MMOOZZGG.com

Glossary

What are AI citations?

In short

AI citations are the ability of an AI system to show in its answer which document, and which place in it, each claim comes from, so a reader can check it.

Search and RAG · citations

MMOOZZGG.com editorsChecked: 1 min read

For a child

When you write a school report, the teacher asks you to name the book where you found the facts. Then anyone can open the book and check. A good computer helper should also say: “I read this here.” AI citations are these honest notes about where each piece of the answer came from.

In plain words

An AI answer, even a correct one, is hard to check when you do not know where it came from. So systems that answer from documents add links to the phrases: the document name, the page, and the exact piece of text. The user clicks the link and sees the source paragraph. But a link alone does not prove the answer is right: the model can cite a document that says something else. For example, an employee asks about the travel allowance limit and gets an answer with a link to an internal order. The employee opens the order and in a few seconds sees that the amount matches.

For business

Source links are needed wherever an assistant’s answer affects money or obligations: notes for lawyers, answers to customers about contract terms, and internal rules. They cut checking time and let staff trust the system for good reasons. But you must measure the quality of the links. A 2023 audit of generative search engines found that on average only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported their sentence. Some model providers offer a built-in citation mode that returns the exact passage of a document. Before launch, compare a sample of answers with the documents: does the cited text match the claim? Nobody launches an assistant on company documents without this test.

For a specialist

In retrieval-augmented generation, a citation (also called attribution) links parts of the answer to specific passages of the context. There are three approaches. You can instruct the model to put chunk numbers after claims. You can train the model to generate citations, as in Self-RAG. Or you can attribute after the fact, when a separate natural language inference (NLI) model matches each claim to a source. The ALCE benchmark (Gao, Yen, Yu, Chen, 2023) rates fluency, correctness, and citation quality. The authors report that on the ELI5 set, even strong models lack full citation support in 50% of cases. Liu, Zhang, and Liang (2023) tested Bing Chat, NeevaAI, Perplexity, and YouChat: on average 51.5% of sentences were fully supported by citations, and 74.5% of citations supported their sentence. Some APIs offer structured citations. In the Claude API, text and PDF files are split into sentences, and the answer returns cited_text with a pointer to the place: character indexes, PDF page numbers, or block indexes. According to the documentation, such citations always contain valid pointers to the documents provided, and cited_text does not count toward output tokens. The metrics are citation recall and citation precision.

Example

A lawyer asks the assistant about the limitation period under a work contract. In the answer, each claim has a link, and a click opens the right article of the code and the paragraph from an internal memo.

Related terms

Questions and answers

Does a source link mean the answer is correct? No. The model can cite a document that says something else, or the right document with a wrong summary.

The link only lets you check this quickly.

How does AI know what to cite? Documents are split in advance into numbered chunks, and the model gives the numbers of the chunks it relies on. Some APIs do this themselves and return the exact quoted piece of text.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us