RAG-BluePrint is designed as a mini-book in the form of 10 curated Jupyter notebooks.
Each notebook focuses on one core component of the Retrieval-Augmented Generation (RAG) pipeline — with clean explanations and minimal, meaningful, fully runnable code.

No heavy frameworks.
No over-engineering.
Just the essential building blocks of modern RAG systems, presented clearly and practically.

RAG is an architecture for grounding LLM outputs using external knowledge.

This handbook walks through each stage of the pipeline:

1. Data Loading – reading and extracting clean text from PDFs, TXT, CSV
2. Chunking – splitting documents into retrieval-friendly units
3. Embeddings – converting text into vector representations
4. Vector Storage – indexing embeddings for efficient search
5. Retrieval – fetching the most relevant chunks
6. Reranking – improving retrieval precision
7. Generation – constructing a final answer using retrieved context
8. Evaluation – checking retrieval accuracy and grounding quality

Each stage is covered in a standalone notebook with examples and experiments.

01 – Introduction to RAG
02 – Loading & Preparing Your Data
03 – Chunking Strategies
04 – Embeddings: Concepts & Implementation
05 – Building a Vector Store with Chroma
06 – Basic Retrieval Techniques
07 – Your First RAG Pipeline
08 – Adding Rerankers
09 – Evaluating RAG (RAGAS)
10 – End-to-End RAG Pipeline (Blueprint)

Each notebook includes:

• A clear written explanation
• Flow diagrams (simple ASCII-style where helpful)
• Clean, reproducible code
• Real output examples
• A short summary
• Optional exercises to reinforce understanding

No clutter. No unnecessary visuals.

Clone and start:

git clone https://github.com/arorarishi/RAG-BluePrint
cd RAG-BluePrint
pip install -r requirements.txt
jupyter lab

Runs entirely on CPU.
No GPU required.
No API key required unless you choose to enable LLM calling via OpenAI / Gemini / Mistral / DeepInfra.

• Python 3.9+
• jupyterlab
• sentence-transformers
• chromadb
• pandas
• nltk
• scikit-learn
• ragas (optional, for evaluation)

All notebooks use small example datasets kept inside /assets for easy reproducibility.

Most RAG content online is either:

• Too shallow — tiny examples that teach nothing, or
• Too abstract — enterprise frameworks hiding all the core logic

RAG-BluePrint fills the missing middle:
A clear, structured, low-level, notebook-driven handbook that teaches
exactly how RAG works under the hood.

You will implement every part manually — chunking, embeddings, vector search, reranking, and evaluation — using only essential libraries.

This builds true understanding, not dependency on frameworks.

This project is actively maintained.
Advanced topics (HyDE, CRAG, RAPTOR, Agentic RAG, etc.) will be introduced in a separate companion repository once the Blueprint foundations are complete.

MIT License

Suggestions and educational improvements are welcome.
Feel free to open issues or submit pull requests.