Document Q&A With Citations: How to Check an AI Answer
To verify AI answer sources, treat every citation as a promise you can test: match each sentence of the answer to the passage it cites, open that passage, and confirm it says the same thing, with the same numbers, conditions and exceptions. It takes two to five minutes per answer once you know what to look for. This guide gives you a five-step method, a worked example with numbered passages, the red flags of hallucinated citations, and the cases where you are better off reading the passages yourself.
Why citations matter in document Q&A
A document Q&A tool answers questions from a fixed set of files: an employee handbook, a contract library, product documentation, a collection of regulatory guidance. Behind the scenes, it retrieves a handful of relevant passages and asks a language model to write an answer from them. We explain that pipeline in how RAG works, step by step.
Citations are what make this different from asking a general chatbot. They give you a path back to the text, so you can decide whether to trust the answer instead of taking it on faith. But a citation only proves that a passage was shown to the model. It does not prove the model read it correctly, kept the conditions attached to a rule, or refrained from adding something it remembered from elsewhere. Builders reduce those errors with grounding and refusal rules (see how to reduce AI hallucinations with grounded answers); this article is about the other side: you, the reader, checking the result.
The one-sentence rule
An answer is only as reliable as the weakest sentence you did not check. On anything you will act on (a payment, a deadline, a policy decision, a reply to a customer), check the sentences that carry numbers, dates, obligations and exceptions.
How to check an AI answer in five steps
- Restate the question. Before reading the answer, say in one line what you need to know. Then check that the answer addresses that exact question, not a neighboring one that was easier to find in the documents.
- Map each claim to a citation. Go sentence by sentence. Every factual statement should carry a marker such as [1] or [2]. Note the sentences with no marker at all: those are either connecting phrases or content the model supplied on its own.
- Open the passage and read it whole. Do not stop at the highlighted words. Read the full passage, and the sentence before and after if the tool lets you. Check that the numbers, units, dates, verbs (must, should, may) and conditions match.
- Check the document behind the passage. Which file is it? Which version, which date, which issuer? A correct quote from a superseded policy is still a wrong answer.
- Look for what is missing. Ask yourself what an expert would add: an exception, a threshold, a different rule for another state or category of employee. If in doubt, ask a follow-up question that targets the gap, or search the passages directly.
Steps 2 and 3 catch most problems. Steps 4 and 5 catch the subtle ones, the answers that are faithful to a passage but still incomplete or out of date.
Worked example: reading an answer with numbered passages
Here is a realistic case, built on a fictional company handbook. An employee asks: "Can I get reimbursed for my home internet if I work remotely?" The tool returns this answer, followed by three numbered passages.
Yes. Employees who work remotely receive a monthly internet stipend of $50 [1]. The stipend is paid with your regular paycheck [2]. You can also claim the cost of a new desk chair up to $300 [3].Generated answer (fictional example)
The passages returned with it read as follows. Passage [1], from the Remote Work Policy (revised January 2026): "Employees approved for remote work at least three days per week receive an internet stipend of $50 per month, subject to manager approval." Passage [2], from the Payroll FAQ: "Recurring stipends are paid with the second paycheck of each month." Passage [3], from the Office Equipment Guide (2023): "Employees working on site may request ergonomic equipment through Facilities."
Checking each claim against its passage
| Claim in the answer | What the passage says | Verdict |
|---|---|---|
| Remote employees get $50 per month [1] | Only if approved for at least three remote days a week, and subject to manager approval | Partly supported: conditions dropped |
| Paid with your regular paycheck [2] | Paid with the second paycheck of each month | Roughly supported, but less precise than the source |
| Desk chair up to $300 [3] | On-site ergonomic requests through Facilities; no amount, no remote rule | Not supported: citation does not back the claim |
The first sentence is the most dangerous kind of error: the number is right, so it looks verified, but the eligibility conditions disappeared. An employee who works remotely one day a week would read "Yes" and be wrong. The third sentence is a classic citation drift: the passage is on a nearby topic (equipment), so the marker looks plausible, but nothing in it mentions remote workers or a $300 cap. That figure came from somewhere else, possibly another company's policy the model has seen, and you should treat it as invented until a passage shows it.
Notice also the dates. Passage [3] comes from a 2023 guide, older than the 2026 remote policy. Even if it had mentioned a chair budget, you would want to confirm it still applies.
Red flags of hallucinated or misleading citations
You do not need to check everything with the same intensity. These signals tell you where to look first:
- A precise figure that does not appear in the cited passage. Amounts, percentages, deadlines and section numbers should be findable word for word. If you cannot find the figure, assume it is not sourced.
- A passage on the right theme but the wrong subject. The topic matches, the actual rule does not (on-site vs remote, employees vs contractors, one state vs another).
