Project Overview

After the initial development of the HKUST Library AI Chatbot, I took over the project and continued improving it over the following months. As a Year 1 student, this was my first time working on a system that combines AI, APIs, and a real user interface, so a big part of the experience was learning by doing.

My main focus was to make the chatbot:

• More useful for real users
• More consistent in its responses
• Easier to expand in the future

During this process, the system gradually developed into a multi-agent chatbot, combining: A RAG (Retrieval-Augmented Generation) agent for answering questions, a Search agent using HKUST’s PowerSearch for retrieving real library resources and a Coordinator agent that decides which agent to use. In addition, an architecture for easy expansion of future agents.

My Approach

Since I was still new to many of the technologies, my approach was simple:

• Break problems into small parts
• Test different solutions
• Improve step by step

Instead of redesigning everything at once, I focused on incremental improvements to both the backend logic and the frontend experience.

What I Worked On

1. Improving the Search Agent (HKUST PowerSearch Integration)

One of my main tasks was refining the search agent, which connects to HKUST PowerSearch system. I worked on fixing formatting issues such as:

• Adding auto-date filtering parameters so results can be more relevant
• Improving how the system handles follow-up searches
• Logging search results for tracking and debugging

These changes made the search results:

• More accurate
• Easier to read
• More consistent across different queries

2. Moving to a Multi-Agent System

Previously, the chatbot behaved more like a single system, also agents were used, the mechnism was only one-agent-at-a-time. During my work, it started to function more like a multi-agent setup, where different parts handle different tasks.

• The RAG agent explains information
• The Search agent retrieves real resources

This allows the chatbot to:

• Provide explanations and actual search results
• Be both informative and practical

3. Frontend and Output Redesign

I also redesigned parts of the frontend to improve usability. Key changes include:

• Changing backend output from HTML to JSON, making it easier to manage
• Designing a new response display using JavaScript and CSS
• Ensuring consistent formatting across different types of responses
• Master the end-to-end RAG AI system architecture, focusing on backend logic and system prompting
• Optimize communication efficiency between the frontend CMS and backend APIs
• Integrate third-party vendor APIs with AI systems to expand data ingestion capabilities

I also added a button that allows users to open search results directly in their browser with pre-set filters.

This connects the chatbot more directly with the actual library system.

4. Feedback System

To support real users and future improvements, I worked on feedback features:

• In addition to the thumbs up / thumbs down feedback, I redesigned the comment option
• Created an option for users to enter their email for direct contact
• Set up automatic feedback emails to staff

What I Learned

From this project, I gained experience in:

• Multi-agent system design (basic level)
• Working with APIs (HKUST PowerSearch integration)
• RAG-based chatbot systems
• Frontend–backend integration (JSON, JS, CSS)
• Debugging and improving real-world systems
• Work on the application development on Linux environment

More importantly, I learned how to:

• Approach unfamiliar problems
• Test and refine ideas
• Build something step by step