Journal of Web Librarianship

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/wjwl20

Aisha: A Custom AI Library Chatbot Using the

ChatGPT API

Yrjo Lappalainen & Nikesh Narayanan

To cite this article: Yrjo Lappalainen & Nikesh Narayanan (2023): Aisha: A Custom
AI Library Chatbot Using the ChatGPT API, Journal of Web Librarianship, DOI:
10.1080/19322909.2023.2221477

To link to this article: https://doi.org/10.1080/19322909.2023.2221477

Published online: 14 Jun 2023.

Submit your article to this journal

Article views: 712

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=wjwl20
Journal of Web Librarianship
https://doi.org/10.1080/19322909.2023.2221477

Brief Report

Aisha: A Custom AI Library Chatbot Using the

ChatGPT API
Yrjo Lappalainen and Nikesh Narayanan
Library and Learning Commons, Zayed University, Dubai, United Arab Emirates

ABSTRACT KEYWORDS
This article focuses on the development of a custom chatbot for chatbots; ChatGPT;
Zayed University Library (United Arab Emirates) using Python OpenAI; GPT-3; GPT-3.5;
and the ChatGPT API. The chatbot, named Aisha, was designed generative pre-trained
to provide quick and efficient reference and support services to transformer; academic
libraries; artificial
students and faculty outside the library’s regular operating
intelligence; AI
hours. The article also discusses the benefits of chatbots in aca-
demic libraries, and reviews the early literature on ChatGPT's
applicability in this field. The article describes the development
process, perceived capabilities and limitations of the bot, and
plans for further development. This project represents the first
fully reported attempt to explore the potential of a ChatGPT-
based bot in academic libraries, and provides insights into the
future of AI-based chatbot technology in this context.

Introduction
Zayed University Library (United Arab Emirates) plays a crucial role in
providing access to various resources and information to support its stu-
dents and faculty’s research and academic needs. The library has been
providing an online chat service managed by Reference librarians to its
users during the working hours of the library. However, extending the
chat service when users require assistance outside of regular operating
hours is difficult. Chatbots offer a promising solution to address this issue.
A chatbot is a computer program simulating communication with human
users, typically using a message interface. Chatbots are designed to com-
prehend user inquiries and provide responses that resemble a human
conversation. They can be utilized for many things, such as work auto-
mation, information retrieval, and customer service. By leveraging chatbots,
libraries can provide quick and easy access to resources, answer research-re-
lated questions, and offer Reference help to students and faculty 24/7.
However, until recently, chatbots have been limited in their ability to
understand and respond to user queries accurately. Traditional chatbot

CONTACT Yrjo Lappalainen [email protected] Library and Learning Commons, Zayed University,
P.O. Box 19282, Dubai, United Arab Emirates
© 2023 Yrjo Lappalainen, Nikesh Narayanan
2 Y. LAPPALAINEN AND N. NARAYANAN

solutions have relied primarily on predetermined rules, such as pattern

matching and keyword-based question answering, while lacking any
advanced logical reasoning capabilities.
In November 2022, the AI landscape changed significantly when OpenAI
introduced ChatGPT, a new type of chatbot that is built upon the highly
advanced GPT-3.5 and GPT-4 (Generative Pre-trained Transformer) large
language models (LLM) (OpenAI, 2022). Although the GPT-3 API had
been in private beta testing since 2020, it was the introduction of ChatGPT
that launched a huge public interest in the technology. According to a UBS
analyst note, ChatGPT was estimated to have reached 100 million monthly
active users in January 2023, just two months after its launch (Hu, 2023).
Alongside the launch of ChatGPT, the OpenAI API (https://platform.openai.
com) was made available to all developers worldwide, leading to a surge
in new projects that utilize the technology. Following the launch of ChatGPT
and the OpenAI API, Zayed University Library started a project to build
a custom chatbot using Python and the OpenAI API. The bot was subse-
quently named Aisha, meaning “alive” or “she who lives” in Arabic.

Objectives and scope

This article describes the development process, perceived capabilities and

current limitations of the bot, and plans for further development. We also
present a brief history of chatbots and review the use of chatbots in the
context of academic libraries. Additionally, we review the early literature
regarding ChatGPT's applicability in this field. At the time of writing, no
custom chatbot utilizing the OpenAI API has been reported in the field
of academic libraries. Therefore, this project represents the first fully
reported attempt to explore its potential in this area.

Literature review
Alan Turing developed the Turing Test (originally called the “imitation
game”) in the 1950s to assess the intelligence of computer programs, and
Mauldin (1994) coined the term “chatbot” to characterize systems that can
simulate human interaction and attempt to pass the Turing test. Midway
through the 1960s, the MIT Artificial Intelligence Laboratory created
ELIZA, the first chatbot capable of locating keywords in a given input
sentence and matching those keywords against predefined rules to produce
appropriate responses (Weizenbaum, 1966). Following ELIZA, the devel-
opment of increasingly intelligent chatbots advanced, most notably with
the creation of PARRY, developed by Kenneth Colby, a psychiatrist, in the
early 1970s to simulate a paranoid patient’s conversational style for use
in therapy and research (Deshpande et  al., 2017). When users interact
 Journal of Web Librarianship 3

with keyword-based chatbots, the system identifies keywords in their

queries and matches them to pre-programmed responses. While key-
word-based chatbots can handle simple queries, they may struggle with
complex, nuanced, or context-dependent questions due to their limited
scope and less-flexible nature.
During the early 1980s, the creation of the Artificial Linguistic Internet
Computer Entity (ALICE) marked a significant milestone in the develop-
ment of Artificial Intelligence Markup Language (AIML) and became a
cornerstone of many chatbot platforms and services in sophisticated chatbot
projects (Wallace, 2009). With the development of new technologies, chat-
bot capabilities grew to include Artificial Intelligence (AI), which has
improved machine learning, data analytics, and natural language processing
(NLP) skills. The sophistication and effectiveness of chatbots have signifi-
cantly increased due to this evolution. AI chatbots can discern context,
semantics, and language nuances, enabling them to manage more complex
and varied queries and continuously learn and improve their performance
based on user interactions, resulting in a more human-like and engaging
conversational experience (Hussain et  al., 2019). Chatbot technology has
grown significantly over the years, and several AI-based models have been
developed to provide more natural and engaging user interactions. Notable
chatbots, such as JABBERWACKY (now known as Cleverbot, 1988), Watson
(2006), ALEXA (2015), Cortana (2015), and Tay (2016), have been devel-
oped to offer either text-to-speech or speech-to-speech interactions, uti-
lizing machine learning techniques and NLP. These chatbots have been
developed for various purposes, ranging from virtual assistants to gaming
and entertainment (Ashfaque, 2022).

