1 of 12

A ChatGPT-Powered Smart Chatbot for Online Library Services

Meng Qu

Aayog Koirala

Nhut (Steve) Do

LIB.MIAMIOH.EDU

2 of 12

History and Need for a Smart Chatbot

Equipment Reservations
Policy Clarifications
Rapid Information Retrieval
Citation Assistance
Subject Guides
Other Questions

SpringShare LibChat

Online reference services have been introduced around late 1980s to early 1990s, offering users an opportunity to interact with human librarians in real-time.

Miami University Libraries employed Springshare LibChat since 2015. Since then, LibChat serves to an average of 3,000 chats annually.

With the blooming of Natural Language Processing technology, the web service librarian has done a brief text analysis, based on the past six years of chat transcripts.

After a set of subject extraction and text classification works on these historical chat transcripts, it allowed us to differentiate between quick fact questions vs. in-depth research questions, leaded to the creation of a statistical report.

This report shows that approximately 20% of all questions were quick fact questions. The most frequently asked questions, which are classified as quick-fact questions, are typically within the following range: facility and equipment reservations, policy clarifications, rapid information retrieval, citation assistance, navigation of subject guides, and other similar questions.

We believed that these questions could have been effectively handled by an auto-answering chatbot if properly embedded into the existing Integrated Library Systems ILS.

3 of 12

A History of Developing a Smart Chatbot at MU Libraries

OpenAI�GPT-4

Google�DialogFlow ES

4 of 12

Embrace GPT-4 LLM into a chatbot system

GPT-4 demonstrated an enhanced understanding of human languages and the ability to generate human-like text. With its introduction in Spring 2023, the Smart Chatbot's development team saw the potential for incorporating GPT-4 into user input and question classification. As a result, the accuracy of assigning the corresponding virtual agent has increased, compared to the early version of the Smart Chatbot.

Incorporating the power of GPT-4 and related innovative technology, this intelligent chatbot embodies a pioneering solution tailored precisely to the distinctive demands of academic libraries. Its application aimed to demonstrate the potential of sophisticated AI capabilities in understanding user questions as well as maintaining continuous conversations. Proper programming structure in Nodejs allowed for user inputs to be accurately classified and directed to the correct virtual agent. In each turn of a dialog, the browser-side caching model saves the context. This enables the user to retain the conversation as if talking to a memorable human. Moreover, a comprehensive error-handling mechanism was put in place. With its power, this system is transforming the way online reference services are rendered, and elevating the overall user experience. We foresee an increase in chat numbers with the introduction of the Smart Chatbot option.

With the goal of leveraging cutting-edge technological advancements to enhance user engagement and satisfaction through efficient task handling and personalized interactions, we named this intelligent system the ChatGPT-powered Smart Chatbot.

5 of 12

Architecture and Functionality

Diving into the architecture of our chatbot, it's a web-based application, serving as a 24/7 available automated response system.

The chatbot leverages the power of ChatGPT's Application Programming Interface (well-known as API), processing these text-based exchanges and assigning them to task-specified virtual agents. Each of these agents is uniquely engineered to fulfill distinct tasks within predefined service domains. I’ll give you one example of our virtual agents, which is EBSCO searching agent. Armed with the capacity to handle diverse searching tasks, this agent interprets and processes various input modalities, such as keywords, authors, titles, subjects, and ISBN numbers. We’ll see a comprehensive demonstration very soon.

These virtual agents are not only designed to fulfill tasks but are also equipped with bespoke error-handling mechanisms. For instance, when a patron seeks to reserve a study room and inputs their preferred date and time. if the requested time slot be unavailable, the Smart Chat would deny their request and guiding the user to reconsider their preferred scheduling, thus ensuring a seamless user experience.

6 of 12

Key Features of the Smart Chatbot

LIB.MIAMIOH.EDU

Modular Design

03

Easy to implement additional features through API calls

Start the Conversation with User’s Choice

02

Start the conversation with:

The Smart Chatbot
A Human Librarian

Multi-Turn Dialog

01

Guides users to provide �missing information

A significant feature of our Chatbot's functionality lies in its capacity for multi-turn conversations. This is best exemplified in the study room reservation process, wherein the chatbot guides users to provide necessary details like date, time, patron's name, and their school email address, seeking missing information in an intuitive manner until the reservation is successfully completed.

