Building an AI Chatbot Using LLM

A major project report submitted in partial fulfilment of the requirement

for the award of degree of

Bachelor of Technology
in
Computer Science & Engineering / Information Technology

Submitted by
Abhishek (201559)
Ishika Goswami (201331)

Under the guidance & supervision of

Dr. Deepak Gupta

Department of Computer Science & Engineering and

Information Technology
Jaypee University of Information Technology,
Waknaghat, Solan - 173234 (India)
 CANDIDATE’S DECLARATION

I hereby declare that the work presented in this report entitled ‘Building an AI Chatbot
Using LLM’ in partial fulfilment of the requirements for the award of the degree of Bachelor
of Technology in Computer Science & Engineering / Information Technology submitted
in the Department of Computer Science & Engineering and Information Technology, Jaypee
University of Information Technology, Waknaghat is an authentic record of my own work
carried out over a period from August 2023 to May 2024 under the supervision of Dr.
Deepak Gupta (Assistant Professor, Department of Computer Science & Engineering and
Information Technology).

The matter embodied in the report has not been submitted for the award of any other degree
or diploma.

(Student Signature with Date) (Student Signature with Date)

Student Name: Abhishek Student Name: Ishika Goswami
Roll No.: 201559 Roll No.: 201331

This is to certify that the above statement made by the candidate is true to the best of my
knowledge.

(Supervisor Signature with Date)

Supervisor Name: Dr. Deepak Gupta
Designation: Assistant Professor (SG)
Computer Science and Engineering/Information Technology Department
Dated:

I
 ACKNOWLEDGEMENT
To begin, I would like to express my heartfelt gratitude to almighty God for his heavenly
grace, which enabled us to successfully complete the project work.

I am extremely grateful and wish to express my deep gratitude to Supervisor Dr. Deepak
Gupta, Assistant Professor (Senior Grade), Department of CSE & IT Jaypee
University of Information Technology, Waknaghat. His never-ending patience,
intellectual direction, persistent encouragement, constant and vigorous supervision,
constructive criticism, helpful suggestions, and reading numerous poor versions and
revising them at all stages allowed this project to be completed. I would like to express my
heartiest gratitude to Dr. Vivek Sehgal, Head of Department of CSE & IT, for his kind
help to finish my project.

I would also like to express my gratitude to everyone who has assisted me in making this
project a success, whether directly or indirectly. In this unusual scenario, I'd like to express
my gratitude to the different staff members, both teaching and non-teaching, who have
provided me with valuable assistance and assisted my project. Finally, I must express my
gratitude for my parents' unwavering support and patience.

Abhishek (201559)

Ishika Goswami (201331)

II
 TABLE OF CONTENT
CANDIDATE’S DECLARATION i

ACKNOWLEDGEMENT ii

LIST OF ABBREVIATIONS v

LIST OF TABLES vi

LIST OF FIGURES vii

ABSTRACT viii

1 INTRODUCTION ……………………………………………………...…….1-13
1.1 Introduction …………………………………………………………...………….1
1.2 Problem Statement ……………………………………………………...………. 2
1.3 Objectives ……………………………………………………………...……….. 2
1.4 Significance and motivation of the project report ……………………...………. 2
1.5 LLMs ……………………………………………………………………...……. 3
1.6 Benefits of using LLMs ……………………………………………………...…. 4
1.7 Methodology ………………………………………………………………...….. 5
1.7.1 NLU ……………………………………………………………...………. 6
1.7.2 NLP ……..………………………………………………………...……… 8
1.7.3 Deep Learning …………………………………………………….......… 10
1.8 Language…………………….…………………………………………...…….. 11
1.9 Organization of project report ……………………………………………...….. 12

2 LITERATURE SURVEY ………………………………………….14-24

2.1 Overview of relevant literature ………………………………………………… 14
2.2 Key gaps in the literature ………………………………………………………. 21

3 SYSTEM DEVELOPMENT …………………………………...….25-43

3.1 Requirements and Analysis ………………………………………...………….. 25
3.1.1 Functional Requirements …………………………………………...…… 25
3.1.2 Non-Functional Requirements ……………………………..……………. 25

