format and print document content
def pretty_print_docs(docs):
    # Print each document in the list with a separator between them
    print(
        f"\n{'-' * 100}\n".join(  # Separator line for better readability
            [f"Document {i+1}:\n\n" + d.page_content for i, d in enumerate(docs)]  # Format: Document number + content
        )
    )
from langchain_community.document_loaders import TextLoader
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter

# Load documents
documents = TextLoader("./data/appendix-keywords.txt").load()

# Configure text splitter
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)

# Split documents into chunks
texts = text_splitter.split_documents(documents)

# # Set up the embedding model
embeddings_model = HuggingFaceEmbeddings(
    model_name="sentence-transformers/msmarco-distilbert-dot-v5",
    model_kwargs = {"tokenizer_kwargs": {"clean_up_tokenization_spaces": False}}
)

# Create FAISS index from documents and set up retriever
retriever = FAISS.from_documents(texts, embeddings_model).as_retriever(
    search_kwargs={"k": 10}
)

# Define the query
query = "Can you tell me about Word2Vec?"

# Execute the query and retrieve results
docs = retriever.invoke(query)

# Display the retrieved documents
pretty_print_docs(docs)
Document 1:
    
    Word2Vec
    Definition: Word2Vec is a technique in NLP that maps words to a vector space, representing their semantic relationships based on context.
    Example: In a Word2Vec model, "king" and "queen" are represented by vectors located close to each other.
    Related Keywords: Natural Language Processing (NLP), Embedding, Semantic Similarity
    ----------------------------------------------------------------------------------------------------
    Document 2:
    
    Token
    Definition: A token refers to a smaller unit of text obtained by splitting a larger piece of text. It can be a word, phrase, or sentence.
    Example: The sentence "I go to school" can be tokenized into "I," "go," "to," and "school."
    Related Keywords: Tokenization, Natural Language Processing (NLP), Syntax Analysis
    ----------------------------------------------------------------------------------------------------
    Document 3:
    
    Example: A customer information table in a relational database is an example of structured data.
    Related Keywords: Database, Data Analysis, Data Modeling
    ----------------------------------------------------------------------------------------------------
    Document 4:
    
    Schema
    Definition: A schema defines the structure of a database or file, detailing how data is organized and stored.
    Example: A relational database schema specifies column names, data types, and key constraints.
    Related Keywords: Database, Data Modeling, Data Management
    ----------------------------------------------------------------------------------------------------
    Document 5:
    
    Keyword Search
    Definition: Keyword search involves finding information based on user-inputted keywords, commonly used in search engines and database systems.
    Example: Searching 
    When a user searches for "coffee shops in Seoul," the system returns a list of relevant coffee shops.
    Related Keywords: Search Engine, Data Search, Information Retrieval
    ----------------------------------------------------------------------------------------------------
    Document 6:
    
    TF-IDF (Term Frequency-Inverse Document Frequency)
    Definition: TF-IDF is a statistical measure used to evaluate the importance of a word within a document by considering its frequency and rarity across a corpus.
    Example: Words with high TF-IDF values are often unique and critical for understanding the document.
    Related Keywords: Natural Language Processing (NLP), Information Retrieval, Data Mining
    ----------------------------------------------------------------------------------------------------
    Document 7:
    
    SQL
    Definition: SQL (Structured Query Language) is a programming language for managing data in databases. 
    It allows you to perform various operations such as querying, updating, inserting, and deleting data.
    Example: SELECT * FROM users WHERE age > 18; retrieves information about users aged above 18.
    Related Keywords: Database, Query, Data Management
    ----------------------------------------------------------------------------------------------------
    Document 8:
    
    Open Source
    Definition: Open source software allows its source code to be freely used, modified, and distributed, fostering collaboration and innovation.
    Example: The Linux operating system is a well-known open source project.
    Related Keywords: Software Development, Community, Technical Collaboration
    Structured Data
    Definition: Structured data is organized according to a specific format or schema, making it easy to search and analyze.
    ----------------------------------------------------------------------------------------------------
    Document 9:
    
    Semantic Search
    Definition: Semantic search is a search technique that understands the meaning of a user's query beyond simple keyword matching, returning results that are contextually relevant.
    Example: If a user searches for "planets in the solar system," the system provides information about planets like Jupiter and Mars.
    Related Keywords: Natural Language Processing (NLP), Search Algorithms, Data Mining
    ----------------------------------------------------------------------------------------------------
    Document 10:
    
    GPT (Generative Pretrained Transformer)
    Definition: GPT is a generative language model pre-trained on vast datasets, capable of performing various text-based tasks. It generates natural and coherent text based on input.
    Example: A chatbot generating detailed answers to user queries is powered by GPT models.
    Related Keywords: Natural Language Processing (NLP), Text Generation, Deep Learning

Now, let's wrap the base_retriever with a ContextualCompressionRetriever . The CrossEncoderReranker leverages HuggingFaceCrossEncoder to re-rank the retrieved results.

Multilingual Support BGE Reranker: bge-reranker-v2-m3

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

# Initialize the model
model = HuggingFaceCrossEncoder(model_name="BAAI/bge-reranker-v2-m3")

# Select the top 3 documents
compressor = CrossEncoderReranker(model=model, top_n=3)

# Initialize the contextual compression retriever
compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor, base_retriever=retriever
)

# Retrieve compressed documents
compressed_docs = compression_retriever.invoke("Can you tell me about Word2Vec?")

# Display the documents
pretty_print_docs(compressed_docs)
Document 1:
    
    Word2Vec
    Definition: Word2Vec is a technique in NLP that maps words to a vector space, representing their semantic relationships based on context.
    Example: In a Word2Vec model, "king" and "queen" are represented by vectors located close to each other.
    Related Keywords: Natural Language Processing (NLP), Embedding, Semantic Similarity
    ----------------------------------------------------------------------------------------------------
    Document 2:
    
    Open Source
    Definition: Open source software allows its source code to be freely used, modified, and distributed, fostering collaboration and innovation.
    Example: The Linux operating system is a well-known open source project.
    Related Keywords: Software Development, Community, Technical Collaboration
    Structured Data
    Definition: Structured data is organized according to a specific format or schema, making it easy to search and analyze.
    ----------------------------------------------------------------------------------------------------
    Document 3:
    
    TF-IDF (Term Frequency-Inverse Document Frequency)
    Definition: TF-IDF is a statistical measure used to evaluate the importance of a word within a document by considering its frequency and rarity across a corpus.
    Example: Words with high TF-IDF values are often unique and critical for understanding the document.
    Related Keywords: Natural Language Processing (NLP), Information Retrieval, Data Mining
Previous11-RerankerNextJinaReranker

Last updated