Fix the knowledge base first
Wrong answers are usually stale, duplicated or contradictory knowledge rather than a weak model. We begin with an audit: which documents are valid, who updates them, which version wins in a conflict.
Then fix ticket structure
If tickets lack structured category, order number and resolution fields, the model cannot judge context and falls back to generic wording. Adding fields beats swapping models.
The model comes last
With trustworthy knowledge and structured context, the model layer is simple: retrieve, generate, escalate. We attach a citation to every answer so agents can verify it.
How to judge it
Not "does it sound human", but "did escalation drop, did first-contact resolution rise, were there zero wrong commitments".