RAG from Scratch

A minimal RAG implementation built from the ground up without framework abstractions. Follows the core loop: documents → chunk → embed → store → query → retrieve → generate. The goal is to understand each step of the RAG pipeline before reaching for tools like LangChain.

What's Inside

• Chunking — configurable text splitting with overlap
• Embedding — sentence-transformers for vector representations
• Vector store — cosine similarity search implemented with NumPy
• Generation — context-augmented prompts sent to Ollama or Gemini
• Modular architecture — separate modules for each pipeline stage