AI-POWERED HEALTH INSURANCE NAVIGATOR USING RETRIEVAL-AUGMENTED GENERATION (RAG)

Abstract

This work presents an AI-powered health insurance navigation system that applies retrieval-augmented generation (RAG) to produce personalized, evidence-grounded plan recommendations. Health plan selection is notoriously complex due to variability in deductibles, co-payments, co-insurance, and out-of-pocket limits, often overwhelming consumers and contributing to suboptimal coverage choices. Our system addresses this challenge through a hybrid architecture that combines structured retrieval over an indexed SQLite database with neural retrieval from a persistent Chroma vector store, enabling both fast filtering and context-rich analysis. A multi-step user profile (demographic, health, and financial data) is converted to targeted questions that drive retrieval; recommendations are generated via a constrained prompt that explicitly grounds outputs in retrieved plan documents. The backend (FastAPI) implements robust validation, monitoring, and REST endpoints, while a React frontend delivers an intuitive multi-step form and side-by-side comparisons with cost breakdowns and visualizations. Performance optimizations, smaller text chunks, MMR retrieval, caching, and indexing yield sub-second response times for 50+ plans. The system supports institutional integration via /compare, /query, and plan catalogue endpoints, making it suitable for hospitals, clinics, and brokers. Results demonstrate accurate, explainable recommendations and actionable cost projections, with clear disclaimers. This architecture shows that RAG can substantially improve transparency and personalization in health insurance selection while remaining more cost-effective and maintainable than fine-tuning approaches.

Key words 

Retrieval-augmented generation (RAG), health insurance, AI, personalized coverage, plan recommendation