Lesson 1: Why chunking matters, and the simple splitters

Module 3 · Lesson 1 · About 40 minutes · Needs: Module 1 (what retrieval returns) and a first look at embeddings (a vector per text, compared by cosine similarity)

The situation

Kwame's prototype embeds every Hollin document whole. Ask it "which C70 programme uses 40 ml of water" and it hands the model the entire C70 user guide: 841 characters, of which "Espresso | 40 ml" is 16. It works on a guide this short. It will not work on a 40-page manual.

So Kwame tried the quick fix: cut every document into 200-character pieces. The same question now retrieves the "Before first use" paragraph, and the answer is gone. Nothing crashed. Grace wants to know how a smaller index made the answer worse. This lesson answers her with the simple splitters every RAG system starts from. The guide and questions are the course's simulated data.