Loading
Not retraining a model. In practice it means selecting the documents and records the assistant may draw on, and setting the boundaries of what it may address.
So the work is curation and configuration rather than machine learning, and the main input is content you already have — or should have.
An assistant answering from an out-of-date help centre will quote it confidently. Content quality is the ceiling on answer quality, and no configuration raises it.
Start by fixing the ten articles covering the most common questions. That is more valuable than any amount of tuning, and it improves the help centre for humans too.
No. You curate sources and set boundaries; the model supplies the language.
Fixing the source content. It is the ceiling on answer quality.
From questions it could not answer. That is a backlog your customers wrote.
Half an hour with your own data usually saves reading three of these. The guides will still be here afterwards.