The assistant and the knowledge base — why it prefers silence to guessing
The assistant answers “how do I do this?” questions from the documentation and provides sources. When the match is too weak, it deliberately does not answer instead of inventing a procedure.
Questions fall into two groups: about data (“how many do we have in stock?”) and about how to use the system (“how do I change the status?”). The assistant answers the latter from the knowledge base — the same documentation you are reading now.
Sources beneath the answer
Clickable links to the articles used to produce the answer appear beneath it. Clicking one opens the full text in a window without leaving the conversation.
It is worth using these links: the assistant's answer is a summary, while the article includes edge cases and warnings.
The threshold below which the assistant does not answer
This is the most important point in this article:
If no article matches closely enough, the assistant receives none and is explicitly instructed not to guess the steps.
There is a practical reason. In a warehouse system, almost every article mentions “order” somewhere. Without a threshold, a question about an undocumented subject would pull in a handful of unrelated texts, and the model would turn them into a confident, entirely fabricated procedure. That is worse than silence because it looks credible.
So “I couldn't find that in the documentation” is a valid answer, not a failure.
Unanswered questions are recorded
A question that finds no article is recorded along with the number of times it has been asked. This creates a list of gaps in the documentation: the more often someone asks something without getting an answer, the higher it moves in the writing queue.
By asking a question that gets no answer, you are doing something useful. There is no need to repeat it over and over — once is enough.
Suggested questions
When the assistant is not sure what you mean, it suggests ready-made questions instead of showing an error. It draws them in this order:
- from articles that partially match your words — “did you mean this?”
- from articles describing the screen you are on,
- from any published articles.
The first non-empty source wins. Two relevant suggestions are better than four, two of which are random.
How to ask a question that finds a match
- Write a complete sentence. Search understands questions, not just keywords.
- Write document codes in uppercase —
PZ,WZ,MM. Capitalization distinguishes a code from an ordinary preposition. - Do not write everything in uppercase — code recognition stops working.
- Inflected forms are fine — “status”, “statuses” and “status's” all lead to the same result.
Want to see this with your orders? We’ll show you NOXTI with your sales channels and warehouse.
Book a demo