Chatbot

The term chatbot covers two very different products. The first is outward-facing: embedded on a website, answering visitor questions, where content is already public and the success metric is conversion.

The second is internal: content is confidential, "who can get an answer from which document" is central, and the success metric is a reduction in repeated internal questions. KobiGPT is built for the second scenario.

Comparisons made without that distinction mislead; the two products may share an interface but solve different problems and deserve different criteria.

In an internal assistant the most serious risk is not a wrong answer but a right answer to the wrong person. A viewer-level user obtaining an answer from the payroll file is a design failure independent of model quality.

Access control therefore belongs at the centre of the product. In KobiGPT, assistants see only the files attached to them, and user roles determine which assistant can be reached.

When evaluating, test that scenario directly: attach a restricted file to an assistant and try to query its content as an unauthorised user.

A chat interface is not the right tool for every problem. For a user who wants a single number, a report screen is faster; chat earns its place on questions that need context and cannot be predefined.

The second limit is currency: the assistant reflects the files attached to it and does not refresh itself. For fast-changing data, the source system is the right place, not a document index.

The third limit is document quality. A chat interface cannot produce knowledge that is not written down; a good interface does not compensate for poor documentation.