Shipping LLM assistants: Lessons from Quantia
Three lessons shipping LLM assistants to production: measure before you build, guardrail before you scale, and treat human handoff as a feature.
LLM assistants are easy to demo and hard to run. Here are three lessons from shipping Quantia’s support assistant to production.
1. Measure before you build
Before writing the RAG pipeline, define an eval set: question–answer pairs and “good answer” criteria. Without it, you can’t tell if you’re improving or regressing.
2. Guardrail before you scale
Two guardrails are non-negotiable: cite sources (every answer links to source docs) and no fabrication (if nothing is found, say “I’m not sure”).
- Cite a source on every response
- Decline gracefully when context is missing
- Log a confidence score per answer
3. Human handoff is a feature
On low confidence, the assistant should hand off to a person — smoothly, with context. This isn’t a failure, it’s safe design.
The best AI systems aren’t the ones that never fail — they’re the ones that know when to defer to a human.
Jordan builds RAG systems and AI assistants for retail and support. Focused on reliability and quality measurement.
