RAG means look up your documents first, then answer. It is how Amazon Quick can quote your own policies instead of guessing.
It learned from public text. Ask about your own policy and it can only guess. Switch between the two and compare.
Your documents
The conversation
What it has not read. Your staff handbook, procurement policy, last week's circular, today's vendor proposal and this morning's meeting notes were never part of its training.
Two bad options without lookup. It says "I don't know", which is useless. Or it writes a plausible answer based on what other organisations usually do, which is a hallucination: confident and made up.
What lookup adds. Answers can point to a document and section. New files can be used as soon as you add them, with no retraining of the model.
Amazon Quick does this for you. Watch your question find the right pages, then click any stop.
Your library
In general RAG systems, your question and every piece of every document are turned into lists of numbers that capture meaning. Pieces whose numbers sit closest to the question's numbers are the best matches. That is why a search for "annual leave" can find text that says "vacation days".
Amazon Quick manages all of this for you. You do not choose any of these settings, and this page does not describe how Quick works inside.
Search works on small pieces, called chunks. Change the cut and see which piece wins, and whether the answer survives.
The AI receives the best-matching piece
Why cut at all? A 200-page handbook as one block would blur every detail. Smaller pieces keep the detail so the right one can be found.
Cutting by structure. Splitting by section heading or paragraph keeps ideas together better than cutting every fixed number of characters, which can slice a word in half.
In Amazon Quick. This is the general idea behind RAG. Quick manages it for you, so you do not set a piece size. What you control is the quality of the documents: clear headings and short sections help.
Rules, the picked pieces and your question are stacked into one prompt. Click a layer to see what it does.
Each layer has a different job. Together they let the AI answer from your pages.
The rules layer asks for a source, and the picked pieces carry their section labels. So the answer can say "per Staff Handbook section 4.3". That makes it checkable, which matters for HR, compliance and procurement work.
Everything shares one limited space, called the context window. More pieces mean less room for others, so a focused set of pieces beats a large pile.
Click each source to see what it controls. Then try the Web Search switch below.
| Situation | Why it is hard | What to do |
|---|---|---|
| Answer is spread out | The facts sit in many places, and only a few pieces are picked | Ask narrower questions, one at a time |
| Totals across many files | Adding up needs every figure, not a few pieces | Use a spreadsheet or finance system |
| Old and new versions | The agent may quote the older policy | Keep only the current version in the Space |
| Not in your files | There is nothing to look up | The agent should say "not covered", so ask it to |
More sources. Quick can also connect to places where content already lives. Which ones are available depends on what your administrator has enabled. For data classification, retention and audit questions, ask your DNA / IT team.