- Stronger wording than the source. The answer says "must" where the passage says "should" or "may", or drops words like "generally" or "unless".
- Conditions and exceptions missing. The rule is quoted, the "only if" clause is not.
- One citation carrying everything. Five claims all point to [1]. Either the passage is very rich, or the model is using it as a blanket reference.
- Quotation marks around text you cannot find. A direct quote should match the passage exactly. A paraphrase in quotation marks is a warning sign.
- References to documents, laws or sections that are not in the passages. If the answer mentions a regulation or a page number that none of the returned passages contain, it came from the model's memory.
- Old or unidentified sources. No document name, no date, or a version older than another passage on the same topic.
- No "I don't know" ever. A tool that always produces a confident answer, even for questions outside its documents, is filling gaps on its own.
Fastest test for a suspicious number
Copy the number or the key phrase and search for it in the passages (Ctrl+F or Cmd+F works on most interfaces). If it is not there, the citation does not support it.
Passages mode vs generated answers: when to read the sources yourself
Some tools let you skip the written answer and get only the retrieved passages. It sounds like a step back, but for certain questions it is the safer and faster way to work.
Which mode to use
| Situation | Generated answer | Passages only |
|---|---|---|
| Quick orientation on a topic | Best: readable summary with pointers | Slower to read |
| Exact figure, deadline or legal wording | Risk of paraphrase | Best: you read the original text |
| Comparing several rules or versions | May merge them | Best: each passage stays separate |
| An AI agent reasoning on its own | Adds a second layer of interpretation | Best: the agent gets raw evidence |
| Checking whether the documents cover a topic at all | May fill gaps | Best: no passage means no coverage |
On Kopik, every knowledge base can answer in both ways. The default answer is written from the retrieved passages only, with numbered citations, and says so when the base does not contain the answer. The passages mode returns just the relevant excerpts, without a written answer. Over MCP, the two tools are `ask_base` (answer plus numbered source passages) and `search_base` (passages only), at the same price, so choosing passages for a high-stakes check costs nothing extra. Connection details are in the developer documentation.
Building trust with the people who use your Q&A tool
If you run a document Q&A tool for colleagues or customers, verification should not depend on everyone reading this article. A few design and process choices make checking natural:
- Always show the passages next to the answer, with the document name and date, not just a bare [1].
- Make "not in the documents" a normal answer. Users trust a tool more once they have seen it decline a question it cannot source.
- Keep sources current. Remove superseded versions so old and new rules do not compete; our guide on preparing documents for AI covers this.
- Spot-check regularly. Pick ten real questions a month, verify them with the five steps, and fix the documents where answers fail.
- Define what needs a human. For HR, legal or safety decisions, the answer is a starting point and a qualified person signs off. The NIST AI Risk Management Framework is a useful reference for writing that policy.
You can practice the method on a public base. The US workplace safety base built on OSHA material is a good test: ask a question about a specific obligation, then check whether each sentence of the answer is supported by the numbered passages.
Ask questions, read the passages
Browse knowledge bases built on official documents and check every answer against its numbered sources, or switch to passages mode and read the text yourself.
A two-minute checklist before you act on an answer
- The answer addresses my exact question.
- Every number, date and obligation has a citation.
- I found each key figure in the cited passage, word for word.
- Conditions and exceptions in the passage also appear in the answer.
- The source document is current and from the right issuer.
- Nothing in the answer refers to a document that is not in the passages.
Frequently asked questions
Can an AI invent citations even when it uses my documents?
Yes. A tool can attach a real passage number to a sentence the passage does not support, or mention a law or section that was never retrieved. Grounded systems make this rarer, but the only way to be sure is to read the cited passage and find the claim in it.
How long does it take to verify an AI answer?
For a short answer with three or four citations, usually a few minutes: map each sentence to its passage, search for the key figures, and check the document date. Long answers on complex topics take longer, which is a good reason to ask narrower questions.
What should I do when a citation does not support the claim?
Discard that sentence, not necessarily the whole answer. Ask a follow-up question targeting the unsupported point, or switch to passages mode and search for it directly. If no passage covers it, the documents probably do not answer that question.
Is a passages-only search better than a generated answer?
For exact wording, figures and comparisons, yes, because you read the original text with no paraphrase. For a quick overview, a generated answer with citations is faster. Many people use the answer first, then passages to confirm the critical points.
Are AI answers with citations reliable enough for HR or compliance decisions?
They are a strong starting point when the documents are current and authoritative, and when someone checks the cited passages. For decisions with legal or financial consequences, keep a qualified person in the loop who reviews the sources before acting.
Get the Kopik newsletter
New knowledge bases, RAG guides and product news. One email every week or two, unsubscribe in one click.
By subscribing you agree to receive our newsletter. We never share your address.