AI chatbots in libraries

The evolution of AI-based chatbots in libraries has seen significant

advancements in the last decade, with a growing number of libraries
around the globe embracing AI technology to augment their services and
provide support to users. In 2010, Kornelia, the inaugural public library
chatbot, was introduced in Bern, Switzerland, marking a pioneering step
in the field (McNeal & Newyear, 2013). European libraries played a vital
role in the nascent stages of chatbot development. The Stella experiment,
an early chatbot, was implemented at Hamburg University, representing
Europe’s first academic library implementation (Allison, 2012). This inno-
vation was closely followed by the deployment of Chatbot Charlie at the
Delft University of Technology in the Netherlands (Ehrenpreis & DeLooper,
2022). In February 2011, the University of Nebraska-Lincoln Libraries in
the United States launched the Pixel project, an AI-based chatbot con-
structed using PHP code and an SQL server for its database to deliver
4 Y. LAPPALAINEN AND N. NARAYANAN

prompt answers to questions concerning library services and resources

(Allison, 2012; McNeal & Newyear, 2013). Mentor Public Library (MPL)
and Akron-Summit County Public Library (ASCPL) in the United States
were among the early adopters of chatbots, both operational by 2012
(Allison, 2012). In Australia, the University of Technology Sydney devel-
oped a chatbot prototype (Mckie & Narayan, 2019). In 2013, the University
of California, Irvine (UCI) initiated the development of its chatbot,
ANTswers, constructed on an open-source platform, envisioned as a point-
of-need reference tool that would complement existing online reference
services without necessitating live staffing (Kane, 2019). Recent chatbot
implementations encompass San Jose State University’s Kingbot and the
University of Oklahoma’s Bizzy, both introduced in 2020. Kingbot was
developed using Kommunicate, a proprietary software that leverages
Google’s Dialogflow tool (Rodriguez & Mune, 2021). The University of
Oklahoma’s Bizzy chatbot employs Ivy machine learning software to answer
routine questions (University of Oklahoma, 2020).

Conversational AI with large language and generative AI models

The last two years have marked a significant breakthrough in generative

AI, with the introduction of ChatGPT and many other LLM such as
LaMDA (https://blog.google/technology/ai/lamda), AlexaTM (Soltan et  al.,
2022), Chinchilla (Hoffmann et  al., 2022), PaLM (Chowdhery et  al., 2022),
PaLM 2 (https://ai.google/discover/palm2), Falcon (Technology Innovation
Institute, 2023), BloombergGPT (Bloomberg, 2023), PanGu-Sigma (Ren
et  al., 2023), GPT-NeoX (Black et  al., 2022), LLaMA (Meta AI, 2023),
Alpaca (Taori et  al., 2023), Cerebras-GPT (Dey et  al., 2023), GPT-J (https://
huggingface.co/EleutherAI/gpt-j-6b), Vicuna (https://vicuna.lmsys.org),
Koala (Geng et  al. 2023) and StableLM (https://github.com/Stability-AI/
StableLM). (For an up-to-date list of over 100 LLM, see https://lifearchitect.
ai/models-table). Simultaneously, there has been another major development
in the field of AI image generation, with ground-breaking large-scale text-
to-image models such as DALL-E 2 (https://openai.com/product/dall-e-2),
Midjourney (https://www.midjourney.com) and Stable Diffusion (https://
github.com/CompVis/stable-diffusion) released in the past two years.
LLM comprehend and produce natural language responses to various
queries using deep learning approaches like transformers. Generative mod-
els can produce completely original content that can be applied in various
ways. A generative chatbot, for instance, can be taught to generate news
articles, poems, and even movie or television screenplays. These models
are more accurate and efficient than traditional rule-based or retriev-
al-based chatbots, providing more personalized and engaging interactions
with users. Google Bard is an experimental conversational AI tool that
 Journal of Web Librarianship 5

originally used the Language Model for Dialogue Applications (LaMDA)

to generate responses to user inputs. In May 2023, Google announced that
Bard had moved from LaMDA to PaLM2, a more advanced language
model. Bard initially had a limited availability through a waitlist, but in
May 2023 it was made publicly available in 180 countries and territories
(Hsiao, 2023). By utilizing data from the internet, Bard can offer up-to-
date and high-quality responses to user queries (Pichai, 2023).
ChatGPT, developed by OpenAI, is a conversational AI model that has
gained widespread popularity for its ability to engage in natural language
conversations with humans. It is based on the GPT-3.5 language model
and has been trained on a large corpus of text data to understand the
nuances of language and generate contextually relevant responses in real
time. ChatGPT can answer a wide range of questions and handle complex
and context-dependent queries. One of the unique features of ChatGPT
is that it can answer follow-up questions, challenge incorrect premises,
and reject inappropriate requests (OpenAI, 2022). The model is based on
a transformer architecture and is trained using Reinforcement Learning
from Human Feedback (Gozalo-Brizuela & Garrido-Merchan, 2023).
Developers can also create custom GPT-based chatbots and virtual assis-
tants that can interact with users in a more natural and intuitive way,
providing them with personalized responses and assistance. A custom
chatbot developed using the ChatGPT API can be trained to understand
and respond to user queries related to the custom environment. In March
2023, OpenAI announced that it is rolling out a web browsing feature
and other ChatGPT plugins that can extend the language model’s func-
tionality by accessing external data sources and services (OpenAI, 2023b).
However, these plugins are still in an early testing phase and are only
available to ChatGPT Plus subscribers at the time of writing (OpenAI, 2023c).

