Team09-Healthcare QA and Information Retrieval System

Indian Institute of Science Summer 2025 DA225o Submission2 Authors

Keywords: Voice QA, Multilingual NLP, Sarvam-M, STT, TTS, RAG, Memory, Cache, LLM, Langchain

TL;DR: Voice-based search and Q&A app for healthcare, aimed at delivering accurate, culturally relevant health info in India with strong safety and ethical standards.

Abstract:

We propose to develop a Healthcare QA and Information Retrieval System that uses deep learning to deliver accurate, accessible, and culturally relevant medical information through both voice and text-based interactions. Designed for the Indian context, the system supports multiple regional languages, addressing challenges related to linguistic diversity and limited access to trustworthy health information.

The system will integrate several deep learning components. User queries - submitted either as spoken input or typed text - will be processed through a speech-to-text (STT) engine (for voice input) and then analyzed using Sarvam-M, a multilingual large language model trained on ten Indian languages. Sarvam-M will perform intent recognition and medical entity extraction as part of a custom Natural Language Understanding (NLU) pipeline.

To generate accurate responses, the system will employ a Retrieval-Augmented Generation (RAG) approach. It will retrieve relevant information from a curated set of healthcare documents and medical resources to supplement the model’s response generation process. The retrieved content, combined with the user query, will guide Sarvam-M in producing a medically informed and contextually appropriate answer. Responses will be delivered either as text or converted to speech using a text-to-speech (TTS) engine, depending on the input mode.

A robust safety and ethics layer will ensure that the system does not offer diagnoses or emergency instructions, issues clear disclaimers, and avoids collecting personally identifiable health data. These safeguards are critical for building trust and ensuring responsible AI deployment in the healthcare domain.

This project is proposed as part of DA 2250 - Deep Learning (Summer 2025 Term course) by Team 9. It leverages publicly available healthcare datasets. The goal is to provide accurate, accessible, and culturally relevant health information to users, particularly in the Indian context, while ensuring robust safety measures and ethical considerations.

Submission Number: 2