Retrieval-Augmented Generation (RAG) From Scratch

A RAG system built from first principles which implements chunking, embeddings, similarity search, and generation directly to show how RAG actually works under the hood.

How it works

1. Chunking: source documents are split into overlapping ~500-character chunks.
2. Embedding: each chunk is embedded into a 384-dim vector using all-MiniLM-L6-v2.
3. Retrieval: a query is embedded the same way, then compared against every chunk using cosine similarity to find the top-(3) most relevant chunks
4. Generation: the retrieved chunks are injected into a prompt and sent to Openai to generate a grounded answer

Setup

uv sync

Create a .env file with your API key:

OPENAI_API_KEY=

Add your source .txt files into data/.

Usage

uv run main.py

Example

q> who invented airplanes?

a< The Wright brothers, Wilbur and Orville Wright, invented the airplane.

Roadmap

• Native brute force retrieval
• [] FAISS/Chroma vector store
• [] Smarter chunking
• [] Evaluation harness measuring retrieval and answer quality.