ChatGPT in the context of academic libraries

Since its public launch in November 2022, ChatGPT has received significant
attention with numerous articles and preprints already published about its
potential impact on various fields. The impact of ChatGPT on academia and
education, particularly in regards to academic integrity, has been an area of
major interest (see e.g. Cotton et  al., 2023; King & ChatGPT, 2023; Lim et  al.,
2023). Some articles have also been published about the role of ChatGPT in
the context of academic libraries. Lund and Wang (2023) examined the potential
impact of ChatGPT on academia and libraries by interviewing ChatGPT itself.
Based on its responses, they identified that ChatGPT has the capability to
improve several library services such as search and discovery, reference and
information services, cataloging and metadata generation and content creation.
However, they also emphasized that the technology needs to be used
6 Y. LAPPALAINEN AND N. NARAYANAN

responsibly and that ethical considerations such as privacy issues and bias need
to be taken into account.
Cox and Tzoc (2023) discussed the potential implications of ChatGPT
for academic libraries from a wide perspective. They suggested that
ChatGPT could complement or even replace existing search methods. They
also commented that ChatGPT can be integrated into library discovery
tools, which may lead into an “arms race” between providers as they
contend to add this functionality into their products. The authors also
highlighted the role of ChatGPT in research, where it could be used for
brainstorming and finding relevant literature. The authors suggested that
as the technology develops, AI tools could function as intelligent research
assistants that conduct virtual experiments, analyze data, do copywriting,
edit texts, and generate citations. In terms of library reference services,
the authors discussed the increasing use of AI chatbots to answer basic
reference questions, which can free up librarian time for more complex
research queries or tasks. The authors also noted that AI tools will make
information literacy and digital literacy more important than ever and
that librarians need to teach critical thinking skills to validate facts and
evaluate the quality of the answers provided by ChatGPT. They concluded
that the introduction of ChatGPT seems similar to other innovative devel-
opments such as the introduction of calculators, cell phones, the World
Wide Web, and Wikipedia, and that libraries should evaluate these new
tools and develop services to support their use.
Chen (2023) conducted a simple test where ChatGPT was asked questions
about library services, and its responses were compared with those provided
by conventional library chatbots. As a result, ChatGPT was able to suggest
specific databases, whereas the conventional chatbots did not understand
the question or only suggested visiting the library’s A-Z database page or
general instructions. The author also noted that a customized ChatGPT
might better answer local questions, such as library hours and local resources.
The author also suggested that past lessons from the adoption of Google
and Web 2.0 can guide how to approach ChatGPT. According to the author,
the library community failed to fully recognize and utilize the potential of
Google when it was first introduced and also failed to anticipate the impact
of social media in spreading misinformation. The author concluded that
the library community should avoid underestimating or underutilizing
ChatGPT's potential to enhance library services but also acknowledge and
address its potential weaknesses and pitfalls, such as plagiarism and the
possibility of erroneous output due to poor data quality.
Panda and Kaur (2023) examined the viability of ChatGPT as an alter-
native to traditional chatbot systems in library and information centers.
According to the authors, ChatGPT represents a significant advancement
over traditional chatbots because it enables more flexible and natural
 Journal of Web Librarianship 7

language conversation. Traditional chatbots rely on predefined rules and

responses to generate answers to user queries, which can limit their flex-
ibility, scalability, and natural language capabilities. In contrast, ChatGPT
is trained on a large corpus of data and specifically designed to generate
natural language responses, making it more flexible and adaptable to var-
ious user needs. Traditional chatbots also often struggle with unexpected
and nonstandard queries, while ChatGPT can use contextual clues to
generate responses even if the question is not phrased in a typical way.
Furthermore, ChatGPT can learn from new data, while traditional chatbots
require regular maintenance to ensure that the questions and answers stay
up to date. Finally, the cost of developing and maintaining traditional
chatbots can be significant, while ChatGPT can be trained on existing
data and fine-tuned for specific tasks, potentially reducing costs.
Adetayo (2023) explored the potential of AI chatbots, particularly
ChatGPT, in academic libraries. According to the author, chatbots can
assist library patrons in accessing materials and completing tasks without
human assistance, freeing up librarians’ time for more in-depth assistance.
ChatGPT has unique features, such as generating diverse and lifelike
responses, and recognizing user intent. It can also be used for language
translation purposes. However, there are potential risks and challenges
associated with ChatGPT's use in academic libraries, including the risk of
job loss and the possibility of misuse. Additionally, ChatGPT may produce
inaccurate query responses during reference transactions and lacks the
ability to comprehend reference queries like a human librarian. The author
concludes that while ChatGPT has the potential to benefit academic librar-
ies, it is essential to carefully assess and address the possible risks and
challenges associated with its use. Libraries must develop clear standards
and guidelines, monitor their performance regularly, and use them ethically
and effectively to provide the best possible user experience.

Developing the bot

Interacting with the OpenAI API

The OpenAI API is an application programming interface that allows

developers to use the GPT large language models in their applications.
The API provides a way to interact with the GPT models and generates
natural language responses to queries. OpenAI’s chat-optimized models,
gpt-3.5-turbo and gpt-4, can discuss almost any topic and perform various
tasks without any additional training. The API’s pricing model (https://
openai.com/pricing) is based on the number of used “tokens”, which can
be considered as pieces of words. The API breaks all inputs into tokens,
which are then converted into numerical representations (vectors) that
8 Y. LAPPALAINEN AND N. NARAYANAN

can be processed by the language model. At the time of writing, the usage
cost of the gpt-3.5-turbo model (the same as the default publicly-available
ChatGPT model) is USD $0.002 per 1,000 tokens. A single question and
answer typically require fewer than 1,000 tokens, which means that the
model can answer at least 1,000 questions for approximately USD $2. This
pricing model offers a very low entry barrier and makes the API accessible
for a wide range of users.
For a custom chatbot, incorporating domain-specific data is crucial.
Currently, there are two ways to use custom data with the GPT models:

1. providing context in the prompt and instructing the model to base

its response on the context; or
2. fine-tuning the model on a custom dataset.

Modifying the prompt (also known as the emerging art of “prompt

engineering”) can be used to limit the responses into the specified context,
making it a good starting-point for a tailored chatbot. Fine-tuning is
another approach that updates the model’s weights toward the custom
dataset. According to OpenAI’s documentation (2023a), the benefit of
fine-tuning over prompt engineering is that once the training is complete,
the model can produce higher-quality results without the need to provide
the context every time in the prompt. In this approach, the responses are
not limited to any context, which potentially makes fine-tuning a more
suitable approach for general-purpose applications. Fine-tuning can also
enable lower-latency requests (OpenAI, 2023a). However, fine-tuning is
priced separately and the price depends on the selected base model.
OpenAI maintains a cookbook (https://github.com/openai/openai-
cookbook) with example scripts for accomplishing common tasks using
the OpenAI API. One of the examples, web-crawl-q-and-a (https://github.
com/openai/openai-cookbook/tree/main/apps/web-crawl-q-and-a), was an
important starting-point for our project. The code shows how to use web
crawling and OpenAI's API to answer questions based on the crawled
data. Crawling is a technique that extracts data from a website and then
stores it in a structured format for further use. The example code first
extracts text from a webpage, saves it into a csv file, and then generates
embeddings using the OpenAI API. Embedding is the process of converting
text strings into vectors (lists of floating-point numbers). These vectors
can then be used to calculate the distance between items to measure their
relatedness. If the distance is small, they are closely related, while large
distances suggest low relatedness. Embeddings are widely used in natural
language processing and the OpenAI API has a dedicated endpoint for
generating embeddings. (OpenAI, 2023a). Using embeddings is a critical
part of the process because it allows the most relevant context to be
 Journal of Web Librarianship 9

identified and added into the prompt. Generating embeddings with the
OpenAI API is also priced separately, and the price is USD $0.0004 per
1,000 tokens at the time of writing. Embeddings can also be created using
free alternatives such as the SentenceTransformers framework (https://
www.sbert.net).