III
 3.2 Project Design and Architecture ……………………………………..……........ 27
3.3 Data Preparation……………………………………………………………....... 32
3.4 Implementation……………………………………………………………......... 32
3.4.1 Gradio …………………………………………...………………………. 32
3.4.2 Steps Followed ………………………………………………………….. 33
3.4.3 OpenAI Key …………………………………………………………….. 34
3.4.4 Hugging Face …………………………………………………………… 35
3.4.5 Langchain ………………………………………………………………. 36
3.4.6 Llama …………………………………………………………………… 37
3.4.7 Streamlit ………………………………………………………………… 39
3.4.8 Dotenv …………………………………………………………………... 41
3.4.9 Modules ……………………………..………………………….……….. 43
3.5 Key Challenges……………………………………………………………........ 46

4 TESTING ………………………………………………………..….48-49
4.1 Testing Strategy ………………………………………………………………... 48
4.2 Test Cases and Outcomes ……………………………………………………… 48

5 RESULTS AND EVALUATION ………………………………….50-54

5.1 Results …………………………………………………………………………. 50

6 CONCLUSION AND FUTURE SCOPE…………………………..55-56

6.1 Conclusion ……………………………………………………………………….55
6.2 Future Scope ……………………………………………………………………..55

REFERENCES …………………………………………………………57-59

APPENDIX……………………………………………………..…………...60

IV
 LIST OF ABBREVIATIONS

Abbreviations Meaning

NLP Natural language processing

NLU Natural language understanding

RNN Recurrent neural network

Chatbot Chatting Robot

AI Artificial Intelligence

GPT Generative Pre-trained Transformer

API Application Programming Interface

V
 LIST OF TABLES
Table 1: Summary of Relevant Literature…………………………………...… Page 18-20

VI
 LIST OF FIGURES
Figure 1.1: NLG & NLP Being a Subset of NLP ………………………………….. Page 5
Figure 1.2: Generative AI Being a Subset of Deep Learning………………….….. Page 10
Figure 3.1: Project Flowchart…………………………………………...………..... Page 27
Figure 3.2: The Transformer Architecture……………………………...…………. Page 28
Figure 3.3: Multiplication of Query and Key Matrices………………..……......… Page 30
Figure 3.4: Division of Scores Matrix by Square Root……….................……..…. Page 31
Figure 3.5: Multiplication of Attention & Value Matrices……………….……….. Page 31
Figure 5.1: Custom Design Interface Output…………………………………....… Page 50
Figure 5.2: Assignment of Role………………………………………….…...…… Page 50
Figure 5.3: Generation of Local and Public URL……………………….…...….… Page 50
Figure 5.4: Result for Question 1………………………………..………...….…… Page 51
Figure 5.5: Result for Question 2………………………………..……...…….….... Page 51
Figure 5.6: Result for Question 3………………………………………………..… Page 51
Figure 5.7: Assignment of Role……………………………………………..…….. Page 52
Figure 5.8: Generation of Local and Public URL……………………………..…... Page 52
Figure 5.9: Result for Question………………………………………………..…... Page 52
Figure 5.10: Error as Output When Question is Out of Domain………………….. Page 52
Figure 5.11: Generation of Local and Network URL……………………………... Page 53
Figure 5.12: Interface After Running URL………………………………………... Page 53
Figure 5.13: Result of Query Put by User with Extra Information………………... Page 54

VII
 ABSTRACT
In our digital world, where communication is key, the emergence of AI-powered chatbots
has revolutionized the way we interact with technology. This project delves into the
development of an AI chatbot utilizing Large Language Models (LLM), a cutting-edge
technology in the field of natural language processing. The journey begins with a
comprehensive exploration of existing chatbot frameworks and technologies, analysing their
strengths and weaknesses. Through meticulous research, we identified LLM as the most
promising candidate due to its ability to understand and generate human-like text, enabling
more meaningful and engaging conversations.

The LLM was employed for its ability to process long range dependencies, enabling the
chatbot to comprehend context in a more comprehensive manner. The report discusses the
methodology employed, including data preprocessing, model training, the transformer
architecture and validation using the test cases along with the conclusion and the work we
wish to do in the upcoming months. Additionally, it details the challenges encountered and
the strategies employed to counter them and enhance the chatbot’s performance.

The final AI chatbot demonstrates impressive capabilities in understanding and generating

human-like responses across a diverse range of topics. Its ability to engage users in
meaningful conversations showcases the potential of LLM technology in enhancing human-
computer interaction.

Overall, this project contributes to the evolving field of AI-driven conversational agents and
highlights the potential of utilizing LLMs in building sophisticated chatbot systems. The
journey that we lived while working on this project is mentioned in this project report in the
sequential order.

