Answer

How does a chatbot answer from our own documents?

Your documents are indexed by meaning, not by keyword. When someone asks a question, the assistant searches the index, retrieves the two or three relevant passages and answers using only those, with a link to the source. If nothing relevant is found, it says so instead of inventing an answer.

Last updated: 22 September 2026

Why answering from sources is the safe design

A model answering from memory will occasionally produce a confident, wrong answer. A model answering from a retrieved passage can be checked, because the passage is shown next to the answer. For internal policy, warranty terms or technical specifications this difference decides whether the tool is usable at all.

The documents matter more than the model

Four versions of a price list produce four possible answers, and the assistant has no way of knowing which one you meant. Deciding what the single valid version of each document is, and deleting the rest from the index, is usually half the project and the half that pays off on its own.

Once the index is clean, keeping it current is automatic: when a document changes in the shared folder, the index updates and the old passage stops being used.

Internal and external versions

The same index can serve staff and customers with different permissions. Internally it can see margins, supplier names and internal procedures; the customer-facing version sees only what you publish. It is one build with two access levels, not two projects.

Asked next

How many documents does it need to be useful?

Often fewer than twenty. A price list, the warranty terms, the main procedures and the product sheets already cover most of what people ask.

What if two documents contradict each other?

The assistant flags the conflict and shows both sources rather than picking one. That flag is usually the first time anyone notices the contradiction.