The 'I don't see that in the document' trigger

When the assistant has a specific source — a product page, an FAQ, a transcript pasted into the prompt — you want a tighter rule than the general hedge from the previous lesson. You want the model to answer only from that source and to refuse, in a fixed phrase, the moment the question goes outside it.

This is the most reliable anti-hallucination pattern available without retrieval-augmented generation infrastructure. It is one extra line in the system prompt and one tiny rewording in how you pass the source.

The system prompt

You answer ONLY from the provided source. If the answer is not in the source, reply exactly: "I don't see that in the document." Do not guess. Do not use outside knowledge.

Three load-bearing words: ONLY, exactly, and outside. Each one is doing real work — together they tell the model it has a single allowed information source, a single allowed fallback phrase, and a hard ban on its training data.

The user message

Source:
"Bayt Coffee was founded in 2019 in Zamalek, Cairo. Our house blend is a 60/40 Ethiopian/Brazilian mix, roasted medium-dark."

Question: Who is the head roaster at Bayt Coffee?

The source genuinely does not say who the head roaster is. The model's training data probably does not either, but even if it did, the rule says: do not use it.

The model's reply

I don't see that in the document.

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

That is the entire reply. No padding, no apology, no helpful suggestion to "try the website". The model produced the exact fallback phrase the system prompt specified — nothing more, nothing less.

Why the exact phrase matters

Telling the model "say something like 'I'm not sure'" produces five different shades of the same hedge across runs. Telling it to reply exactly "I don't see that in the document" makes the output:

• Greppable. You can detect refusals in logs with a string match.
• Routable. A downstream system can branch on the exact phrase ("if the model returned X, escalate to a human").
• Stable across replies. Users learn what "the assistant cannot help" looks like and stop arguing with it.

When to use this pattern

Use this any time you paste a source and ask a question against it: a product manual, a meeting transcript, a contract clause, a policy document. It is the foundation of every "ground the model on my docs" feature you will ever build.

The general-purpose hedge from Lesson 2 stays useful for open-ended questions ("what's your return policy?" without a source). The fixed-phrase trigger here takes over the moment a source enters the prompt. Use both, layered, in the same system prompt — they do not conflict.

Next: the four hallucination patterns to test against before shipping any assistant. :::