How does RAG cite its sources?

A RAG system cites its sources by tracking which retrieved chunks were fed to the model and attaching their metadata — document title, URL, page, or section — to the answer. Because every response is built from specific passages the system pulled, it can point back at those exact passages, either as inline markers next to each claim or as a source list at the end.

How are citations actually attached?

The key is that each chunk carries metadata from the moment it's indexed. When retrieval hands a handful of chunks to the model, the system remembers where each one came from, then links the generated text back to them. Common approaches:

• Inline markers — numbered references like [1] or [2] placed next to individual sentences, so each claim maps to a source.
• Appended source list — a "Sources" block at the bottom listing every document that fed the answer.
• Chunk metadata — title, URL, page number, and section stored alongside each embedding and surfaced in the output.
• Highlighted passages — the exact retrieved text shown so a reader can verify the answer against the original.

Do citations mean the answer is grounded?

Not automatically. Good grounding means the model was told to answer only from the retrieved passages and to say "I don't know" when they don't cover the question. Citations are strongest when they're generated from the retrieval step itself rather than written freehand by the model — otherwise you get the hallucinated-citation problem, where the model invents a plausible-looking reference that doesn't support the claim, or points at a real document that never says what's quoted.

Why do citations matter so much?

For enterprise use, they're often the whole point. A citation lets a person verify the answer instead of trusting it blindly, turns the system from a black box into something auditable, and gives legal, medical, and finance teams a paper trail. An answer nobody can check is hard to act on — a cited answer you can click through to is what makes RAG trustworthy enough to deploy.