Collecting and preparing the data

Library websites and guides often contain a wealth of information about

the library’s services, so these served as a natural starting-point for our
data collection. Zayed University Library maintains over 100 guides and
a large number of webpages, so collecting this data manually would be a
tedious and time-consuming project. To speed things up, we created a
custom script to crawl the guides automatically. We also used ChatGPT
as a coding assistant during the project and found it extremely useful due
to its capability to generate, debug, and comment code. This speeded up
the development process significantly.
The scraping script collects all the links from the list of guides and
then follows every link under the domain https://zu.libguides.com. The
list of LibGuides contains links that are created dynamically by JavaScript,
so the Selenium library was used to capture all the data. Selenium launches
a browser window and finds all the links on the page. The script then
extracts the headings and contents from every page, and stores them in
a dataframe. Zayed University Library’s guides have a fairly consistent
structure so we were able to collect all the content. The script also col-
lected in-text links and embedded content (e.g., videos and other pages).
Web crawling can be challenging due to the variability of data and lack
of standardization across webpages. Information can be presented in var-
ious styles and formats, making it more difficult to extract automatically.
Initially, we also planned to crawl the library’s main website, but since it
has a less-structured format than our LibGuides, we decided to collect
the information manually instead. This included basic information such
as library hours, available services, and contact details. In addition, we
used ChatGPT to generate a list of 100 typical questions and answers
regarding academic libraries. We then revised the list and updated it with
Zayed University-specific information. Furthermore, we used a set of
previously asked questions and answers from LibAnswers. The last step
was to review all data and to remove all duplicates and inconsistencies
manually. The final dataset consisted of a total number of 2,200 rows.
The Q&A script provided by OpenAI is a simple solution that creates
the embeddings, saves them into a csv file, loads them into a dataframe,
and then calculates the distances between the question and the custom
dataset. This is fine for smaller-scale implementations or testing purposes
10 Y. LAPPALAINEN AND N. NARAYANAN

but it can also lead to slow performance when dealing with larger datasets.
This is why OpenAI recommends using a vector database for searching
over many vectors quickly (OpenAI, 2023a). Based on this recommendation
and extremely helpful instructions published by Kim (2023), Yang (2023),
and Chase (2023), we decided to set up a vector database using Chroma
(https://www.trychroma.com), a toolkit designed for building AI applica-
tions with embeddings. It uses an in-process (serverless) DuckDB database,
allowing the storage and querying of embeddings and their metadata
without having to set up a dedicated server environment. At the same
time, we started using LangChain (https://python.langchain.com), a frame-
work for interfacing and working with various LLM. It facilitates data
ingestion, prompt management, embedding creation, and output parsing.
Above all, it can also be used to create chains, i.e., sequences of multiple
LLM calls, and advanced agents that use LLMs to interact with other
systems and tools. Chroma has a LangChain integration which makes it
possible to create a chain that queries the vector database first, before
passing the data to the OpenAI API. LangChain can also be used with
other LLM, which gives us more possibilities for further development.

Instructing the bot

The Q&A script in OpenAI’s cookbook is designed to answer a single

question based on the context, so it does not save the conversation history.
Since our main goal was to build a conversational bot for customer service
purposes, it was essential that the bot can keep up the conversation and
ask follow-up questions. To achieve this, we created a conversational mem-
ory that adds one previous question and response as part of the prompt.
Later, we also added a setting where the number of previous conversations
used in the prompt can be selected.
Our source data is in English, and initially the bot only responded in
English. Since ChatGPT can also do translations and have conversations
in different languages, we wanted to take advantage of this feature in our
bot as well. We also wanted the bot to sound natural and ask follow-up
questions. The GPT models can be instructed by using natural language,
so the bot was given the following instructions:

• You are Aisha, a friendly and helpful library assistant at Zayed University Library.
Provide clickable links for any URLs. Answer the questions from the perspective of
Zayed University Library. Translate responses to the language of the question. Ask
follow up questions. If you don’t know the answer, say that you don’t know. Ask
for clarifications if you don’t understand the question. Provide direct links to the
mentioned library databases and services. Remind users that all databases can also
be accessed from https://zu.libguides.com/az.php. Don’t respond if the question is
not related to Zayed University Library or its resources and services. If you cannot
answer the question, recommend contacting a relevant subject librarian.
 Journal of Web Librarianship 11

Finding the right balance between a ChatGPT-style general chatbot and

a library-specific bot has been a challenge throughout the project. We
wanted to limit the bot to answer questions about our library, but also
give it some room for improvization. OpenAI’s example Q&A script limits
the response to the given context using the following wording: Answer
the question based on the context below, and if the question can’t be answered
based on the context, say “I don’t know”. However, after initial testing, we
found this approach too restrictive because it often resulted in the bot
providing “I don’t know” responses. The bot also kept mentioning the
word “context” in its responses. To fix this, we changed the instruction
to: “Answer the question using only the information below”. In addition,
we explicitly instructed the bot to “act as a library assistant and remain
in this role throughout the conversation”. These instructions, together with
the conversational memory, significantly improved the quality of the output
and overall user experience (see Figure 1). This resulted in conversations
that were more natural, and the bot was able to remember the previous
conversations and ask follow-up questions (see our chats in Figure 1 and
the Appendix). After including the instructions to translate responses, the
bot was also able to have conversations in different languages based on
the English source material.
The GPT models are susceptible to a phenomenon known as “halluci-
nation”, where the model makes up content when it encounters gaps in

Figure 1. A conversation with Aisha.

