Why your assistant should show its sources

A company knowledge assistant has one job: answer staff questions from the company’s own documents. Ask it how many days of leave carry over, or which clause covers late delivery in a supplier contract, and it should give the right answer quickly.

The problem is that a wrong answer looks exactly like a right one. Language models write fluently and confidently whether or not they’re correct. If people can’t tell the difference, they end up either trusting everything, which is dangerous, or trusting nothing, which makes the assistant pointless.

Source links fix this. Every knowledge assistant we build links each answer to the document and passage it came from. Here’s why we think that matters more than almost any other feature.

People can check the answer in seconds

When an answer comes with a link to section 4.2 of the leave policy, checking it takes one click. People do click, especially at first and especially when the stakes are high. Over time they learn which kinds of question the assistant handles well, and they check less often. That’s trust built on evidence rather than hope.

Without sources, the only way to check an answer is to go and find the document yourself, which is the job the assistant was supposed to save.

It keeps the assistant honest

Assistants like ours work by retrieval-augmented generation, or RAG. Before answering, the system searches your documents for relevant passages and gives them to the model along with the question. The model is told to answer only from those passages.

Requiring a citation for every claim makes that rule much easier to enforce and test. If the model writes something it can’t link to a retrieved passage, that’s a sign it’s filling gaps from its general training, which is where many wrong answers come from. A well-built assistant will say “I couldn’t find this in the documents I have access to” instead of guessing, and the citation requirement is what makes that behaviour possible.

It shows you where your documents are the problem

Once answers carry sources, wrong answers become easy to diagnose. Usually the cause is one of these:

• The document is out of date. The assistant correctly quoted the 2023 travel policy, but nobody removed it when the 2025 version was published.
• Two documents disagree. The handbook says one thing and a team wiki page says another.
• The answer isn’t written down anywhere. It lives in someone’s head or an old email thread.

None of these is a model problem, and no amount of prompt tuning will fix them. Source links point you straight at the document that needs attention. Many of our clients find that the assistant’s biggest benefit is a cleaner document library.

It respects who’s allowed to see what

Sources also make permissions visible. An assistant should only retrieve documents the person asking is allowed to open, and the citation shows exactly which document it used. If someone in sales gets an answer citing a board paper, you’ve found an access problem before it becomes a leak.

What good citations look like

Not all source links are equal. A useful citation:

• links to the specific passage or page, not just the file;
• shows the document’s title and date, so people can spot outdated sources;
• appears next to the claim it supports, not in a list at the bottom;
• opens the document in the system where it already lives, with the user’s normal permissions.

The trade-off

Citations make answers slightly longer, and they force the system to say “I don’t know” more often. Some people find that disappointing at first. We think it’s the right trade. An assistant that admits gaps gets used for years, while one that bluffs tends to get quietly abandoned after its first confident mistake.

If you’re comparing knowledge assistants, ask every vendor to show you an answer it got wrong and how a user would have noticed. The answer tells you a lot.