Lesson 12
Tools
The model doesn’t get smarter. Its window gets better text.
Every “the AI checked the web” is the same trick. The model writes a request as text, ordinary software runs it, and the result gets pasted into the window.
Two ways to get knowledge in: bake it, or hand it over
Training (lesson 05) melts facts into the dials. Powerful, slow, and frozen solid the moment it finishes. Retrieval (“RAG”) is the document lookup you just stepped through. It changes nothing about the model and everything about what’s in the window. The difference shows up the instant a fact changes. So change one:
① Baked in by training
you: What's our refund window?
② Handed over by retrieval (RAG)
retrieved into the window just now
you: What's our refund window?
| Training | Retrieval (RAG) | |
|---|---|---|
| To update it | Retrain or fine-tune (GPUs, weeks, money) | Save the new document. Done in seconds |
| Freshness | Frozen at the cutoff (lesson 11) | As fresh as the document store |
| Can it cite a source? | No. Knowledge is smeared across a trillion dials | Yes. The paragraph is sitting right in the window |
| Best for | Skills: language, reasoning, broad world knowledge | Your facts: policies, docs, prices, anything that changes |
| Failure mode | Confidently stale, or invented (lesson 09) | Retrieval grabs the wrong doc, and the model reads it anyway |