A Multi-Agent RAG System for Personalized Wellness Guidance

Abstract

This project introduces the Holistic Health AI Assistant, a modular, AI-powered wellness advisor that leverages Retrieval-Augmented Generation (RAG) and GPT-3.5 to deliver personalized, evidence-based health guidance. The system retrieves real-time information from PubMed and synthesizes it using specialized GPT agents, enabling it to provide credible wellness recommendations tailored to user input. By combining the flexibility of multi-agent architecture with scientific grounding, the assistant bridges the gap between reliable medical information and user-friendly interaction.

Introduction

Millions of users turn to digital tools for health and wellness guidance, yet most systems lack personalization, evidence-based insights, or both. Conventional chatbots often generate superficial advice without source validation. To address these gaps, we developed a holistic AI assistant designed to respond with medically grounded, personalized recommendations using real-time scientific data and natural language generation.

System Overview

The Holistic Health AI Assistant is powered by a multi-agent architecture, each agent performing a dedicated function within a shared RAG pipeline. The system’s design allows flexibility, traceability, and scalability while ensuring that the advice provided is not only natural but also backed by up-to-date biomedical literature.

Core Agents

Retriever Agent: Queries the PubMed API using semantic and keyword matching to gather relevant literature.

Planner Agent: Analyzes user intent, decomposes complex queries, and routes sub-tasks to appropriate agents.

Responder Agent: Synthesizes findings into coherent, user-friendly responses using GPT-3.5.

Interface Agent: Manages user interaction through a clean, web-based interface.

Technology Stack

LLM Backbone: OpenAI GPT-3.5 (gpt-3.5-turbo)

Knowledge Source: PubMed API (real-time biomedical data)

Framework: ReadyTensor Agent Framework

Frontend: Web-based chat interface (React + Flask)

Use Case Examples
General Health Queries: “What are the health risks of low vitamin D levels?”

Lifestyle Guidance: “How can I improve my sleep quality naturally?”

Fitness & Nutrition: “What’s an effective diet plan for prediabetes?”

Mental Wellness: “How to manage stress without medication?”

For each of these, the system retrieves relevant studies and composes a personalized response that reflects both the user’s needs and scientific consensus.

Results and Insights

Evaluation Metric Outcome
PubMed Retrieval Accuracy ~88% relevance to user queries
Response Fluency 4.8/5 (user feedback from 10 testers)
Synthesis Accuracy High when grounded in relevant documents
Response Time 5–7 seconds per query

Pilot users praised the system for its clarity, medical accuracy, and inclusion of cited sources. The assistant was especially effective in queries involving lifestyle improvements and chronic condition management.

Limitations

While promising, the system has a few constraints:

Not Clinically Validated: It is not approved for diagnostic or emergency use.

Dependence on Retrieval Quality: The accuracy of outputs depends on the relevance of PubMed query results.

LLM Hallucination Risk: Although grounded in data, GPT-3.5 may still generate unsupported statements in edge cases.

Future Work

To improve the assistant further, we aim to:

Integrate medical-specialized LLMs (e.g., Med-PaLM, GPT-4 with tuning)

Support wearable data integration for real-time biometric feedback

Conduct clinical validation studies

Offer voice-based and multilingual support

Develop a mobile app version with offline capabilities

Conclusion

The Holistic Health AI Assistant demonstrates how a modular RAG system can transform the way individuals access health information. By grounding GPT-3.5 responses in live biomedical literature and maintaining a flexible, agent-driven design, the system delivers medically relevant, conversational guidance. With further development, it has the potential to become a trustworthy companion for preventive health and lifestyle improvement.

Authors
Sampanna Timlasina
Gaurav Koirala
Sushant Subedi