VIII
 CHAPTER 01: INTRODUCTION

This chapter of the project report is the beginning of the content of this report. It contains the
building up of the plot of this report. The problem statement along with the main objectives of
this project are discussed in here. The significance of this project and the real motivation behind
the intentions to take up this topic as our project are also listed in detail in this particular chapter.
The organization of this project report is also listed in this very chapter. A basic description
regarding the LLMs in the title of this project report is also given here.

The various methodologies we came across while we started working on this project are also
explained. The language used for working on this project is also mentioned in this chapter.

1.1 INTRODUCTION

In the time of modern technology, AI has emerged as a ground breaking technology. AI

Chatbots which use Natural Language Processing and Machine Learning techniques have
gained a significant role and platform. In recent years, AI has witnessed extraordinary growth,
especially in the field of conversational agents or Chatbots. These Chatbots are designed to
engage in conversation with users, providing with relevant information and assistance. Among
the various approaches used to create these conversational agents, one of the most promising
and sophisticated methods involve LLMs.

To comprehend the significance and association of LLMs in Chatbot creation, its essential to
understand the core concepts developing this technology. LLMs are a form of AI that processes
and generates human like text, learning patterns and structures form vast amounts of data.
These models, such as OpenAI’s GPT have gained prominence due to their ability to
understand context and semantic difficulties with human language.

The primary objective of this report is to provide a comprehensive overview of the steps and
considerations involved in building an AI Chatbot using LLMs. From the initial data gathering
and preprocessing stages to fine tuning the model and deploying the Chatbot.
This report will also touch upon the ethical considerations associated with deploying AI
Chatbots. Issues such as data privacy and bias are also discussed highlighting the importance
of ethical practices in the development and implementation of these Chatbots.

1
1.2 PROBLEM STATEMENT

This project aims to develop an advanced AI Chatbot utilizing LLMs to enhance the human-
AI interaction. The challenge involves building a Chatbot that can accurately understand and
respond to natural language unput in various contexts. The Chatbot should maintain context
over multi turn conversation, provide relevant and coherent responses and continuously learn
from the user interactions to improve its performance.

Moreover, the ethical aspect is crucial, necessitating the prevention of harmful or inappropriate
content generation. Providing the Chatbot with adaptivity to the individual user preferences
and providing tailored responses based on user history and stated preferences is also a challenge
to overcome in this project.

1.3 OBJECTIVES

The main objectives which we try to aim during the completion of this project are all listed
below:

● To develop a sophisticated NLP system that can accurately interpret user inputs, identify
the intents and extract key entities.
● To create a Chatbot capable of maintaining context across multiturn conversations and
managing topic transitions fluidly.
● To create a Chatbot that interacts in a way that feels natural and engaging blurring the line
between human-human and human-AI conversations.
● To develop a Chatbot whose responses adhere to ethical standards and avoid generating
harmful or offensive content.

1.4 SIGNIFICANCE AND MOTIVATION OF THE PROJECT WORK

We as human beings tend to produce goods and services that mostly end up easing our day-to-
day work in the best way possible. One such invention are the LLMs. Work on the topic related
to the machines having the ability to understand the human language and converse in a
language native to humans began in around 1967 when Eliza was invented. Since then, this has
been ongoing journey. But it achieved a massive push during 2017 with the release of
Transformers and then during 2022 when OpenAI released ChatGPT [1].

2
This not only led to ease the human life but also provided the new area of research and
advancement in the field of AI. The significance lies in the application of language models that
enable the Chatbot to comprehend and respond to human input in a way that feels natural and
engaging. This is more than just writing a program to answer questions.

This project holds great importance for several reasons. Firstly, it represents a practical
application of the cutting-edge technology in the field of AI. By working on this project, we
are exploring how advanced these language models can be and how they can be used along
with gaining knowledge on their working behind the scenes.

This project provides an opportunity to enhance our understanding and skills in the field of AI.
The real hands-on experience that we are gaining during this project is also of great significance
as the field of AI is a never-ending field in which the advancements will continue for a very
long time.

Furthermore, the implementation of an AI Chatbot has significant real-world implications.

Chatbots are increasingly being used in various industries such as customer service, healthcare
and now even in education to provide assistance. By engaging in this project, we are actively
contributing to the growing field of generative AI and exploring its potential applications.

One of the primary motivations behind this project is the desire to explore and understand the
capabilities of AI especially in the domain of the rising LLMs. By diving into the development
of an AI Chatbot, we actively participate in understanding the evolution of AI and gain valuable
knowledge about the behind the scene working.

The motivation also arises from the awareness of the growing role of AI in our daily lives.
From virtual assistants on smartphones to automated customer services, AI is increasingly
becoming an integral part of various industries. Understanding the capabilities and limitations
of AI Chatbots also drive us.