12 Y. LAPPALAINEN AND N. NARAYANAN

its knowledge (Alkaissi & McFarlane, 2023). Unfortunately, this has been
an issue in our project as well, with the bot occasionally promoting non-ex-
istent links and library services. We tried to reduce this by providing the
following additional instructions after the context: “When providing links,
prefer those that start with https://zu.libguides.com or https://zulib.idm.
oclc.org. Do not invent non-existent links or services that are not listed in
the context”. The OpenAI API also has a “temperature” setting that controls
the randomness in the generated text. Higher values will make the output
more random and lower values more predictable, closer to the training
data. Another approach was to provide one example question and the
correct answer to the bot before the actual prompt (also known as “one-
shot learning”). Revising the instructions, setting the temperature to 0 and
incorporating one-shot learning slightly reduced the frequency of the
hallucinations, but Aisha continues to occasionally generate non-existent
links and other minor hallucinations (e.g., personal names that do not
appear anywhere in the embedded source data).

Creating the interface

There are many options for deploying Python apps online. Based on the
instructions published by Biswas (2023), we chose Streamlit (https://
streamlit.io), which is an open-source Python library for creating interactive
web applications. In Streamlit, a chatbot interface can be created with just
a couple of lines of code using the streamlit-chat component (https://pypi.
org/project/streamlit-chat). However, we decided to print the outputs using
Streamlit’s “st.markdown” function instead since this gave us more options
for customizing the look and feel of the chat. We added a custom avatar
generated by another AI model, Stable Diffusion. After initial testing, we
also added a debug mode that prints the prompt history, number of tokens,
and costs of usage. Based on Streamlit’s documentation, we also created
a Google Drive integration to record all questions and answers in a spread-
sheet that is only accessible to a few selected developers (Streamlit, 2023).
This allows us to monitor the bot’s performance and to modify the settings
based on the outputs. While Google Drive is convenient for logging the
conversations, we only intend to use it as a temporary solution during
testing as it may not meet the necessary data privacy standards in a pro-
duction environment.

Summary

To summarize, the following steps were required to set up and customize

the bot:
 Journal of Web Librarianship 13

1. Crawling content from LibGuides and website.

2. Adding more content manually.
3. Checking and cleaning the dataset.
4. Creating and storing the embeddings (using Chroma).
5. Creating a script that identifies the correct context and queries the
OpenAI API (using LangChain).
6. Instructing the bot.
7. Creating the chat interface.
8. Deploying the application (using Streamlit).

Each chat query was then processed using the following steps (Figure 2):

1. Creating an embedding for the question.

2. Identifying the nearest context from the vector database and adding
it into the prompt.
3. Querying the OpenAI API with both the question and the context.

Figure 2. Processing chat queries.

14 Y. LAPPALAINEN AND N. NARAYANAN

Discussion and future development

Perceived capabilities

Zayed University Library is dedicated to offering exceptional information

and support services to its students and faculty members. As digital tech-
nologies become more prevalent, libraries must seek out innovative solu-
tions to improve their services. Our initial experience has shown that
chatbots hold great potential for providing personalized, accessible, and
cost-effective support to library users. Based on initial testing, the chatbot
can provide very realistic and human-like responses, keep up the conver-
sation by asking follow-up questions, and respond in different languages
based on English source data. Since the bot is using the gpt3.5-turbo
model, it can improvise its responses, write poems, tell jokes, etc., just
like its “bigger cousin” ChatGPT, although the number of tokens and the
context window size are smaller in our solution.
The chatbot has tremendous potential to speed up reference services
and make them more accessible, providing library users with round-the-
clock assistance. The chatbot could also be taught to recognize specific
users and tailor its responses to them based on their previous experiences
with the system by utilizing machine learning methods. This could produce
a smoother and more exciting user experience, increasing user satisfaction
and library service use. Additionally, by including features like speech-to-
text or text-to-speech conversion or alternative methods to engage with
the system, the bot could be made to deliver accessible responses for
people with impairments. This can ensure that everyone who uses the
library can get assistance and knowledge from it, regardless of their abil-
ities. The bot could also be integrated into other library systems, which
could increase the visibility and accessibility of all library services.

Current limitations

Although the bot has performed well in initial testing, the implementation
still has certain limitations. First of all, the OpenAI API has token restric-
tions based on the selected model. OpenAI's gpt-3.5-turbo model currently
has a limit of 4,096 tokens, including both the input and the output. This
is not a major issue when it comes to chats, because the questions are
typically brief and require a total of 500-1,500 tokens per question, includ-
ing the question, context and the output. However, the full chat history
cannot be preserved for long since the token limit would be reached quickly.
Another issue is that the bot could be “tricked” by providing additional
data in the prompt (also known as “prompt injection”). This can result
in unreliable or questionable responses and the bot could start performing
tasks beyond its intended scope. Privacy issues may also arise if personal
 Journal of Web Librarianship 15

data are passed to the OpenAI API. However, these issues could be pre-
vented by adding measures for detecting and filtering unwanted prompts
before even passing them to the OpenAI API.
One major downside of the implementation is that the bot currently
has no real-time access to information online, for example the library
website. However, this could be solved by loading and embedding specific
information automatically (such as opening hours and library events) on
a regular basis. We also expect that the ChatGPT plugins, which are cur-
rently under development, will be incorporated into the OpenAI APIs in
the coming months, making it easier to retrieve real-time information
from websites and other sources.
Librarians often receive complex Reference questions about specific papers
and topics. Since the bot currently has no real-time access to online informa-
tion, it cannot answer specific questions about individual papers. It is possible
to build functionality that allows users to upload their own documents, embed
them, and ask questions about their contents (e.g., Dara, https://www.dara.
chat). However, copyright and privacy issues could arise if all data are processed
on OpenAI's servers. One possible solution would be to embed the paper
locally and then query the OpenAI API with the relevant question and context
only, instead of passing the full text to the API. When it comes to interpreting
a paper’s content, a ChatGPT-based bot could easily outperform human librar-
ians, because it can process an entire scientific paper in a matter of seconds.
A ChatGPT-based bot could also provide general guidance in research methods
and citations, especially if such materials are available in the source data.
However, one of the most important aspects of human librarians is the ability
to provide search assistance and recommend specific library resources, which
requires a thorough and up-to-date understanding of the subject area and the
library’s collections. Achieving human-level recommendations with a chatbot
would require at least full access to the library discovery service and possibly
a memory function that keeps track of recommended resources, latest acqui-
sitions and perhaps even the latest trends in different fields. This is an inter-
esting area that calls for further research.
So far, we have only tested the bot informally among library staff - about
15 people representing all library teams. The bot saves all questions and
answers in an access-restricted Google spreadsheet, and we have used this
data to refine the bot’s instructions and source materials. After reviewing
around 500 unique questions and answers, we identified three main issues:
1) the bot often generates non-existent links, as described earlier; 2) the
bot may mistake a link from the source data as a (subscribed) library
service, although it is only mentioned as an additional resource, and; 3)
the bot cannot answer questions that require real-time data or access to a
specific resource (for example “When is the next library workshop?” or
“Can you recommend a good research article about AI?”). Despite these
16 Y. LAPPALAINEN AND N. NARAYANAN

issues, we were pleased to notice that there were very few factual errors,
and most of them were due to outdated or erroneous source data. The
generation of non-existent links remains a challenge at the time of writing,
but the latter two issues can be corrected by revising the source materials
and ingesting certain content (e.g., library event calendar) on a regular basis.

