Perspectives
Your documents do not need a smarter chatbot
They need answers that name their source.
By Ehsan Gazar, 22 October 2026
Somebody on your team asks where the latest travel policy is. Three people answer, with three different files.
One is the version from last spring. One is a draft that was never approved. One is right, and nobody can say which.
So the obvious fix gets proposed: put every document into a chatbot and let people ask it. The demo goes well. A month later, nobody quite trusts what it says.
The chatbot was not the problem. The problem is that an answer with no source is just another opinion in the room.
Why the first version gets ignored
A model asked about your documents will always produce something fluent. That is what it is built to do.
Fluent is not the same as found. When the answer comes back with no document attached, the reader has two choices. They can trust it, or they can go and look it up themselves.
Most people look it up. After a few weeks they stop asking the chatbot first, because it saves no time over searching the shared drive.
The answer to that is not a bigger model. It is a search that finds the right passage, and an answer that shows it.
Search by meaning and by the exact words
Questions about documents come in two shapes. "How much can I spend on a hotel?" is a question about meaning. The words "hotel allowance" may not appear anywhere.
"What does clause 14.2 say?" is the opposite. Only the exact words will do, and a search that understands meaning may return a paragraph that is merely similar.
Vatan Knowledge searches both ways at once. Each document is split into passages and indexed for meaning and for its exact words, and a query is matched against both.
That is why it finds the hotel limit when nobody typed "allowance", and finds clause 14.2 when somebody did.
An answer that shows where it came from
Every result Knowledge returns carries the document and the collection it came from. It is not a summary of the source. It is the passage itself, with its name on it.
When the Assistant answers from Knowledge, it marks each claim with the source behind it. The reader sees a short answer and, beside each part of it, the document that supports it.
That changes what the reader does next. They do not need to take the answer on trust, because the source is named beside it and they can check it.
It also changes what a wrong answer looks like. If the answer cites last spring's policy, the reader can see that it did. The fix is to retire the old file, not to argue with the model.
Each team sees its own documents
A company's documents are not one pile. Finance has files that sales should not read, and HR has files almost nobody should.
In Knowledge, documents live in collections, and each collection is opened only to the teams that need it. A search runs over the collections the person asking can see, and nothing else.
So the same Assistant can serve the whole company without putting every file in front of every person. Somebody in sales asking about pay bands gets nothing back from an HR collection they cannot open.
This is the part a quick demo skips, and the part an enterprise rollout cannot.
One set of documents, used everywhere
Most teams that try this end up with copies. The chatbot has one copy of the handbook, the website widget another, and the agent that writes weekly reports a third.
The copies drift apart within a month. The handbook gets updated in one place and the old version keeps answering in two others.
Knowledge is the one place documents are kept. The Assistant, agents, workflows and website widgets all read from it, rather than each holding its own copy.
Documents arrive by upload, from Notion pages, or from a website you point it at. When a document is updated in Knowledge, every tool that reads it reads the new version.
Where a team starts
The useful first step is small. Pick one collection people already ask about every week, such as the IT guides or the expenses policy.
Put that collection into Knowledge, open it to the teams who ask, and let them use the Assistant for a fortnight. Then ask them which answers they checked against the source, and which ones they found wrong.
The wrong answers are the useful ones. Each points at a document that is out of date, duplicated or unclear, and fixing it fixes every answer that reads it.
That tells you more than any demo. It shows whether people trust answers they can check, and which documents need rewriting before anything else is added.
An answer that names its source is one people can stop double-checking.