How our insurance reference tenant serves 1,200 FAQs at sub-second latency
A walk-through of the deployment we use ourselves to validate every release.
1,200+
indexed FAQs
across finance, insurance, healthcare — retrieved via hybrid RAG in <350 ms median
Challenge
Most insurance sites ask a visitor to compare 4–6 plans across 15 variables: sum insured, room rent, co-pay, sub-limits, waiting periods, OPD. The typical outcome: 3 tabs open, decision fatigue, and an unattended contact form.
The reference tenant needed to handle those comparison questions without sacrificing the compliance language the regulator requires.
Setup
- Uploaded policy PDFs via /knowledge/import-document — auto-chunked into ~400-word KB entries and indexed into Qdrant.
- Configured the Insurance tone preset plus a custom snippet: "Use Indian English. Always quote premiums in ₹. Never promise coverage — say 'subject to policy terms'."
- Enabled the orchestrator's rules engine to hand off to a human advisor whenever the retrieval confidence fell below 0.55.
- Turned on the embedded widget on the Hanvitt marketing site plus a staged WhatsApp Business flow for inbound leads.
Result
<1s
First answer in
84%
KB hit-rate on real visitor questions
32% of started convos
Lead captures from chat sessions
100%
Advisor handoffs (vs. hallucinated answers)
Every Hanvitt release ships only after the reference tenant passes a 50-query regression. It's how we catch regressions before you do.