# A knowledge assistant support agents actually trust

URL: https://mechaneforge.tech/case-studies/support-knowledge-assistant/
Sector: SaaS support team (representative example engagement, anonymised)

> Agents spent time hunting through help articles, past tickets and internal notes. We built an assistant that answers from those sources and cites them, with a feedback loop to keep it accurate.

## Challenge
- Answers lived in wikis, old tickets and individual agents’ heads, so replies varied by who picked up the ticket.
- New agents took weeks to become productive.
- Earlier AI experiments produced confident but unsourced answers the team could not rely on.

## Approach
- **Index the knowledge**: Help centre, past resolved tickets and internal runbooks were cleaned, chunked and indexed, with permissions respected.
- **Answer with citations**: Retrieval-augmented answers that quote and link their sources, and say “not found” rather than guess.
- **Evaluate before launch**: A test set of real tickets with agent-approved answers, scored on every change to prompts and retrieval.
- **Close the loop**: Agent thumbs-up and thumbs-down, and corrections, feed back into the test set and knowledge base.

## What we measure
- Answer accuracy on the test set: Agreed threshold before rollout
- Time to first draft reply: Seconds instead of minutes
- New-agent ramp-up: Measurably shorter than baseline

Built with: Retrieval (vector + keyword), LLM with citations, Helpdesk API, Evaluation harness