AI answers with citations and no hallucination: how Duaer does it

Why does AI make things up?

A general chat model writes what sounds likely. It doesn’t know your company’s policies, and it rarely says “I don’t know”. Ask it about something internal and you often get a plausible, wrong answer with nothing to check it against.

The fix isn’t a smarter model. It is making the model answer only from your own documents and show what it relied on.

How does Duaer keep answers tied to sources?

• Search first, then answer. Duaer finds the passages the person may read and gives only those to the model.
• Every statement is numbered. [1] and [2] in the answer link to the original documents below it. In Feishu and WeCom, the reply ends with Sources.
• No source, no answer. If nothing matches, Duaer doesn’t call the model and shows “Nothing you may see matches.” An answer without a valid citation is not shown either.
• Follow-ups work the same way. With Ask a follow-up, when the documents can’t answer, Duaer says “What you may read does not answer this follow-up.”

What do you do when an answer is off?

1. Click Helpful or Not helpful under the answer. Duaer records it.
2. Open the source. If the original is out of date, fix it where it lives, and Duaer reads the new version on its next pass.
3. For a question that comes up often and is easy to get wrong, write a verified answer. Duaer shows it first, labeled Verified answer, and marks it due for review when its review date arrives (180 days by default).

How can you tell whether an AI answer tool is trustworthy?

During a trial, ask questions you already know the answers to and check four things:

• Can you open the original behind every answer?
• Does it say so when the documents don’t cover the question?
• Does it ever use content the person isn’t allowed to see?
• How long after a document changes does the answer change?