MedPolicy-RAG: A Multi-Domain AI Chatbot for Healthcare Compliance Interpretation

Abstract

This paper presents a domain-aware multilingual Retrieval-Augmented Generation (RAG) framework designed for regulatory document interpretation across five policy domains, including healthcare delivery, research funding, digital health data, food safety, and administrative directives. The proposed system supports both English and Devanagari Hindi documents and integrates optical character recognition (OCR) with dynamic language detection using the Google Vision API to process scanned and poorly formatted documents. Cross-lingual transformer embeddings enable semantic retrieval across languages, while a domain-aware routing mechanism restricts retrieval to relevant policy corpora, thereby reducing retrieval noise. A hybrid retrieval strategy combining dense vector search with cross-encoder and decoder-based re-ranking is employed to improve relevance, mitigate hallucination, and strengthen evidence grounding. The framework generates citation-aware responses in the user’s preferred language using structured prompting. Experimental evaluation on a gold-standard question-answer dataset demonstrates improved retrieval accuracy and answer fidelity compared with monolingual retrieval and conventional RAG baselines, establishing the framework as an effective multi-lingual regulatory assistant for the Indian healthcare and public policy environments.

Ähnliche Arbeiten