If the Chatbot cannot provide an answer, the user will be navigated to start a live chat with a human librarian. Recognizing the critical role of human involvement, our system allows users to initiate direct conversations with a human librarian during working hours. For times when librarians are unavailable, an embedded form lets users submit queries to receive a response when the librarian is back on duty.

From a structural standpoint, our chatbot is designed with an unwavering focus on cost efficiency, thanks to the implementation of caching mechanisms that optimize the usage of ChatGPT resources. Moreover, the modular design structure allows hassle-free integration of additional features, aligning with the evolving needs of the users as well as from the developers.

I’ll pass it to my team member, Steve, for a technological explanation and a more detailed introduction about this product.

7 of 12

ReAct

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022). ReAct: Synergizing Reasoning and Acting in Language Models.

8 of 12

ReAct

9 of 12

Architecture

Diagram Link

10 of 12

Let’s watch a live demonstration.

LIB.MIAMIOH.EDU

The public view of testing study room reservation:

https://muohio.libcal.com/reserve/spaces/chatbot-testing

Demo examples:

Talk to a human librarian on-duty
Book a study room
EBSCO search by title, author, keyword
Check King Library building hour
And more..

11 of 12

Challenges and Effective Solutions

LIB.MIAMIOH.EDU

Challenges	Problems	Solutions
Misuse	Unforeseen responses	Extra response screening
Non-consistent responses	Poor user experiences	Adopt advanced LLMs
Resource scarcity	Hindered development	Acquire more supports
Service efficiency	High online queries	Implement caching system

A conversational tool like this has several benefits, but it doesn't come without its challenges which span across different areas. The first one is the potential for misuse of this tool, leading to unforeseen responses. This is primarily because AI lacks the ability to distinguish right from wrong, potentially producing unwanted textual outcomes. In order to mitigate this, our developers have included additional protection to screen out such responses.

Another issue is the inconsistent reliability of AI-generated responses. In the early stages of our development, AI responses deviated significantly from the intended track. However, things have significantly improved since the adoption of GPT-3.5-Turbo, a model provided by OpenAI. Through fine-tuning, we have managed to make the responses more accurate and direct the virtual agent to perform the given tasks more precisely.

The third challenge is the inadequacy of resources for development, ranging from technical support to financial assistance to server-side help. We are continuously striving to acquire more resources to ensure that the development process progresses smoothly.

The final challenge lies in enhancing the efficiency and efficacy of academic library services, given the rise in online queries. The cost of implementing the GPT-4 Model during development was substantial, and we anticipate a significant increase once the Smart Chatbot goes live. Increased use of the GPT-4 API is likely, driven primarily by curiosity and easy access to answers. To tackle this, our developers are devising a solution to store some question-answer pairs within a caching system before they are directed to appropriate virtual agents. This will not only reduce the cost of deploying the GPT-4 model but also expedite the process of providing answers.

12 of 12

Future Project Development

Implement to other platforms

Higher emotional intelligence capabilities

Provide multilingual services

Incorporate to �Multiple Platforms

Higher Emotional Intelligence

Provide Multilingual Services

As we look into the future, there are numerous potential projects involving smart chatbots and library services. One such project could be the development of chatbots with higher emotional intelligence capabilities. Currently, most chatbots can only respond based on pre-defined phrases. Enhancing their emotional intelligence would involve programming the chatbots to detect and interpret the emotional state of the user based on their queries and respond appropriately. This could greatly improve user satisfaction and the overall user experience.

Another potential improvement could encompass incorporating intelligent chatbot technology into other resources, such as enhancing information literacy, addressing complex research inquiries, and integrating with broader teaching and learning platforms, like Canvas or Blackboard. The chatbot could lead users across these multi domain tools, clarifying their functionality and helping users access the information they require.

Additionally, future development could also entail the creation of multilingual smart chatbots. Such chatbots could serve to break language barriers in library services, providing assistance to users in multiple languages and thus enhancing inclusivity.