In words
What it is, why it matters, and what it is like.
Why am I learning this?
You're learning this so that when an AI answers a question, you can see *where* that answer came from. This is what turns a chatbot that confidently says something false into a tool you can actually trust in your work. Once you master this, you'll be ready to study RAG deeper — how to retrieve the right passages in the first place, how to measure if retrieval quality is good, and how to prevent hallucination. Without citations, every AI answer is a black box. With them, you can check, question, and verify. This is also a critical requirement in regulated fields like law, medicine, and finance — where an answer that can't point to its source might as well be wrong.
The idea, in plain terms
Imagine you're writing a school essay or a report at your job. You make a claim like 'India's GDP grew by 7% last quarter.' Where did you get that number? If you can't point to the government report, the news article, or the spreadsheet where you saw it, your boss or teacher has every right to be suspicious. You might have misremembered, or the source might be wrong, or you might have just made it up to sound knowledgeable. Now imagine you asked an AI assistant the same question. A normal AI model just predicts the next word based on patterns it learned during training — it has no actual memory of 'facts'. So when it says '7%', it's not reading it from any document; it's generating a word that *sounds* plausible. It can be completely wrong, and it won't know. That's called hallucination. Citations fix this problem. Instead of the AI just saying the answer on its own, we give it a stack of real documents (through retrieval), and we *force* it to show its work: 'Here is my answer, and here are the exact sentences from the documents that support it.' This makes the mistake visible. If the citation doesn't match the answer, you immediately see the lie. If there's no citation at all, you know to be suspicious. It turns trust from a leap of faith into a simple check.
An analogy
Think of a courtroom witness. A good witness doesn't just say 'I saw him commit the crime.' They must show evidence: 'I saw him at 7pm, and this is the CCTV footage at that time.' The court doesn't trust a bare claim. It demands the witness *point* to the specific piece of evidence that backs each statement.
A citation in an AI answer works exactly the same way. The AI is the witness. The retrieved passages are the CCTV footage, the fingerprints, the signed document. When the AI makes a claim, it must raise its hand and say 'This claim is supported by this exact sentence in this document.'
But the analogy breaks down in one important way. In court, the witness is *expected* to remember what they saw and connect it to evidence. The AI model does not remember anything in that sense. It has no memory of 'facts'. It only knows patterns of words. So the citation is not the AI *recalling* where it learned something — it is the system *showing you the document it read* at the same moment it generated the answer. The document is right there in front of it, and the citation is just a pointer saying 'this is the sentence that I am basing my next word on.' This is more like a witness who is allowed to keep the CCTV footage on the table in front of them and point to it: they can see it, you can see it, and everyone can verify the connection.
Definition
A system that makes the model's answer point directly to the specific retrieved passage that supports each claim, so that the chain from evidence to answer is transparent and checkable.
Where this sits
You have mastered nothing yet, so this is your very first concept. This is a *young* topic — it sits right at the heart of something called retrieval-augmented generation, a technique where a model is given a set of retrieved documents and asked to answer based on what they say. That is your parent concept, listed in your library. From here, you will naturally move on to learning about *chunking* (how documents are split into retrievable pieces of evidence — this determines what a citation can legally point to), *retrieval quality* (measuring whether the right passages were even found), and then *why hallucinations happen*. This concept connects to your notes on *Admissible Evidence Spans* — in regulated setting, citations aren't just about transparency, they're about *permission*: the model may only cite passages that are legally allowed to influence an answer.
