Retrieval-Augmented Generation (RAG) Project

🔜 Check out bragai.dev (launching soon)

---

This repository contains a comprehensive exploration of Retrieval-Augmented Generation (RAG) for various applications. Each notebook provides a detailed, hands-on guide to setting up and experimenting with RAG from an introductory level to advanced implementations, including multi-querying and custom RAG builds.

Project Structure

If you want to jump straight into it, check out the file full_basic_rag.ipynb -> this file will give you a boilerplate starter code of a fully customizable RAG chatbot.

Make sure to run your files in a virtual environment (checkout section Get Started)

The following notebooks can be found under the directory notebooks/.

[1]_rag_setup_overview.ipynb

This introductory notebook provides an overview of RAG architecture and its foundational setup. The notebook walks through:

• Environment Setup: Configuring the environment, installing necessary libraries, and API setups.
• Initial Data Loading: Basic document loaders and data preprocessing methods.
• Embedding Generation: Generating embeddings using various models, including OpenAI's embeddings.
• Vector Store: Setting up a vector store (ChromaDB/Pinecone) for efficient similarity search.
• Basic RAG Pipeline: Creating a simple retrieval and generation pipeline to serve as a baseline.

[2]_rag_with_multi_query.ipynb

Building on the basics, this notebook introduces multi-querying techniques in the RAG pipeline, exploring:

• Multi-Query Setup: Configuring multiple queries to diversify retrieval.
• Advanced Embedding Techniques: Utilizing multiple embedding models to refine retrieval.
• Pipeline with Multi-Querying: Implementing multi-query handling to improve relevance in response generation.
• Comparison & Analysis: Comparing results with single-query pipelines and analyzing performance improvements.

[3]_rag_routing_and_query_construction.ipynb

This notebook delves deeper into customizing a RAG pipeline. It covers:

• Logical Routing: Implements function-based routing for classifying user queries to appropriate data sources based on programming languages.
• Semantic Routing: Uses embeddings and cosine similarity to direct questions to either a math or physics prompt, optimizing response accuracy.
• Query Structuring for Metadata Filters: Defines structured search schema for YouTube tutorial metadata, enabling advanced filtering (e.g., by view count, publication date).
• Structured Search Prompting: Leverages LLM prompts to generate database queries for retrieving relevant content based on user input.
• Integration with Vector Stores: Links structured queries to vector stores for efficient data retrieval.

[4]_rag_indexing_and_advanced_retrieval.ipynb

Continuing from the previous customization, this notebook explores:

• Preface on Document Chunking: Points to external resources for document chunking techniques.
• Multi-representation Indexing: Sets up a multi-vector indexing structure for handling documents with different embeddings and representations.
• In-Memory Storage for Summaries: Uses InMemoryByteStore for storing document summaries alongside parent documents, enabling efficient retrieval.
• MultiVectorRetriever Setup: Integrates multiple vector representations to retrieve relevant documents based on user queries.
• RAPTOR Implementation: Explores RAPTOR, an advanced indexing and retrieval model, linking to in-depth resources.
• ColBERT Integration: Demonstrates ColBERT-based token-level vector indexing and retrieval, which captures contextual meaning at a fine-grained level.
• Wikipedia Example with ColBERT: Retrieves information about Hayao Miyazaki using the ColBERT retrieval model for demonstration.

[5]_rag_retrieval_and_reranking.ipynb

This final notebook brings together the RAG system components, with a focus on scalability and optimization:

• Document Loading and Splitting: Loads and chunks documents for indexing, preparing them for vector storage.
• Multi-query Generation with RAG-Fusion: Uses a prompt-based approach to generate multiple search queries from a single input question.
• Reciprocal Rank Fusion (RRF): Implements RRF for re-ranking multiple retrieval lists, merging results for improved relevance.
• Retriever and RAG Chain Setup: Constructs a retrieval chain for answering queries, using fused rankings and RAG chains to pull contextually relevant information.
• Cohere Re-Ranking: Demonstrates re-ranking with Cohere’s model for additional contextual compression and refinement.
• CRAG and Self-RAG Retrieval: Explores advanced retrieval approaches like CRAG and Self-RAG, with links to examples.
• Exploration of Long-Context Impact: Links to resources explaining the impact of long-context retrieval on RAG models.

