Citations link each claim in the answer back to the chunk it came from.
They turn a system a user must trust into one a user can check.
That difference is usually what makes RAG acceptable in a business at all.
Each chunk is given an identifier in the prompt and the model is asked to reference it inline.
Numbered markers are the common pattern, resolved to links in the interface.
Requiring a citation per claim also improves the answer itself.
A model that must attribute is less inclined to invent.
Always validate the identifiers before displaying them.
Models cite chunks that do not exist, and a broken citation is worse than none.
Check that the cited chunk actually supports the sentence, which a small model can verify.
Citing a real chunk that says something else is a subtle and common failure.
Show the quoted passage on hover or expansion so verification takes one click.
Cite at the sentence level rather than dumping sources at the end.
An answer with no citations should be treated as ungrounded and flagged.
Citations are also your debugging tool, since they show exactly what retrieval delivered.