TL;DR

This publication introduces an open-source toolkit for building Retrieval-Augmented Generation (RAG) systems using Wikipedia data. It provides a reproducible pipeline for data ingestion, chunking, embedding, vector indexing, and LLM-powered retrieval, enabling rapid experimentation and deployment of RAG architectures for research and production.

1. Introduction

1.1 Purpose

This toolkit demonstrates how to build a scalable RAG system using Wikipedia as a knowledge base, leveraging modern embedding models and LLMs. It aims to accelerate research and prototyping in retrieval-augmented NLP.

1.2 Why It Matters

RAG systems are foundational for trustworthy, up-to-date, and context-aware AI applications. This project lowers the barrier to entry for practitioners and researchers by providing a ready-to-use, modular, and extensible codebase.

2. System Overview

• Data Source: Wikipedia dumps (configurable)
• Chunking: Configurable chunk size and overlap
• Embedding: Pluggable sentence-transformers (default: all-MiniLM-L6-v2)
• Vector Store: FAISS-based index for efficient retrieval
• LLM Integration: Supports Groq API (Llama 3 models)
• Config: All parameters via .env file

3. Technical Details

3.1 Architecture Diagram

(Wikipedia → Chunker → Embedder → Vector Store → Retriever → LLM)

3.2 Installation & Setup

git clone https://github.com/SosiSis/Deep-Learning-Wikipedia-RAG
cd deep-learning-wikipedia-rag
pip install -r requirements.txt
cp .env.example .env

Edit .env with your API keys and preferences

3.3 Usage Example

from rag_pipeline import RAGPipeline

pipeline = RAGPipeline()
answer = pipeline.query("What is the history of deep learning?")
print(answer)

3.4 Extensibility

• Swap embedding models via config
• Plug in different LLM providers
• Adapt for other corpora

4. Evaluation

4.1 Technical Validation

• Unit tests included (tests)
• Example queries and expected outputs
• Performance benchmarks (retrieval latency, accuracy on sample questions)

4.2 Limitations

• Wikipedia-only by default (other corpora require -preprocessing)
• LLM responses depend on external API quality
• No built-in fine-tuning for embeddings

4.3 Future Directions

• Add support for multi-lingual corpora
• Integrate more LLM providers (OpenAI, Anthropic, etc.)
• Enhance chunking strategies for better context

5. Documentation

• README.md: Quickstart, architecture, usage, and troubleshooting
• API Docs: Docstrings and Sphinx-generated documentation
• Examples: Jupyter notebooks for end-to-end demos

6. Contact & Support

GitHub Issues: https://github.com/SosiSis/Deep-Learning-Wikipedia-RAG/issues
Maintainer Email: sosinasisay29@gmail.com

7 References

RAG: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
FAISS: Facebook AI Similarity Search
Sentence Transformers

8 Appendix

.env.example: Configuration template
Sample Data: Example Wikipedia chunks and embeddings