Answers with receipts: why citations matter at work.
Before veridive helped companies, it answered questions from videos with a citation to the exact second. The lesson carried over: people trust answers they can check.
veridive5 min read
veridive began as an answer engine for the spoken web. You asked a question; it searched YouTube videos, podcasts, lectures and interviews live, and answered with a citation to the exact second in each source. The timestamp mattered as much as the answer. One click, and you could hear the claim in the speaker’s own words, in context, and decide for yourself.
Building that product taught us something we now apply to every system we build for organizations: people trust answers they can check. At work this matters even more than on the open web, because answers lead to decisions, and decisions have owners.
What did the answer engine teach us?
Three lessons from building it shaped how we build for clients.
- The citation is part of the answer. It is not a footnote for the curious. The answer and its evidence belong on the same screen, at the same moment.
- Precision beats volume. Pointing to a whole four-hour recording is barely better than no source at all. Pointing to the second where the claim is made changes what people can do with it.
- Sources need judgment. Not every source deserves equal weight. Deciding which sources count, and which don’t, is a design decision that belongs with the people who own the knowledge.
Why does checking have to be fast?
Nobody verifies an answer if verification takes longer than doing the work from scratch. A citation that points to a sixty-page policy is technically a source and practically useless. A citation that points to one paragraph, with the sentence quoted, takes seconds to confirm. In the answer engine, that was the difference between a link to a video and a link to the moment.
The speed of checking changes behavior. When checking is fast, reviewers actually check, errors get caught, and trust grows on evidence rather than hope. When checking is slow, people either stop using the system or, worse, stop checking and start trusting it blindly.
What makes a good citation?
- It points to the exact place. The paragraph, the table row, the clause or the second in a recording, not just the document.
- It quotes the words. A short excerpt lets the reader confirm the claim without opening anything.
- It shows which version. Policies change. A citation should show the document’s date or version, so nobody relies on an outdated rule.
- It respects permissions. People are only shown sources they are allowed to open.
- It supports one claim. Each important statement gets its own source, rather than a list of documents at the end.
People trust answers they can check.
What should happen when there is no source?
The system should say so. If no approved source supports an answer, the honest response is “I couldn’t find this in the documents I can access”, followed by the closest passages it did find and a suggestion of who to ask. That answer is less impressive in a demo. In daily use it is worth more than a confident paragraph with nothing behind it, because it tells people exactly when to be careful.
Designing for “no source found” also shows where the knowledge base has gaps. Every unanswered question is a note for the content owner: something people need that nobody has written down yet. It also guards against a subtler risk: an answer that is mostly right, with one invented detail, is much harder to spot than one that is plainly wrong.
How do citations change the review?
Without citations, reviewing an AI answer means redoing the research. With them, the reviewer checks the evidence: does this passage say what the answer claims? That is faster, and it catches a different kind of error. A citation that does not support its claim is a visible warning sign, even when the answer sounds right.
Citations also help after the fact. When a decision is questioned weeks later, the record shows what the system said, which source it relied on and who approved it. Accountability becomes a matter of looking things up, not of reconstructing memories.
How do you know the citations are right?
Measure them, as you would measure answer quality. For each example in the evaluation set, check two things: is the answer correct, and does the cited passage actually support it? The second check matters because a system can be right for the wrong reason, or cite a real document that says something slightly different. Track the share of answers whose citations fully support them, review a sample by hand every week, and treat an unsupported citation as seriously as a wrong answer.
What changes for audio and video?
Calls, meetings and training videos need the same discipline, with two additions. Transcripts must keep their timestamps and speaker labels, so every claim links back to the moment and the person who said it. And the playback link must open at that moment, for people who are allowed to hear the recording. Retention matters more here too: agree up front how long recordings and transcripts are kept, and where they are processed.
From videos to the working day
The same principle applies wherever knowledge lives. In voice and meeting intelligence, a finding about a sales call links to the exact second it happened. In knowledge and document intelligence, an answer about a contract quotes the clause behind it. The medium changes. The rule doesn’t: every answer should come with a receipt. If your team already asks “where does that number come from?”, you are halfway there.
Ask an assistant about this note