With RAG, an assistant receives specific passages from its knowledge base before answering. Source citation displays this information beneath the answer: a document title, a section, sometimes a link to the original page.
What it looks like
A visitor asks: “Is the museum open on Mondays?” The assistant replies: “No, it is closed on Mondays; it opens Tuesday to Sunday, from 10 am to 6 pm” and displays “Visitor information › Opening times and prices” underneath. The visitor can check with one click, and the team can see the associated passage.
Why it matters
- For the visitor, it is a way to check. They can read the original rule rather than a paraphrase, which matters for a price, a deadline or a condition
- For the team, it is a way to review answers. Reading back through conversations shows which document was used and which needs correcting
- For the business, it is a record. Linking an answer to an approved document makes it possible to trace what was said and on what basis. In Canada, in February 2024, a tribunal held Air Canada responsible for inaccurate information given by its website chatbot (Moffatt v. Air Canada decision)
What a source does not guarantee
A cited source makes an answer verifiable, not infallible. The document may be outdated, or the passage misunderstood: this is why sources have a review date. An answer with no source is also a signal of a possible hallucination or a question outside the intended scope.
Source citation and answer engines
The same principle applies to generative answer engines, which cite the pages their answers come from. A website is more likely to be cited when it provides short, dated answers that make sense out of context.
KernChat assistants cite the source of their answers, and each source is versioned within its knowledge space (see the features).