What RAG is and why it grounds AI

A language model knows a lot, but its knowledge is frozen at training time, generic to the whole internet, and — critically — it can't tell you where an answer came from. Ask it about your company's refund policy and it will confidently invent something plausible. Retrieval-Augmented Generation (RAG) fixes this by retrieving relevant information from a knowledge base you control and appending it to the prompt, so the model answers from supplied evidence rather than fuzzy memory.

The mechanic is simple: when a question comes in, you search your data for the most relevant passages, put those passages in the prompt, and ask the model to answer using them. That's it. The sophistication is all in retrieving the right passages — the rest of this course.

Why RAG matters, concretely:

• It reduces hallucination. The model is steered to answer from text you provided and can verify against, instead of guessing from training data that may be wrong or outdated.
• It gives you citations. Because every answer traces to a retrieved source, you can show the user where it came from — a trust feature and a correctness check in one.
• It updates without retraining. Change a document, re-index it, and the model's answers change. No fine-tuning, no model retraining.
• It uses private, current data. Your internal wiki, today's records, this customer's contract — none of which the model was trained on.

This is the same reliability goal as the enterprise-prompting course ("cite or abstain"), but automated: instead of pasting sources by hand, you build a system that fetches them. For anything where an answer has to be accurate, current, and defensible — support, internal knowledge, compliance, docs — RAG is the standard architecture. By the end of this course you'll build one that retrieves well, proves what it knows, and holds up in production.