Making RAG good

Why it matters

A RAG demo takes an afternoon. A RAG system people trust takes measurement. Most bad answers aren't the model's fault: the right passage never made it into the prompt. The fixes in this lesson (hybrid search, reranking, checked citations) are what separates a support bot that gets 6 in 10 questions right from one that gets 9 in 10, and you can only tell which you have if you measure.

What you'll be able to do

• Measure retrieval with a small labeled test set (hit rate and MRR)
• Combine keyword and embedding search with reciprocal rank fusion
• Explain reranking and when it's worth the extra call
• Check that an answer's citations point at documents you actually retrieved

Before this lesson

• Retrieval-augmented generation

Next in this phase: Hallucinations and grounding