Ahmed-Nezar/RAGPlayground

RAGPlayground

RAGPlayground is a project for experimenting with Retrieval-Augmented Generation (RAG) pipelines. It integrates PostgreSQL + pgvector for semantic search, Hugging Face models for embeddings and generation, and an MCP server that allows GitHub Copilot Agent Mode to interact with the knowledge base.

The project demonstrates two workflows:

1. Classic RAG: Hugging Face LLMs are used to generate responses based on retrieved context.
2. Agent Mode: GitHub Copilot uses MCP (Model Context Protocol) to query the knowledge base.

---

Features

• End-to-end RAG pipeline (data ingestion → embeddings → database → retrieval → LLM response)
• PostgreSQL with pgvector for vector similarity search
• SentenceTransformers for embeddings
• Hugging Face Inference API for LLM chat completion
• MCP server for Copilot Agent integration
• Automatic database and table creation if not present
• Interactive CLI loop for queries with colored prompts

---

RAG Pipeline Flow

---

Project Structure

RAGPlayground/
├── README.md
├── LICENSE
├── main.py
├── requirements.txt
└── src/
    ├── db/
    │   └── rag.py          # Database setup, table creation, embeddings insertion
    ├── extract/
    │   └── processing.py   # Dataset download, preprocessing, embeddings generation
    ├── llm/
    │   └── llm.py          # Hugging Face client, context retrieval, LLM querying
    └── mcp/
        └── server.py       # MCP server for GitHub Copilot Agent

---

Installation

1. Clone the repository git clone https://github.com/ahmednezar/RAGPlayground.git cd RAGPlayground

2. Install dependencies pip install -r requirements.txt

3. Install PostgreSQL and enable the pgvector extension Inside psql: CREATE EXTENSION IF NOT EXISTS vector;

4. Configure environment variables in a .env file:

   HF_TOKEN=your_huggingface_token_here
   

---

Dataset

The project uses the AG News Classification Dataset from Kaggle. It is downloaded automatically using kagglehub when you run the pipeline, then preprocessed into embeddings.

---

How to Use

Running with Hugging Face Models (Classic RAG)

python main.py --llm mistralai/Mistral-7B-Instruct-v0.3 --embedding BAAI/bge-base-en-v1.5

Example:

[RagPlayground] Enter your query: Tell me about soccer news
[RagPlayground] Response: <contextual answer based on retrieved docs>

Running in MCP (Copilot Agent Mode)

python main.py --mcp --embedding BAAI/bge-base-en-v1.5

• Starts an MCP server at localhost:5001.
• Copilot Agent will use the retrieve_documents tool to fetch context before answering.
• The agent is guided by instructions defined under .github/instructions/.

---

Architecture

1. Data Layer: Downloads AG News dataset, generates embeddings.
2. Storage Layer: Embeddings stored in PostgreSQL with pgvector.
3. Retrieval Layer: Similarity search using cosine distance.
4. Generation Layer: Either Hugging Face LLM or Copilot Agent with MCP.

---

Screenshots

Classic RAG

RAG + MCP

---