Testing and quality
Check your assistant before you publish and keep an eye on answer quality once it's live.
Test before you publish
An eval is a set of questions you prepare with the answers you'd expect. Before you publish a change, you run the eval and comxbot checks whether the assistant covers the right topics, avoids anything you've banned and uses the right tools.
If the score drops compared with the last version, you know not to publish yet. It's a safety net, so a tweak that helps one answer doesn't quietly break another.
Keeping knowledge healthy
Every knowledge source gets a health score based on how fresh it is, how often it's actually used in answers and whether it had any problems being read. A source that goes stale or starts causing weak answers is flagged and relied on less.
Check the source health centre now and then. If something's gone red, refresh it or replace it. Good answers depend on good, current information.
Closing the gaps
When the assistant can't confidently answer something, that question goes into your unanswered queue. It's the single most useful list you've got, because it shows exactly what your customers or students want that you haven't covered yet.
Click a question and the AI drafts an answer from your existing knowledge. You check it and approve, and the new answer becomes a knowledge source itself. So the next person who asks gets it right, and your assistant gets better every week.