You already met two limitations of LLMs: they can hallucinate confident-sounding falsehoods, and their knowledge has a fixed cutoff date baked into their training. Retrieval-Augmented Generation, or RAG, is the standard fix when you need answers grounded in a specific, possibly-changing body of knowledge.
The idea: instead of trusting the model's memory, look up the relevant information first, hand it to the model as context, and have the model answer from that -- rather than from what it happened to learn during training.
Try it: ask the course a question
This retrieves from the lessons you've just read, so you can see exactly which passages fed the answer.
Knowledge Check
5 questions — answer all, then submit
1. What problem does RAG primarily solve?
2. What does a text embedding capture?
3. Why is chunking strategy important in a RAG system?
4. What is a vector database optimized for?
5. How is RAG different from fine-tuning for adding knowledge?