Future development

In the next phase of our project, we plan to move forward with more
formal testing by engaging Zayed University students and faculty. We are
excited to study how the bot performs in larger-scale testing and to hear
feedback and development ideas from library users. As a follow-up study,
we intend to do a more-formal analysis of the bot’s outputs and compare
it with a keyword-based chatbot solution.
During testing, we noticed that managing the embeddings can be challenging
in the long run, and another interface is needed for creating and updating
embedded contents. By developing an interface, a larger number of library
staff will be able to manage the embedded contents. We are also planning to
create a feedback mechanism in the bot that allows users to indicate their
opinions about the bot’s performance (for example, upvoting or downvoting
the response). In case of downvoting, the user could be instructed to provide
textual feedback or contact a liaison librarian for further questions.
The bot is currently instructed to provide contact details when it cannot
answer a question. This could be developed further, for example by con-
necting the user to a live chat with a librarian automatically during library
hours or by providing a form to ask further questions or report the issue.
This feedback would be valuable for further development.
Another development idea is adding a cache to reduce the number of
LLM calls and increase the bot’s performance. A solution called GPTCache
(https://gptcache.readthedocs.io) has already been developed for this pur-
pose. Another interesting direction to explore is the implementation of
AI agents which have the potential to facilitate interactions with other
library systems. With the use of agents, users could potentially carry out
various tasks such as searching library databases, renewing loans, or mak-
ing other requests, directly through the chatbot. This could greatly enhance
the user experience and streamline the overall process of accessing library
resources. Projects such as BabyAGI (https://github.com/oliveirabruno01/
babyagi-asi), Auto-GPT (https://github.com/Significant-Gravitas/Auto-GPT),
and AgentGPT (https://github.com/reworkd/AgentGPT) are already avail-
able to help with the development of intelligent agents. One particularly
interesting possibility would be to connect the chatbot to the library
discovery service, allowing it to query and recommend specific library
materials. We have also begun testing speech-to-text and text-to-speech
 Journal of Web Librarianship 17

capabilities to improve accessibility and to provide an alternative way to

interact with the bot.

Assessment of the method

The field of generative AI and LLM has evolved rapidly in early 2023.
The emergence of new tools and frameworks (e.g., LangChain and various
vector databases) has facilitated this growth, but the fast pace of updates
has also made it challenging to keep up with the latest developments.
Documentation is still scarce and developers seem to rely heavily on
experimenting and scattered instructions posted on discussion forums and
other platforms. As a result, much of our development process has been
driven by trial and error. Finding the right system messages and prompts
for OpenAI’s API has been particularly challenging due to the lack of
clear instructions. Despite trying different prompting options, we have not
been able to completely eliminate the issue of hallucination, which can
lead to inaccurate or confusing bot responses.
Remarkably, if we do not put a price on our own work, the implemen-
tation has not cost us anything yet. The OpenAI API provides free tokens
worth $18.00 USD and we have not used all of them at the time of writing.
At current API prices, answering questions using the ChatGPT API (gpt3.5-
turbo) costs about $0.002 USD per question. As a result, a GPT-based chatbot
could provide 24/7 support at a significantly lower cost than hiring additional
staff or implementing other support systems. Nevertheless, we do not con-
sider chatbots as a threat to library jobs. On the contrary, we believe that
chatbots have the potential to free up staff time and resources, allowing
library personnel to focus on more complex and specialized Reference que-
ries, as well as other critical duties within the library. This shift in respon-
sibilities can ultimately lead to a more productive and efficient library.
Using ChatGPT as a coding assistant during the project significantly helped
the coding process. The fact that we can now use ChatGPT for writing and
reviewing code and to give us development ideas shows that we have entered
a fascinating new era in software development and human-computer interaction
in general. We are expecting a wave of new GPT-based solutions in the upcom-
ing months. However, as the field of generative AI continues to evolve, user-
friendly tools, accessible documentation and standardized practices are required
to help the creation of LLM-based applications.

Conclusion
In this article, we described the development of Aisha, a custom ChatGPT-
powered chatbot at Zayed University Library. We also reviewed the history
of chatbots and early literature on ChatGPT in the context of academic
18 Y. LAPPALAINEN AND N. NARAYANAN

libraries. In conclusion, we believe that chatbots based on ChatGPT and

other LLM have great potential to transform library Reference services by
offering personalized, accessible, and cost-effective support to users.
While chatbots are one potential use case for LLM, we also acknowledge
that ChatGPT and other LLM have a lot more to offer for libraries and
many other fields. In the context of academic libraries, some potential use
cases include personalized search assistance, research analysis, supporting
cataloging and metadata generation, analyzing customer feedback, and gen-
erating ideas and texts for library’s outreach, marketing, instruction and
support materials. LLM-based AI research assistants could be particularly
helpful for students and researchers, since they can summarize academic
texts in an instant and enable users to ask focused questions about the texts.
The last two years have marked a significant breakthrough in the field
of generative AI. Several new LLM and other generative AI models have
been introduced, resulting in hundreds, if not thousands, of new AI-based
services, tools, frameworks and projects in a relatively short period of
time. While there are still many technical, ethical, legal and other chal-
lenges to overcome, it is clear that a major transformative change is under
way and there is no turning back. As academic librarians, we are excited
to explore the ways in which AI can improve research, enable better access
to information and advance scholarly communication, while also being
mindful of the ethical implications and challenges that come with this
new technology.

