Citations in UI

In a RAG-based copilot, every answer is composed from retrieved passages. Citations expose that: numbered markers or source cards next to the answer that open the exact article section, contract clause, ticket or database record used. Good citations are precise (a section, not a 60-page PDF), current (the version that was actually retrieved) and honest (no citation is shown for a claim the sources do not support).

Technically, citations depend on chunk metadata captured at ingestion (chunking with title, section, page, version and URL), on prompting the model to attribute claims to specific passages, and on a groundedness check that removes or flags unsupported statements. Rendering them well is a UX task: hover previews, highlighted passages and a "searched but found nothing" state when retrieval returned no usable source.

Citations are not the same as listing the top retrieval results below an answer. That shows what was found, not what was used, and users learn to ignore it. Nor are they a legal guarantee of correctness; they are the mechanism that makes checking cheap enough that people actually do it.

• Product manager: Citations are the single UI element most correlated with users trusting and continuing to use a copilot.
• Compliance officer: A cited, versioned source for every answer is what makes AI output defensible in an audit or complaint.
• Support manager: Agents can paste the cited article link to the customer, which turns AI drafts into resolutions rather than guesses.
• CTO / Head of Engineering: Citation precision is a direct test of ingestion quality; vague citations mean the metadata and chunking need work.

An answer without a source asks the user to trust the machine; an answer with a source asks them to trust the document, which they already do. Citations exist to shift trust to something verifiable and to make wrong answers cheap to catch. The trade-off is engineering and honesty: citations require metadata discipline at ingestion, a groundedness check to prevent decorative or fabricated references, and design work to keep them readable. A citation that opens the wrong page erodes trust faster than no citation at all.

• SaaS help copilot that cites the exact help-centre section and opens it in a side panel
• Bank staff assistant citing the circular number, clause and effective date behind each policy answer
• Clinical assistant linking each recommendation to the hospital protocol paragraph it came from
• AI tutor citing the textbook chapter and page for every explanation given to a student
• Logistics operations copilot citing the contract SLA clause when explaining a penalty

Technique / practiceA UX and retrieval engineering practice that depends on chunk metadata and groundedness checks. It is standard in Eazyware's SaaS Copilots and Retrieval & Knowledge Engineering deliveries and is verified in evals rather than assumed.

Eazyware service that covers it: AI Copilot Development for SaaS. Starting prices are on the pricing page.