Support Assistant

An assistant that answers from your records, or escalates

The same forty questions arrive every week, and the answer to nearly all of them already exists somewhere: in a help centre, in a wiki, and in four years of resolved tickets that nobody has time to read twice.

A model on its own knows none of that. Retrieval is most of the build.

So the system is a retrieval pipeline first and a model second. Tickets, documents and product records are chunked, embedded and indexed in Postgres, the question is answered only from what that index returns, and every answer carries the passages it was built from so anyone can check it.

What makes it safe to put in front of a customer is what happens at the edges:

• a confidence threshold that is tuned, not guessed

• a handover that carries the whole conversation

• citations on every answer, with no exceptions

• a refusal path for anything outside the index

Escalation is a feature rather than a failure. When retrieval returns nothing close enough, the assistant says so and passes the case to a person with the question, the sources it did find and its own uncertainty attached.

Grounding is only as good as the thing it is grounded in, so the index is rebuilt on a schedule and withdrawn documents are retired rather than left to compete with the current ones. An answer quoting a policy dropped last year is worse than no answer.

Every conversation is logged with the passages it used and the decision it made, which turns it into a reading of where the documentation is thin. The questions it escalates most are the pages worth writing next.

The target is not an assistant that answers everything. It is one that answers the routine half from your own records and knows, reliably, when it is out of its depth.