Use this when

An AI answers from documents or data, and a person may need to trust, check or defend the answer.

The pattern

QUESTION
   ↓
AI ANSWERS
   ↓
ANSWER + "based on: [source, version, date]"
   ↓
person can open the source and verify

Why it works

An answer with a citation can be checked in seconds. An answer without one has to be trusted or redone. Attaching the provenance, which source, which version, when, turns a claim into something verifiable.

It also changes behavior for the better. A model instructed to cite its source is nudged toward answering from retrieved material rather than from memory, which reduces invented answers.

How to apply it

  1. Have the workflow retrieve passages before answering. See retrieve then answer.
  2. Return the source reference with every answer: title, section, version and date where possible.
  3. Make the citation easy to open, not just a name.
  4. Instruct the model to say when it has no source, rather than answering anyway.

Good fit

  • Policy, product and documentation questions.
  • Any answer with consequences, where a person must be able to verify it.
  • Situations where you will later need to trace where an answer came from.

Poor fit

  • Casual, low-stakes answers where a citation adds clutter and no value.

What this does not solve

Citing a source shows where an answer came from; it does not make the source correct. A well-cited answer from a wrong or outdated source is still wrong. Citation makes errors catchable, not impossible.