• Virtual and Augmented Acoustic Reality
• mmWave-based Continuous Monitoring
• Edge video analytics
• Multi-modal human motion synthesizing and understanding
• RAG for Older Adult Mobility and Health Information

Retrieval-Augmented Generation (RAG) for Older Adult Mobility and Health Information

Author: Jia Yang

Supervisor: Dr. Rong Zheng

Institution: McMaster University

---

Table of Contents

1. Project Description and Approach
   • Motivation
   • System Overview
   • Key Design Features
   • Architecture Components
2. Results
   • Retrieval Performance (BEIR Benchmark)
   • Local vs Web Retrieval
   • Qualitative Evaluation
3. Source Code
4. References

---

1. Project Description and Approach

Motivation

Older adults increasingly rely on online resources for health information related to mobility, fall prevention, balance training, and rehabilitation. However:

• Search engines often return fragmented or unreliable results.
• Large Language Models (LLMs) may generate fluent but unverified responses.
• Health misinformation can negatively impact safety and decision-making.

This project develops a reliable, citation-grounded Retrieval-Augmented Generation (RAG) system designed specifically for older-adult mobility and health information.

---

System Overview

The system integrates:

• BM25 lexical retrieval
• Cross-encoder semantic reranking
• Relevance-gating mechanism (CE threshold τ = 0.6)
• Google Search API web fallback
• LLM-based answer generation with explicit citations

Pipeline Summary

1. User submits a natural-language query.
2. BM25 retrieves top-k passages from a curated MongoDB corpus.
3. A cross-encoder reranks candidates.
4. A relevance gate determines:
   • High CE score → use local database evidence
   • Low CE score → trigger web fallback
5. The LLM generates a plain-language answer with citations.

---

Key Design Features

• Curated corpus of trusted health documents and passages
• Explicit source citations in every answer
• Plain-language output tailored for older adults
• Safety-aware prompts with medical disclaimers
• Modular Python implementation (BM25, Sentence-Transformers, Flask, Streamlit)

---

Architecture Components

• Crawler Layer (PDF + Web)
• Retrieval Layer (BM25)
• Reranking Layer (Cross-encoder)
• Decision Layer (Relevance Gate)
• Generation Layer (LLM synthesis)
• User Interface Layer (CLI, Flask API, Streamlit UI)

---

2. Results

Retrieval Performance (BEIR Benchmark)

Semantic retrieval outperformed BM25:

Dataset	Method	nDCG@10	MAP@10	Recall@100
SciFact	BM25	0.5178	0.4760	0.7896
SciFact	Semantic	0.6451	0.5959	0.9250
NFCorpus	BM25	0.3804	0.3415	0.6827
NFCorpus	Semantic	0.4686	0.4317	0.8541

A hybrid BM25 + cross-encoder balances computational efficiency and semantic precision.

---

Local vs Web Retrieval

Query Type	Local CE	Web CE
In-domain	≈ 0.997	≈ 0.892
Out-of-domain	≈ 0.000018	≈ 0.958

These results demonstrate strong domain sensitivity and reliable fallback behavior for out-of-scope queries.

---

Qualitative Evaluation

Evaluation using real-world older-adult mobility queries demonstrates:

• Evidence-grounded responses
• High in-domain relevance
• Appropriate fallback behavior
• Clear citation display
• Inclusion of safety disclaimers

---

3. Source Code

The full implementation, retrieval pipeline, and deployment instructions are maintained in a private GitHub repository:

GitHub Repository: View GitHub Repository

---

4. References

• Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
  Paper link
• Izacard, G., & Grave, E. (2021). Leveraging Passage Retrieval with Generative Models for Open-Domain QA. EACL 2021.
  Paper link
• Nogueira, R., & Cho, K. (2019). Passage Re-ranking with BERT. arXiv:1901.04085.
  Paper link
• Thakur, N., et al. (2021). BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of IR Models. NeurIPS 2021.
  Paper link
• World Health Organization. (2022). Healthy Ageing.
  WHO page

Back to top ↑