Why RAG exists
Ask a model ten veterinary questions with no documents to help it, and keep a record of where it is wrong, vague or out of date. In Module 4 you ask the same ten questions of the assistant you build, and compare.
Colab opens a fresh copy each time. Save your own with File, then Save a copy in Drive.
What you do in this module¶
The questions are about Riverside Veterinary Hospital, a made-up clinic whose practice protocols come with the course. They include a drug dose, a protocol that changed, a species trap, a "no history of" condition, clinic shorthand, and one question that no document covers.
The model has never seen the clinic's protocols, just as it has never seen your own. Your notes go on the sheet below, not in the notebook: a notebook forgets everything when Colab disconnects.
Before you start: which question will a bare model handle worst?
Probably the one no document covers. A model rarely says "I don't know". A gap in what it learned often comes out as a confident, plausible answer. Watch for that in question 10, then check whether your guess held.
Three ways to give a model your documents¶
- Put them in the prompt. Fine for one or two documents. It does not scale to hundreds.
- Retrieval-augmented generation (RAG). For each question, first find the few passages most likely to hold the answer, then have the model answer only from those, citing them. This course builds one.
- Fine-tuning. Train the model further. Good for style and habits, poor for facts that must stay exact and current, and it cannot say where an answer came from.
The notebook explains when each one fits.
What the bare model got wrong
Note the question numbers under each heading, with a few words on each. The notebook prints a numbered summary you can copy from. Download your copy when you finish: Module 4 asks you to compare against it.
What you type here stays in this browser on this device only. Nothing is sent to AVI or anyone else. Do not include client names or anything that could identify a client or patient.
The answer contradicts the clinic's answer, or would harm a patient.
Nothing false, but too general to act on.
Right once, but not now. Question 2 is about a protocol that changed.
Include the predictions you made before asking, and whether they held.
Where next
Building a RAG system: Module 2: Documents, chunking and embeddings
A shared question channel is on the way. When it opens, each answer will be written once and shared with everyone taking the course.