Document Q&A With Citations: How to Check an AI Answer

To check an AI answer and its citations, work sentence by sentence: find the passage each claim points to, read that passage in full, and confirm it says the same thing, including the figures, the conditions and the exceptions. With a little practice it takes a few minutes per answer. Below you will find a five-step method, a worked example with numbered passages, the warning signs of invented or misleading citations, and the situations where reading the raw passages is the better option.

What a citation does, and does not, prove

Document Q&A tools answer from a defined set of files: a staff handbook, a library of contracts, technical documentation, a collection of HSE or HMRC guidance. For each question, the system retrieves a few relevant passages and has a language model write the answer from them. If you want to see that process end to end, read how RAG works, step by step.

The citation is the bridge between the answer and the evidence. It lets you decide for yourself instead of trusting the tone of the reply. What it does not guarantee is that the model understood the passage, kept every qualification, or avoided adding something from its general training. Teams building these systems tackle that with strict grounding and the right to refuse, as covered in reducing AI hallucinations. This guide is for the person on the receiving end: the reader who has to decide whether the answer is good enough to act on.

A five-step method for checking an AI answer

1. Pin down the question. Write in one line what you actually need to know. Then confirm the answer responds to that question, not to an easier one the documents happened to cover.
2. Pair each claim with its marker. Every factual sentence should carry a reference such as [1] or [2]. Underline the sentences without one: they are either glue or the model's own contribution.
3. Read the full passage, not the snippet. Look at the whole excerpt and, if the interface allows, the text around it. Compare figures, units, dates and modal verbs one by one.
4. Identify the document. Note its title, version, date and author. An accurate quote from a policy that was replaced last year is still the wrong answer today.
5. Ask what is missing. Would a specialist add an exception, a threshold, a different rule for Scotland or for part-time staff? If so, ask a targeted follow-up or search the passages directly.

Most problems surface at steps 2 and 3. Steps 4 and 5 catch the quieter failures: answers that faithfully reflect a passage but leave out the part that applies to you.

Worked example: an answer with three numbered passages

The following case uses a fictional organisation's staff handbook. An employee asks: "Can I claim for my home broadband if I work from home?" The tool replies:

Yes. Staff working from home receive a broadband allowance of £25 a month [1]. It is paid through payroll [2]. You can also order a monitor of up to £200 [3].Generated answer (fictional example)

The three passages shown underneath say this. Passage [1], Hybrid Working Policy (updated February 2026): "Colleagues on a contractual home-working arrangement of three or more days a week may claim a broadband contribution of £25 per month, with line manager approval." Passage [2], Payroll Guide: "Allowances are paid in arrears in the following month's salary." Passage [3], IT Equipment Procedure (2022): "Office-based staff can request a second monitor via the IT service desk."

The first claim is the trap. The figure matches, so a quick glance says "verified", yet the conditions have gone. Someone who works from home one day a week would take the "Yes" at face value. The third claim shows citation drift: the passage sits on a related subject (IT equipment), so the reference looks reasonable, but it concerns office staff and contains no £200 limit. Until a passage shows that figure, treat it as made up.

The dates matter too. The 2022 IT procedure predates the 2026 hybrid policy, so even a matching sentence would need confirming.

Warning signs of invented or misleading citations

• A figure you cannot find in the cited text. Sums, percentages, deadlines and paragraph numbers should appear verbatim. If they do not, they are not sourced.
• Right theme, wrong rule. The passage is about the same area but a different group, place or situation: employees versus contractors, England versus Scotland, office versus home.
• Firmer language than the source. "Must" in the answer, "should" or "may" in the passage. Qualifiers such as "normally" or "unless" quietly disappear.
• Exceptions dropped. The main rule is there, the "only where" clause is not.
• Everything cites [1]. One passage used as a blanket reference for several unrelated claims.
• Quotation marks around words that are not in the passage. A quote should match exactly; a reworded "quote" is a red flag.
• Legislation or sections that appear nowhere in the passages. If the answer names an Act, a regulation or a page the retrieved text does not contain, it came from the model's memory.
• Undated or anonymous sources. No title, no date, or an older version competing with a newer one.
• A tool that never says it does not know. If every question gets a confident reply, gaps are being filled somewhere.

Passages only or a written answer?

Some tools can skip the written answer and return only the retrieved excerpts. For several kinds of question, that is the more reliable route.

Kopik offers both on every knowledge base. The standard answer is written only from the retrieved passages, with numbered citations, and says when the base does not hold the answer. Passages mode returns the relevant excerpts with no written answer. Through MCP the two tools are `ask_base` (answer plus numbered source passages) and `search_base` (passages only), billed at the same price, so choosing the raw text for a sensitive check costs nothing extra. The developer documentation explains how to connect.

Building trust with your users

If your organisation offers a document Q&A tool to staff or customers, make checking easy rather than relying on goodwill:

• Display the passages beside the answer, with document title and date, not just a bare number.
• Treat "not covered by the documents" as a valid reply. People trust a tool more once they have seen it decline.
• Retire superseded documents so old and new versions do not contradict each other.
• Audit a sample each month. Take ten real questions, run the five steps, and correct the source documents where answers fall short.
• Say where a human decides. For HR, legal or safety matters the answer is a starting point. If the tool processes personal data, the ICO publishes guidance on AI and data protection under UK GDPR that is worth reading.

To practise, try a public base such as UK workplace health and safety: HSE guidance for employers. Ask about a specific duty, then check each sentence against the numbered passages.

Checklist before acting on an answer

• The answer responds to my precise question.
• Each figure, date and obligation carries a citation.
• I found every key figure in the cited passage.
• Conditions and exceptions from the passage appear in the answer.
• The document is current and comes from the right author.
• No Act, section or document is mentioned that the passages do not contain.