ORCID
Yrjo Lappalainen http://orcid.org/0000-0003-0942-6377
Nikesh Narayanan http://orcid.org/0000-0002-2005-1177

References
Adetayo, A. J. (2023). Artificial intelligence chatbots in academic libraries: The rise of
ChatGPT. Library Hi Tech News, 40(3), 18–21. https://doi.org/10.1108/LHTN-01-2023-
0007
Alkaissi, H., & McFarlane, S. I. (2023). Artificial hallucinations in ChatGPT: Implications
in scientific writing. Cureus, 15(2), Article e35179. https://doi.org/10.7759/cureus.35179
Allison, D. (2012). Chatbots in the library: Is it time? Library Hi Tech, 30(1), 95–107.
https://doi.org/10.1108/07378831211213238
Ashfaque, M. W. (2022). Analysis of different trends in chatbot designing and develop-
ment: A review. ECS Transactions, 107(1), 7215–7227. https://doi.org/10.1149/10701.7215ecst
Biswas, A. (2023, March 19) How to build a chatbot with ChatGPT API and a conver-
sational memory in Python. Medium. https://medium.com/@avra42/how-to-build-a-
chatbot-with-chatgpt-api-and-a-conversational-memory-in-python-8d856cda4542
Black, S., Biderman, S., Hallahan, E., Anthony, Q., Gao, L., Golding, L., He, H., Leahy,
C., McDonell, K., Phang, J., Pieler, M., Prashanth, U. S., Purohit, S., Reynolds, L., Tow,
 Journal of Web Librarianship 19

J., Wang, B., & Weinbach, S. (2022). GPT-NeoX-20B: An open-source autoregressive

language model. ArXiv:2204.06745v1. https://arxiv.org/abs/2204.06745
Bloomberg. (2023, March 30). Introducing BloombergGPT, Bloomberg’s 50-billion parame-
ter large language model, purpose-built from scratch for finance. https://www.bloomberg.
com/company/press/bloomberggpt-50-billion-parameter-llm-tuned-finance/
Chase, H. (2023). Chat over documents with chat history. LangChain documentation.
Retrieved April 6, 2023, from https://python.langchain.com/en/latest/modules/chains/
index_examples/chat_vector_db.html
Chen, X. (2023). ChatGPT and its possible impact on library reference services. Internet
Reference Services Quarterly, 27(2), 121–129. https://doi.org/10.1080/10875301.2023.
2181262
Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung,
H. W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A.,
Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., … Fiedel, N. (2022). PaLM: Scaling language
modeling with pathways. ArXiv:2204.02311v5. https://arxiv.org/abs/2204.02311
Cotton, D. R., Cotton, P. A., & Shipway, J. R. (2023). Chatting and cheating: Ensuring
academic integrity in the era of ChatGPT. Innovations in Education and Teaching
International, 1–12. Advance online publication. https://doi.org/10.1080/14703297.2023
.2190148
Cox, C., & Tzoc, E. (2023). ChatGPT: Implications for academic libraries. College &
Research Libraries News, 84(3), 99–102. https://crln.acrl.org/index.php/crlnews/article/
view/25821
Deshpande, A., Shahane, A., Gadre, D., Deshpande, M., & Joshi, P. M. (2017). A survey
of various chatbot implementation techniques. International Journal of Computer
Engineering and Applications, 11(Special Issue, May 17). https://www.ijcea.com/survey-
various-chatbot-implementation-techniques
Dey, N., Gosal, G., Chen, Z., Khachane, H., Marshall, W., Pathria, R., Tom, M., & Hestness,
J. (2023). Cerebras-GPT: Open compute-optimal language models trained on the Cerebras
wafer-scale cluster. ArXiv:2304.03208v1. https://arxiv.org/abs/2304.03208
Ehrenpreis, M., & DeLooper, J. (2022). Implementing a chatbot on a library website. Journal
of Web Librarianship, 16(2), 120–142. https://doi.org/10.1080/19322909.2022.2060893
Geng, X., Gudibande, A., Liu, H., Wallace, E., Abbeel, P., Levine, S., & Song, D. (2023,
April 3). Koala: A dialogue model for academic research. The Berkeley artificial intel-
ligence research blog. http://bair.berkeley.edu/blog/2023/04/03/koala/
Gozalo-Brizuela, R., & Garrido-Merch́an, E. C. (2023). ChatGPT is not all you need. A
state of the art review of large generative AI models. ArXiv:2301.04655v1. https://arxiv.
org/abs/2301.04655
Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas,
D. D. L., Hendricks, L. A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K.,
Driessche, G., van den, Damoc, B., Guy, A., Osindero, S., Simonyan, K., Elsen, E., …
Sifre, L. (2022). Training compute-optimal large language models. ArXiv:2203.15556v1.
https://arxiv.org/abs/2203.15556
Hsiao, S. (2023, May 10). What’s ahead for Bard: More global, more visual, more inte-
grated. Google Blog. https://blog.google/technology/ai/google-bard-updates-io-2023
Hu, K. (2023, February 2). ChatGPT sets record for fastest-growing user base - Analyst
note. Reuters. https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-
user-base-analyst-note-2023-02-01
Hussain, S., Ameri Sianaki, O., & Ababneh, N. (2019). A survey on conversational agents/
chatbots classification and design techniques. Advances in Intelligent Systems and
Computing, 927, 946–956. https://doi.org/10.1007/978-3-030-15035-8_93
20 Y. LAPPALAINEN AND N. NARAYANAN

