Evaluated project

QuoteGuard

Guardrailed, grounded RAG chatbot over an insurance policy

RAG Guardrails BM25 + rerank recall@5 0.80 Legal-aware
What's under the hood
01Retrieval-augmented search over a 100-page policy
02A language model that answers only from the document
03Guardrails that keep every reply in scope
04Proven on a gold-standard question set
📄 Allianz Business PDS · 100 pp
Live
Guardrail Trace
Measured on a 66-question evidence-validated gold set · bootstrap 95% CIs
0.80
recall@5 — right passage in top 5
1.00
Adversarial safe-rate across 40 attacks
0%
Legitimate questions wrongly refused
7
Attack categories — all at 1.00 after one tighten pass
Want the full story?
See how it actually works
Retrieval bake-off · guardrail pipeline · adversarial eval · live demo
Try the live demo
williamcatt.dev/projects/quoteguard
← Back to home
0:00 / 0:00