RAG is a sandwich. Find a few passages. Put them next to the question. Make the model answer from the paste. If the passages are wrong, the answer is wrong with better posture.
Aside. The robot does not get your badge. Your manager still reads the email with your name on it.
Fine-tuning is sending the intern to a semester abroad. RAG is putting the binder on the desk. Most jobs need the binder. Most jobs can start with you pasting the two pages. A vector database is for when the binder is huge and the same questions return.
When this is the job
- You have a document you are allowed to use (SOP, handbook excerpt, ticket macros)
- Answers must stay inside it
- “I don’t know” is better than a confident shrug
Skip it when there is no document, or the question is “should we fire them,” or you will ask once — just paste the pages.
The packet that fails closed
Answer only from the passages below.
If they do not contain the answer, reply NEED and quote nothing.
Cite the heading or page I pasted.
Question: [question]
Passages:
[paste]
If your company has a blessed “chat with our wiki” tool, use that instead of uploading the handbook to a random bot.
You will not
Build a startup that “does RAG for SMBs.” This course will not help you sell that. You will learn enough to use the internal tool without treating it as gospel.
Try this today
Paste one real policy section. Ask a question you know is in it, and one you know is not. The second answer should be NEED. If it answers anyway, the packet is weak or the tool is sloppy.
You are done when
You have one refuse-to-guess prompt and you have seen it refuse.