RAG from Scratch

Building a Retrieval-Augmented Generation pipeline from the ground up — chunking, embedding, vector search, and LLM-powered Q&A with RAGAS evaluation metrics.

---

Why This Project?

Most RAG tutorials use high-level abstractions that hide the mechanics. This project builds every component from scratch so you understand exactly what happens at each stage — from raw text to cited answers.

---

Architecture

graph LR
    A[PDF / Text Documents] --> B[Document Loader]
    B --> C[Text Chunking]
    C --> D[Embedding Generation]
    D --> E[ChromaDB Vector Store]
    
    F[User Query] --> G[Query Embedding]
    G --> H[Similarity Search]
    E --> H
    H --> I[Context Assembly]
    I --> J[LLM Generation]
    J --> K[Answer + Sources]
    
    K --> L[RAGAS Evaluation]

Loading

---

Key Features

Feature	Description
Document Loading	PDF and text file ingestion with metadata extraction
Semantic Chunking	Recursive text splitting with configurable overlap
Vector Embeddings	OpenAI/HuggingFace embedding models
Similarity Search	Cosine similarity over ChromaDB vector store
LLM Generation	Context-grounded answer generation with source citations
RAGAS Evaluation	Automated metrics — faithfulness, relevance, context precision
Flask Chat UI	Interactive web interface for document Q&A

---

Pipeline Deep-Dive

1. Document Ingestion

• Load PDFs and text files using custom parsers
• Extract metadata (filename, page number, section)
• Handle encoding edge cases

2. Chunking Strategy

• Recursive character text splitter
• Configurable chunk size (default: 1000 tokens) and overlap (200 tokens)
• Preserves paragraph boundaries where possible

3. Embedding & Indexing

• Generate dense vector embeddings (OpenAI text-embedding-3-small or HuggingFace alternatives)
• Store in ChromaDB with metadata for filtered retrieval
• Persistent storage for production use

4. Retrieval & Generation

• Embed user query → cosine similarity search → top-k retrieval
• Context window assembly with source tracking
• LLM generates grounded answers with inline citations

5. Evaluation (RAGAS)

• Faithfulness: Is the answer supported by retrieved context?
• Answer Relevancy: Does the answer address the question?
• Context Precision: Are the retrieved chunks relevant?
• Context Recall: Did retrieval capture all needed information?

---

Quick Start

Prerequisites

• Python 3.11+
• OpenAI API key

Setup

git clone https://github.com/Nagavenkatasai7/rag-from-scratch.git
cd rag-from-scratch/rag-project
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

Configure

export OPENAI_API_KEY="your-key-here"

Run

python app.py

Open http://localhost:5000 to use the chat interface.

---

Project Structure

rag-from-scratch/
├── rag-project/
│   ├── app.py              # Flask chat UI
│   ├── rag_pipeline.py     # Core RAG pipeline
│   ├── chunking.py         # Text splitting logic
│   ├── embeddings.py       # Embedding generation
│   ├── retriever.py        # Vector search & retrieval
│   ├── evaluation.py       # RAGAS evaluation metrics
│   ├── templates/          # HTML templates for Flask UI
│   └── data/               # Sample documents
├── RAG-FROM-SCRATCH-GUIDE.md  # Detailed implementation guide
├── requirements.txt
└── README.md

---

Tech Stack

Component	Technology
Language	Python 3.11+
Framework	LangChain, Flask
Vector Store	ChromaDB
Embeddings	OpenAI / HuggingFace
LLM	GPT-4o / GPT-3.5-turbo
Evaluation	RAGAS
Frontend	Flask + HTML/CSS

---

What I Learned

• How chunking strategy directly impacts retrieval quality
• Why overlap matters — without it, answers miss context at chunk boundaries
• RAGAS evaluation reveals failure modes invisible to manual testing
• ChromaDB's metadata filtering enables precise document-scoped queries

---

License

MIT — see LICENSE for details.