Why RAG gets hard
From five notes to a whole handbook
RAG has three steps: retrieve the parts of your documents that match a question, augment the prompt with them, and let the model generate an answer from them. With five short notes, almost anything works. With real documents, each step can quietly go wrong:
| Problem | What happens | You'll fix it in |
|---|---|---|
| Messy text from PDFs | Odd characters, page headers and "Page 3 of 9" mixed into the content | Module 1 |
| Bad chunks | A rule cut in half, or a chunk about three different things | Module 2 |
| Old versions | The 2024 handbook says pay day is Thursday; the 2026 one says Wednesday | Module 2 |
| Different words | "How long is lunch?" vs "a 30-minute unpaid meal break" | Module 3 |
| Exact names and codes | "KeyPay", "FSS-208841" — meaning search can miss them | Module 4 |
| No proof it works | It seems fine on the three questions you tried | Module 4 |
| Made-up answers | The model fills gaps, or mixes facts up | Module 5 |
The big lesson of this course: RAG quality is mostly about retrieval. If the right text reaches the model, even a small model answers well. If it doesn't, no model can.
What you'll build
The Corner Café Staff Handbook Assistant. Staff ask things like "How do I swap a shift?" or "A customer has a nut allergy — what do I do?", and get a short answer with numbered sources that link to the exact page of the right document. It searches seven documents:
- The Staff Handbook 2026 (a 9-page PDF with a pay-rates table) — and the old 2024 handbook, which disagrees with it
- The Food Safety Program and the Espresso Machine Guide (PDFs with tables)
- The Refunds and Complaints and Privacy policies (Word files)
- A Staff FAQ (Markdown)
They're made up, but written like the real thing — and the techniques work the same on your workplace's documents.
Try it
- Think of the documents at a job you know: where are they kept, how many are there, and what questions do people ask about them?
- Download the course files from this lesson.
Downloads
1 file to download once you enrol.