MODULE 02 — HOW AI THINKS · 09 AI Concepts

Look it up before you answer

The fix for the last page is not a better model. It is fetching the right page of your handbook and putting it on the desk before the model says anything — a trick usually called RAG.

Ask the company handbook

Every dot is a chunk of the handbook, placed by meaning — the same kind of map as page 2, built from documents instead of words. Nobody sorted these into folders.

question → point→ find nearest chunks→ put them on the desk→ answer from what is there

Pulled onto the desk

The answer it builds

The problem it solves

Your handbook was never in the training data

WHAT IT READ — THE PUBLIC INTERNET your files behind a login no path in so it answers with everyone else’s average Retrieval does not teach it anything — it changes what is on the desk before you ask.

How it works

Three steps, and only the middle one is new

chunks the coordinate store
Chop the documents into pointsSame trick as page 2, applied to files instead of words.
“time off” lands by “annual leave” no shared words needed
Land the question on the mapMeaning, not keywords, decides what counts as near.
your question pasted in above, invisibly handed to the model together
Paste the nearest few inFrom here it is ordinary reading comprehension.

See it break

Citations are not proof

ON THE MAP — VERY CLOSE “policy on leaving” §2.1 §2.4 §2.7 retrieval worked perfectly IN MEANING — VERY FAR what was fetched time off work what was meant ending employment nothing in the system compares these no notice-period chunk exists anywhere on the map so “not covered here” was never reachable

Fluent, genuinely cited, and answering a question nobody asked. After adopting retrieval, hallucination gets rarer and confident irrelevance takes its place.

So what

Retrieval moved the hard problem. It is no longer what does the model know — it is whether the right paragraph got picked.