Introduction to retrieval-augmented generation concepts: Review

The ‘Exercise - Implement a basic RAG pattern’ unit made my heart sink. It’s not an exercise. It’s a link to a GitHub repo with a fully baked solution, and the instructions are just to read the code. There's no sandbox, no local setup, not even a 'fill in the blank' line. This is the part I was waiting for, the moment a concept clicks because you’re forced to build it, but it’s just a reading assignment. I clicked through the repo, skimmed the Python script, and felt like I'd just been handed the answer key to a test I never got to take.

The rest of the module is a series of well-made PowerPoint slides converted to a webpage. The diagrams explaining the flow from user prompt to vector database to LLM are clean, probably the best part of the whole thing. But the text feels like it was written for a sales deck about Microsoft Foundry. Every other paragraph seems to hint at a Microsoft product. It teaches the *what*—what RAG is, what grounding means—but completely skips the *how*. How do you choose an embedding model? What are the trade-offs between different chunking strategies? What happens when the retrieval step fails? None of that is here.

It’s a 35-minute vocabulary lesson. I now know the official definition of RAG and can recognize the diagram. For that, it’s fine. But if you’re hoping to walk away ready to try building one, you'll be disappointed. This is pre-learning for the real learning.