How is the assistant trained?

The assistant learns from your own sources: website and subpages, documents and PDFs, data sheets, a list of frequent questions and price lists. From this material the knowledge base is built, on which its answers rely.

You do not need to program anything for this. We prepare the content and agree with you what the assistant should answer and where it prefers to hand over to a human. We discuss the basis in the initial consultation.

No model is trained here in the sense of changed weights. Your text is split into passages of 800 characters with 120 characters of overlap, uploaded files more coarsely; each passage receives a weight by source type and a detected language. The keywords that make a passage findable come from the passage itself, no model invents them. The language model later only formulates from the passages that were found.

This has a practical consequence: the search is literal, not guessed by meaning. A question that shares no word with your texts often finds no matching section, and the assistant is instructed to say so instead of formulating. Such cases are recorded as a knowledge gap and, from the Pro package, appear in the evaluation together with the question that triggered them. Which content is worthwhile for this and in what form is described in the article building a knowledge base.