SimpleRAG

A simple RAG (Retrieval-Augmented Generation) implementation using LlamaIndex and OpenAI. This project demonstrates different approaches to building a RAG system with increasing levels of sophistication.

Features

• Basic RAG implementation with document indexing and chat interface
• Advanced version with text splitting and chat history
• Dynamic data version with automatic reindexing on data changes
• Persistent storage of vector indices
• Streaming responses for better user experience

Prerequisites

• Python 3.10+
• OpenAI API key

Installation

1. Clone the repository:

git clone git@github.com:sajithamma/simplerag.git
cd simplerag

2. Install dependencies:

pip install -r requirements.txt

3. Create a .env file in the project root:

OPENAI_API_KEY=your_api_key_here

Project Structure

The project contains three main implementations:

1. app.py - Basic RAG implementation
2. app-advanced.py - Advanced version with text splitting and chat history
3. app-dynamic-data.py - Full version with automatic reindexing

Directory Structure

simplerag/
├── data/               # Directory for your documents
├── storage/           # Directory for storing vector indices
├── app.py            # Basic implementation
├── app-advanced.py   # Advanced implementation
├── app-dynamic-data.py # Full implementation
└── requirements.txt   # Project dependencies

Usage

Basic Version (app.py)

Simple RAG implementation with basic document indexing:

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.llms.openai import OpenAI

# Load and index documents
data = SimpleDirectoryReader(input_dir="./data").load_data()
index = VectorStoreIndex.from_documents(data)

# Create chat engine
chat_engine = index.as_chat_engine(
    llm=OpenAI(model="gpt-4"),
    verbose=False
)

# Chat interface
response = chat_engine.stream_chat("Your question here")

Advanced Version (app-advanced.py)

Includes text splitting and chat history:

from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.base.llms.types import ChatMessage, MessageRole

# Configure text splitting
text_splitter = SentenceSplitter(chunk_size=512, chunk_overlap=10)

# Create index with text splitting
index = VectorStoreIndex.from_documents(
    data,
    transformations=[text_splitter]
)

# Chat with history
chat_history = []
user_message = ChatMessage(role=MessageRole.USER, content="Your question")
chat_history.append(user_message)
response = chat_engine.stream_chat(user_message.content, chat_history=chat_history)

Dynamic Data Version (app-dynamic-data.py)

Includes automatic reindexing when data changes:

# Check if data has changed
needs_reindex = has_data_changed() or not os.path.exists("storage/docstore.json")

if needs_reindex:
    # Create new index
    index = VectorStoreIndex.from_documents(data)
    index.storage_context.persist(persist_dir="storage")
else:
    # Load existing index
    index = load_index_from_storage(storage_context)

Running the Application

1. Place your documents in the data/ directory
2. Run any of the implementations:

# Basic version
python app.py

# Advanced version
python app-advanced.py

# Dynamic data version
python app-dynamic-data.py

3. Start chatting with your documents!

Features by Version

Basic Version (app.py)

• Simple document indexing
• Basic chat interface
• Streaming responses

Advanced Version (app-advanced.py)

• Text splitting for better context
• Chat history support
• Persistent index storage
• Improved response quality

Dynamic Data Version (app-dynamic-data.py)

• All advanced features
• Automatic reindexing on data changes
• File state tracking
• Efficient index management