Getting Started

Pre-requisites

Ensure Python 3.11.11 (preferred) is installed on your system. Follow the platform-specific instructions below to install it if not already installed.

macOS

1. Install Homebrew if not already installed:

   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
   

2. Install Python 3.11.11:

   brew install [email protected]
   

3. Verify installation:

   python3.11 --version
   

Linux

1. Update your package manager:

   sudo apt update
   

2. Install Python 3.11.11:

   sudo apt install python3.11 python3.11-venv
   

3. Verify installation:

   python3.11 --version
   

Windows

1. Download the Python 3.11.11 installer from Python.org.
2. Run the installer and ensure you check the box "Add Python to PATH".
3. Verify installation:

   python --version
   

---

Installation Instructions

1. Clone the Repository

git clone https://github.com/bRAGAI/bRAG-langchain.git
cd bRAG-langchain

2. Create a Virtual Environment

Use Python 3.11.11 to create a virtual environment:

python3.11 -m venv venv

Activate the virtual environment:

• macOS/Linux:

  source venv/bin/activate
  

• Windows:

  venv\Scripts\activate
  

3. Verify and Fix Python Version

If the virtual environment defaults to a different Python version (e.g., Python 3.13):

1. Verify the current Python version inside the virtual environment:

   python --version
   

2. Use Python 3.11 explicitly within the virtual environment:

   python3.11
   

3. Ensure the python command uses Python 3.11 by creating a symbolic link:

   ln -sf $(which python3.11) $(dirname $(which python))/python
   

4. Verify the fix:

   python --version
   

4. Install Dependencies

Install the required packages:

pip install -r requirements.txt

---

Additional Steps

5. Run the Notebooks

Begin with [1]_rag_setup_overview.ipynb to get familiar with the setup process. Proceed sequentially through the other notebooks:

• [1]_rag_setup_overview.ipynb
• [2]_rag_with_multi_query.ipynb
• [3]_rag_routing_and_query_construction.ipynb
• [4]_rag_indexing_and_advanced_retrieval.ipynb
• [5]_rag_retrieval_and_reranking.ipynb

6. Set Up Environment Variables

1. Duplicate the .env.example file in the root directory and rename it to .env.

2. Add the following keys (replace with your actual values):

   # LLM Model - Get key at https://platform.openai.com/api-keys
   OPENAI_API_KEY="your-api-key"
   
   # LangSmith - Get key at https://smith.langchain.com
   LANGCHAIN_TRACING_V2=true
   LANGCHAIN_ENDPOINT="https://api.smith.langchain.com"
   LANGCHAIN_API_KEY="your-api-key"
   LANGCHAIN_PROJECT="your-project-name"
   
   # Pinecone Vector Database - Get key at https://app.pinecone.io
   PINECONE_INDEX_NAME="your-project-index"
   PINECONE_API_HOST="your-host-url"
   PINECONE_API_KEY="your-api-key"
   
   # Cohere - Get key at https://dashboard.cohere.com/api-keys
   COHERE_API_KEY=your-api-key
   

---

You're now ready to use the project!

Usage

After setting up the environment and running the notebooks in sequence, you can:

1. Experiment with Retrieval-Augmented Generation: Use the foundational setup in [1]_rag_setup_overview.ipynb to understand the basics of RAG.

2. Implement Multi-Querying: Learn how to improve response relevance by introducing multi-querying techniques in [2]_rag_with_multi_query.ipynb.

Star History

Contact

Do you have questions or want to collaborate? Please open an issue or email Taha Ababou at [email protected]

If this project helps you, consider buying me a coffee ☕. Your support helps me keep contributing to the open-source community!

The notebooks and visual diagrams were inspired by Lance Martin's LangChain Tutorial.