An AI answer with three citations looks more convincing than an answer with none. It is also easier to inspect. Those are useful properties, but neither proves the answer is right.
A document can be real, the link can work, and the sentence beside it can still claim something the document does not say. Before relying on a document assistant, we recommend checking the connection between each important claim and the evidence behind it.
This is especially relevant to retrieval-augmented generation, or RAG. RAG searches a document collection and gives relevant passages to a model before it writes an answer. That helps ground the response in your material. It does not remove the need to test what the model does with that material.
A working citation is the start of the check
There are three separate things to establish: the cited passage exists, it supports the statement, and it is an appropriate source for this decision.
Structured citation features can provide valid pointers into documents, including page or character locations. That solves a useful navigation problem. It does not mean every business conclusion beside that pointer follows from the text.
For a hypothetical example, imagine a supplier’s policy says unopened standard items may be returned within thirty days, but custom orders are excluded. An assistant replies that all items may be returned within thirty days and cites the policy.
The citation is real. The deadline is right. The answer is still wrong because it dropped the conditions that determine eligibility. A citation check needs to catch that, not merely count a link as a pass.
Check the whole claim, including the small words
Start with sentences that could change what someone does. Dates, amounts, exclusions, eligibility, and statements containing words such as “all,” “only,” or “must” deserve particular attention.
Open the passage and compare it with the answer. Mark a claim as supported only when the evidence backs its full meaning. A source supporting half a sentence does not rescue the unsupported half.
For the return-policy example, a useful reviewer note would say: “Deadline supported; item eligibility overstated.” That is more actionable than “bad answer” and more honest than a passing score based on the correct deadline.
We recommend keeping the answer narrow when the evidence is narrow. If the retrieved text establishes a deadline but not eligibility, the response should preserve that distinction rather than supply a likely-sounding rule.
Look beyond the highlighted sentence
A passage can lose meaning when separated from its heading, table, footnote, or neighboring paragraph. Inspect the surrounding material when an answer depends on a qualification.
In our hypothetical policy, an exception might sit under a separate heading on the next page. If the search returns only the general return window, even a faithful summary of that passage can leave the reader with the wrong impression.
This is why groundedness and completeness should be separate checks. Groundedness concerns whether the answer stays within its evidence. Completeness concerns whether it covers what the question requires. A short, fully supported answer can still omit the part that matters.
Do not fix every omission by asking for a longer response. First establish whether the missing material was available to the model at all.
Confirm which source governs the answer
An accurate summary of an obsolete policy is not a reliable answer about today’s policy. Likewise, a document for one product does not automatically apply to another.
For each important test case, we recommend recording the applicable document, its version or effective date, and the scope it covers. Compare those details with the source the assistant actually retrieved. A recent upload date alone should not settle which policy governs.
If two documents conflict and there is no established precedence rule, the assistant should expose the conflict rather than silently choose one. The person responsible for the material needs to resolve the underlying ambiguity. Better wording cannot settle an unresolved business rule.
Separate a search problem from an answer problem
When a response fails, save the question, retrieved passages, answer, and citations together. Then inspect where the evidence was lost.
If the required paragraph exists in the collection but never reached the model, investigate retrieval. If the paragraph reached the model but its exception disappeared from the answer, investigate how the response was generated. If the original file never contained an answer, do not classify the result as a failure to retrieve one.
Those distinctions keep a team from changing the wrong component. Rewriting a prompt will not restore a page that failed text extraction. Improving search will not, by itself, prevent a model from overstating a passage it already received.
Test the cases where a confident answer would be a mistake
A useful review set should contain more than straightforward questions with one obvious source. Include a policy with an exception, two conflicting versions, a question requiring multiple passages, and a question the collection cannot answer.
Write the expected outcome before examining the model’s response. Sometimes that outcome is a supported answer. Sometimes it is a request for missing context or a clear statement that the available documents do not establish the answer.
Automated scoring can help screen results, but we recommend having someone familiar with the source material review important failures and a sample of passing answers. Treat the score as evidence to investigate, not permission to stop reading.
Make verification part of the product
Useful citations should take readers to the relevant passage, not leave them searching through a long file. Show enough context to inspect qualifications, and make the document identity and version understandable.
We build document-based AI systems with retrieval, evaluation, and the interface considered together. For a business choosing its first use case, the practical goal is not an assistant that always sounds certain. It is an assistant whose useful answers can be checked and whose limits are visible.
If your documents are few and staff can already find the governing answer quickly, custom RAG may not be necessary. If the search and verification work is becoming the bottleneck, talk to us about the workflow. Start with the questions your team needs answered and the evidence a good answer must carry.