FAQ Assistant — answers with sources and human handoff

Context

Customer-facing assistants lose trust quickly when they guess. Business owners worry about wrong answers most.

The problem

How do you show a customer assistant that behaves responsibly: grounded answers, visible sources and a graceful “I’ll ask the team”?

Objective

Build a small demo of the behaviour we design into production assistants.

Role

Interaction design and build by Orbit Hive.

Constraints

Uses a fictional café’s information. Rules-based matching stands in for a language model in this demo.

Approach

• A small approved knowledge base with named sources
• Every answer shows where it came from
• Sensitive topics (allergies, complaints, refunds) always go to a person
• Quick-question chips for common questions

Technical notes

Front-end prototype. In production: retrieval over approved documents, a language model, guardrails, logging, evaluation sets and a handoff channel.

Outcome

No measured outcome to report. This is a internal prototype, not a client deployment. We only publish results we have actually measured.

What we learned

Visible sources and an explicit handoff do more for trust than a clever personality.

Next-step ideas

Connect to a language model with retrieval, and add an evaluation set of real questions.