RAG with Memory¶

This guide demonstrates how to implement a RAG system with conversation memory using AdalFlow, based on our github_chat reference implementation.

Overview¶

The github_chat project is a practical RAG implementation that allows you to chat with GitHub repositories while maintaining conversation context. It demonstrates:

Code-aware responses using RAG
Memory management for conversation context
Support for multiple programming languages
Both web and command-line interfaces

Architecture¶

The system is built with several key components:

Data Pipeline¶

Input Documents → Text Splitter → Embedder → Vector Database

The data pipeline processes repository content through:

Document reading and preprocessing
Text splitting for optimal chunk sizes
Embedding generation
Storage in vector database

RAG System¶

User Query → RAG Component → [FAISS Retriever, Generator, Memory]
                  ↓
              Response

The RAG system includes:

FAISS-based retrieval for efficient similarity search
LLM-based response generation
Memory component for conversation history

Memory Management¶

The memory system maintains conversation context through:

Dialog turn tracking
Context preservation
Dynamic memory updates

This enables:

Follow-up questions
Reference to previous context
More coherent conversations

Quick Start¶

Installation:

git clone https://github.com/SylphAI-Inc/github_chat
cd github_chat
poetry install

Set up your OpenAI API key:

mkdir -p .streamlit
echo 'OPENAI_API_KEY = "your-key-here"' > .streamlit/secrets.toml

Run the application:

# Web interface
poetry run streamlit run app.py

# Repository analysis
poetry run streamlit run app_repo.py

Example Usage¶

Demo Version (app.py) - Ask about Alice (software engineer) - Ask about Bob (data scientist) - Ask about the company cafeteria - Test memory with follow-up questions
Repository Analysis (app_repo.py) - Enter your repository path - Click “Load Repository” - Ask questions about classes, functions, or code structure - View implementation details in expandable sections

Implementation Details¶

The system uses AdalFlow’s components:

core.embedder.Embedder for document embedding
core.retriever.Retriever for similarity search
core.generator.Generator for response generation
Custom memory management for conversation tracking

For detailed implementation examples, check out the github_chat repository.