Building RAG from Scratch — Embedding, Cosine Similarity, Chunk Search (with Korean vs English Comparison)
Carving out the inside of RAG (Retrieval-Augmented Generation) with actual code, from what embedding even is to debugging suspicious results. I pulled 1536-dimension vectors with OpenAI's text-embedding-3-small, compared them with cosine similarity, compared Korean vs English performance, and worked through chunk splitting. Surprisingly, Korean embedding similarity turned out to be lower.
