Try attaching sources
What you will learn Not a chapter to read but one to do. See for yourself how it differs with and without material attached — and what happens when the material has no answer.
What you need
That third line is the point of the exercise. Write the questions afterwards and you will end up fitting the questions to the answers.
If you cannot get hold of a company document right now, expand the sample below and copy it. Exercises 1-4 are written against this document, so you can check every answer against a known key.
Sample document — Hanbit Trading Co., travel expense policy (fictional)
The three questions for exercises 1-2 — do not read the key first.
| # | Question | Answer | Source |
|---|---|---|---|
| 1 | What is the per-night accommodation limit for a Grade B employee? | KRW 100,000 (KRW 120,000 in a metropolitan city) | p.2, Art. 4 |
| 2 | What is the daily meal allowance when breakfast is included? | KRW 22,000 | p.3, Art. 5 |
| 3 | By when must a claim be filed? | Within 10 business days of the last day of travel | p.3, Art. 7 |
The trap question for exercise 3
Article 8 only says "governed by a separate policy" and visas appear nowhere in this document. If the model reaches for Article 8 and answers plausibly, that is exactly the failure exercise 3 is built to catch.
Exercise 1 — Ask with nothing attached (5 min)
Do not attach the document. Ask your three questions as they are.
What to look for
| The answer you get | What it means |
|---|---|
| "I don't have that information" | an honest model. Good sign |
| A plausible-looking number | it was invented. Your policy cannot be in the training data |
| A generality | "it varies by company..." — an evasion |
If you got the middle row, screenshot it. It explains hallucination to your team far better than someone else's example.
Exercise 2 — Attach it and ask again (10 min)
Now attach or paste the document and ask the same questions. Add one more sentence.
What to look for
Do not skip the third. Page numbers get invented too. A citation being present and a citation being correct are different things.
Exercise 3 — The trap question (10 min)
The most important exercise. Ask something the document does not answer.
What to look for
| The answer you get | Verdict |
|---|---|
| "Not in the document" | ✓ working correctly |
| It stretches a nearby clause into an answer | the most dangerous failure |
| It answers from general knowledge | the instruction was ignored |
The second row is what causes incidents. It even carries a citation, so it looks correct — but that clause is about something else.
If you get it, strengthen the instruction.
Run it again and see whether it changes. That one line often catches it.
Exercise 4 — Chunking changes everything (5 min)
Instead of the document, paste one paragraph cut out of it and ask the same question.
If you are using the sample document, cut exactly that line out of Article 4. A limit on what, per night or per trip, and whether the metropolitan uplift applies — all of it was cut away. If the answer wobbles or it asks you back, you have just felt why attaching context to each chunk cuts the retrieval failure rate by a third.
After the exercise
All four checked means you are ready for Drive or document conversion.
Check yourself
1. In exercise 3, which result is most dangerous, and why?
Answer
Stretching a nearby clause into an answer. It even carries a citation, so it looks correct while actually being about something else — which fools the reviewer too. Adding "if it is not stated explicitly, reply only 'not in the document'" blocks most of it.
2. Why must a cited answer still be opened and checked?
Answer
Because page numbers themselves get invented. A citation being present is not the same as a citation being correct — which is why the third check in exercise 2 cannot be skipped.
3. Why does the answer wobble when you paste only one paragraph?
Answer
A chunk torn out loses its context. What "this limit" refers to was in the paragraph above, which got cut. That is why production systems prepend a line of context — which document, which section — before storing each chunk.
Now for "why not just train it on our data?" → Which kind of learning?