Back to Projects
🌟 Featured Case Study 🎓 Licensed for Continued Use

Agentic Multimodal RAG-Enhanced Dialogue System for University Sports Administration

Master's Thesis · Institute of Information Management, National Taipei University
Cheng-Yun Wu · Advisors: Min-Yuh Day, Yen-Chun Lin · 2026
Live System Full Text Slides Demo Video
Problem Architecture Tech Stack Trade-offs My Role Challenges Contributions Evaluation Demo

Problem Background

System Architecture

Agentic Multimodal RAG system architecture
Agentic Multimodal RAG core: multimodal input, agentic reflection loop, domain tool routing, hybrid retrieval & reranking, and the indexed knowledge base

Tech Stack

Tech stack and deployment architecture
Actual deployed system: 4-layer architecture from frontend to infrastructure

LLM & Agent

OpenAI Responses API GPT-5.4-mini GPT-4o-mini Tool Calling

Retrieval

FAISS BM25 HyDE RRF LLM Reranking text-embedding-3-small

Multimodal

Whisper-1 (STT) tts-1 (TTS) GPT-4o Vision

Frontend & Infra

Gradio Cloudflare Tunnel Zero Trust

Engineering Trade-offs

Cost
    Speed
      Accuracy
        Stability

          My Role

          This project was completed independently by me, covering research planning, system development, and empirical evaluation. Key contributions include:

          Challenges & Solutions

          1 Administrative data is scattered and inconsistently formatted

          Problem

          Solution

            2 A single RAG pipeline struggles with complex administrative questions

            Problem

            Solution

              3 A single automated score doesn't fully reflect system quality

              Problem

              Solution

                System Contributions

                Adopted for Continued Use

                01 Lowering the cost of finding information and repeated back-and-forth

                02 Making complex administrative questions faster to answer

                03 A more natural, multimodal front door to administrative services

                User Feedback & Evaluation

                User Interviews

                Well-suited for standardized queries

                  Positive acceptance, trust built on evidence

                    Effectiveness tracks knowledge base completeness

                      Post-Use Survey
                      4.17/5.00
                      N = 39 respondents · overall positive rating
                      Information Quality
                      4.38
                      Communication Intelligence
                      4.23
                      Trust & Acceptance
                      4.22
                      Task Effectiveness
                      4.21
                      System Usability
                      3.81

                      Top-Rated Items
                      Response Appropriateness — 4.56 Naturalness — 4.54 Faithfulness — 4.46
                      Focus for Improvement
                      Response Speed — 3.38 Intent Understanding — 3.77 Interaction Ease — 3.95

                      Demo