Helpdesk RAG
An IT helpdesk assistant that answers staff questions from the company's own knowledge base, with a source for every claim and a clean "I don't know" for everything else.
Results
- Answerable questions answered correctly
- 6/6
- Citations verified against the source
- 18/18
- Out-of-scope questions declined
- 5/5
- Correct questions wrongly declined
- 0
cat problem.md
The problem
Helpdesk teams answer the same questions every day, and the answers are already written down somewhere in the knowledge base. Staff can't find them, so they open a ticket. An assistant can answer instantly, but only if it never invents an answer.
What we built
A web app that answers questions over an IT helpdesk knowledge base, streaming the answer as it is written and listing the sources underneath, each with the exact quote it relies on.
It has three layers of protection against made-up answers. First, a relevance check stops questions the documents don’t cover before the AI is called. Second, the AI is told to answer only from the documents. Third, every claim carries a citation copied from the source, so anyone can check it.
Built to run in production
It runs as a proper service, not a notebook: access keys, rate limits, input checks, and a database that can be swapped without changing the answers. It is packaged with Docker and deployed to a public address.
cat project.json
- industry:
- IT support
- built:
- September 2026
- stack:
- PythonFastAPIClaudeVoyage embeddingsBM25Postgres + pgvectorDockerRender