Kane, D. (2019). Creating, managing and analyzing an academic library chatbot. BiD, 43
https://doi.org/10.1344/BiD2019.43.22
Kim, S. (2023, March 20) How to ensure OpenAI's GPT-3 provides an accurate answer
using embedding and semantic search. Dev Genius, https://blog.devgenius.io/how-to-
ensure-openais-gpt-3-provides-an-accurate-answer-and-stays-on-topic-af5da300ba81
King, M. R, ChatGPT. (2023). A conversation on artificial intelligence, chatbots, and
plagiarism in higher education. Cellular and Molecular Bioengineering, 16(1), 1–2. https://
doi.org/10.1007/s12195-022-00754-8
Lim, W. M., Gunasekara, A., Pallant, J. L., Pallant, J. I., & Pechenkina, E. (2023). Generative
AI and the future of education: Ragnarök or reformation? A paradoxical perspective
from management educators. The International Journal of Management Education, 21(2),
100790. https://doi.org/10.1016/j.ijme.2023.100790
Lund, B. D., & Wang, T. (2023). Chatting about ChatGPT: How may AI and GPT impact
academia and libraries? Library Hi Tech News, 40(3), 26–29. https://doi.org/10.1108/
LHTN-01-2023-0009
Mauldin, M. L. (1994). Chatterbots, TinyMUDs, and the Turing Test: Entering the Loebner
Prize competition. In Hayes-Roth, B. & Korf, R. E. (Eds.), AAAI-94: Proceedings of the 12th
national conference on artificial intelligence (pp. 16–21). AAAI Press. https://aaai.org/
papers/00016-chatterbots-tinymuds-and-the-turing-test-entering-the-loebner-prize-competition/
Mckie, I. A. S., & Narayan, B. (2019). Enhancing the academic library experience with
chatbots: An exploration of research and implications for practice. Journal of the
Australian Library and Information Association, 68(3), 268–277. https://doi.org/10.1080
/24750158.2019.1611694
McNeal, M. L., & Newyear, D. (2013). Introducing chatbots in libraries. Library Technology
Reports, 49(8), 5–10. https://www.journals.ala.org/index.php/ltr/article/view/4504/5281
Meta, AI. (2023, February 24). Introducing LLaMA: A foundational, 65-billion-parameter
language model. https://ai.facebook.com/blog/large-language-model-llama-meta-ai/
OpenAI. (2022, November 30). Introducing ChatGPT. Blog. https://openai.com/blog/chatgpt
OpenAI. (2023a). OpenAI Documentation. Retrieved April 6, 2023, from https://platform.
openai.com/docs
OpenAI. (2023b, March 23). ChatGPT plugins. OpenAI Blog. https://openai.com/blog/
chatgpt-plugins
OpenAI. (2023c, May 12). Web browsing and Plugins are now rolling out in beta. ChatGPT
Release Notes. https://help.openai.com/en/articles/6825453-chatgpt-release-
notes#h_9894d7b0a4
Panda, S., & Kaur, N. (2023). Exploring the viability of ChatGPT as an alternative to
traditional chatbot systems in library and information centers. Library Hi Tech News,
40(3), 22–25. https://doi.org/10.1108/LHTN-02-2023-0032
Pichai, S. (2023, February 6). Google AI updates: Bard and new AI features in search.
Google Blog. https://blog.google/technology/ai/bard-google-ai-search-updates
Ren, X., Zhou, P., Meng, X., Huang, X., Wang, Y., Wang, W., Li, P., Zhang, X., Podolskiy,
A., Arshinov, G., Bout, A., Piontkovskaya, I., Wei, J., Jiang, X., Su, T., Liu, Q., & Yao,
J. (2023). Pangu-Sigma: Towards trillion parameter language model with sparse hetero-
geneous computing. ArXiv:2303.10845v1. https://arxiv.org/abs/2303.10845
Rodriguez, S., & Mune, C. (2021). Library chatbots: Easier than you think. Computers in
Libraries, 41(8), 29–32. https://www.infotoday.com/cilmag/oct21/Rodriguez-Mune–
Library-Chatbots-Easier-Than-You-Think.shtml
Soltan, S., Ananthakrishnan, S., FitzGerald, J., Gupta, R., Hamza, W., Khan, H., Peris, C.,
Rawls, S., Rosenbaum, A., Rumshisky, A., Prakash, C. S., Sridhar, M., Triefenbach, F.,
Verma, A., Tur, G., & Natarajan, P. (2022). AlexaTM 20B: Few-shot learning using a
 Journal of Web Librarianship 21

large-scale multilingual seq2seq model. ArXiv:2208.01448v2. https://arxiv.org/

abs/2208.01448
Streamlit. (2023). Connect streamlit to a private Google Sheet. Streamlit Documentation.
https://docs.streamlit.io/knowledge-base/tutorials/databases/private-gsheet
Taori, R., Gulrajani, I., Zhang, T., DuBois, Y., Li, X., Guestrin, C., Liang, P., & Hashimoto, T. B.
(2023, March 13). Alpaca: A strong, replicable instruction-following model. Stanford University
Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html
Technology Innovation Institute. (2023, March 15). Abu Dhabi-based technology innovation
institute introduces Falcon LLM: Foundational Large Language Model (LLM) outperforms
GPT-3 with 40 Billion Parameters. https://www.tii.ae/news/abu-dhabi-based-technology-
innovation-institute-introduces-falcon-llm-foundational-large
University of Oklahoma. (2020). Introducing “Bizzy”. University of Oklahoma Libraries.
https://libraries.ou.edu/content/introducing-bizzy
Wallace, R. S. (2009). The anatomy of A.L.I.C.E. In Parsing the Turing Test: Philosophical
and methodological issues in the quest for the thinking computer (pp. 181–210). https://
doi.org/10.1007/978-1-4020-6710-5_13
Weizenbaum, J. (1966). ELIZA 2013 A computer program for the study of natural language
communication between man and machine. Communications of the ACM, 9(1), 36–45.
https://doi.org/10.1145/365153.365168
Yang, S. (2023, March 27). The easiest way to work with large language models? Towards
Data Science, https://towardsdatascience.com/the-easiest-way-to-interact-with-language-
models-4da158cfb5c5

Appendix:  An example conversation with Aisha

Authors: How are you doing, Aisha?
Aisha: I'm doing great, thank you for asking! How can I assist you with your library
needs today?
Authors: I have a question about Scival
Aisha: Of course, I'd be happy to help! What would you like to know about Scival?
Authors: What can I do with it?
Aisha: Scival is a research analytics tool that can help you analyze research perfor-
mance, identify research trends, and find potential collaborators. You can use it to search
for publications, authors, institutions, and funding opportunities. Is there anything specific
you would like to know about Scival?
Authors: How can I access it?
Aisha: You can access Scival by visiting the Scival login page. Please note that each
researcher needs to register for an account before using it. We also have a guide available
that provides guidance on how to use Scival effectively. Let me know if you have any oth-
er questions about Scival.
Authors: So I need to register?
Aisha: Yes, you need to register for an account before using Scival. Once you are on
the login page, click on “Register Now” and follow the instructions to create your account.
If you need any further assistance, please let me know.
Authors: Can you give me step by step instructions?
Aisha: Certainly! To register for an account on Scival, please follow these steps:
1. Go to the Scival login page.
2. Click on “Register Now” located below the login form.
3. Fill in the required information such as your name, email address, and password.
4. Click on “Create Account” to complete the registration process.
22 Y. LAPPALAINEN AND N. NARAYANAN

5. Once you have registered, you can log in to Scival using your email address and
password.
I hope this helps! Let me know if you have any further questions or if there’s anything
else I can assist you with.
Authors: Thank you so much!
Aisha: Sure, I'm glad I could help! If you have any other questions or need further
assistance, feel free to ask.