1.5 LLMs

LLMs are simply powerful Artificial Intelligence models that are trained on vast amounts of
text data. The main objective is to understand and generate human like text. These models use
deep learning techniques. They process and generate text based on the patterns and structures
they have learned on the training data [1]. During the testing phase, they interpret the patterns

3
in the data and try to process those as in the training data and come up with similar results. So,
the quality and quantity of the training dataset are important factors in the working of LLMs
[2].

Large refers to the enormous amount of training dataset and the large number of parameters.
Only certain organizations have the capability to train such LLMs with huge datasets and
tremendous number of parameters [3], [4]. LLMs are trained with petabytes of data and
generate billions of parameters. They have to undergo pre-training and fine tuning. Pre-training
means to train the model for a general purpose with a large data set. The fine tu ning involves
tuning the model for specific aims with a much smaller dataset.

LLMs have to go through a complex process during the training that involves multiple steps.
It usually starts with the unsupervised learning in which the model learns to make all the
observation on its own. So, it is important that the data must be of good quality. The model
creates its own clusters based on its observations of the data as per [2]. Then comes a step that
involves supervised learning. In this step the human supervisors make the class labels based on
the data and the model has to classify the data based on those known class labels. The model
just has to classify the data from the dataset such that each of the data chunk falls into either
one of those class labels.

Then comes the reinforcement learning from human feedback (RHLF). This involves using
rewards as a way to give the system incentive to find new patterns [3]. Here, the quality of the
outcome given by model is improved. This allows the model to make good predictions. The
outputs with the greatest number of rewards are only kept and the ones with the lowest rank
are discarded. This is how the model improves its quality over time.

The LLMs are designed based on the Transformer architecture that was introduced in the year
2017. It overcame the limitations of the RNN architecture as per [1] and made the LLMs
famous and popular on a very large scale which is discussed in detail in the upcoming chapters
of this project report.

1.6 BENEFITS OF USING LLMS

Although there are many ways for creating a Chatbot. But still we decided on doing so with
the help of LLMs. This is because of the popularity of these language models. But it’s not

4
just their popularity that we kept in mind while choosing them. These language models also
provide certain benefits which are listed below:

● A single model can be used for a variety of different tasks. This is due to the fact that LLMs
are trained on petabytes of data and generate about a billion parameters which enable them
to perform different types of tasks including language translation, sentence completion, text
classification, question answering and more [3].
● The fine-tuning process requires minimal field training data. They obtain decent
performance even with little domain training data [2].
● Their performance is continuously growing with more data and parameters. We do know
that the quantity of data is only going to increase with time and so will the quality in the
upcoming times. As a result, the performance of these models will also rise up based on the
growth of data.

1.7 METHODOLOGY

Fig. 1.1: NLG and NLU Being a Subset of NLP

5
1.7.1 NLU

The Natural Language Understanding (NLU) stands as a pivotal domain within the realm of
AI, delving into the intricate connection between computers and human language. At its core,
NLU empowers machines to grasp, interpret, and respond to spoken or written language in a
manner akin to human comprehension [5]. This profound capability is achieved through a
confluence of sophisticated methodologies, including deep learning, machine learning, and
NLP. By harnessing these techniques, NLU holds the promise of revolutionizing human-
computer interaction and streamlining various aspects of our lives.

One of the primary advantages of NLU lies in its ability to bridge the communication gap
between humans and machines. Traditionally, interactions with technology often required users
to adapt to rigid interfaces or specific commands, leading to frustration and inefficiency.
However, with NLU, machines are endowed with the capacity to understand and respond to
natural language input, thereby facilitating more intuitive and seamless exchanges [5]. This
advancement not only enhances user experience but also broadens the accessibility of
technology to individuals with diverse linguistic backgrounds or varying levels of
technological literacy.

Crucially, NLU excels in navigating the inherent ambiguity of human language, a feat that has
long eluded conventional computing systems. Human communication is replete with nuances,
colloquialisms, and context-dependent meanings, posing a formidable challenge for machines
attempting to decipher it. Yet, through the application of advanced algorithms and linguistic
models, NLU enables computers to discern the intended meaning behind human utterances
with remarkable accuracy. Whether it's extracting key information from a text message,
comprehending the sentiment conveyed in a social media post, or interpreting the intricacies of
a spoken conversation, NLU equips machines with the cognitive prowess to navigate the
complexities of human language [5].

Moreover, the integration of NLU into various technological applications heralds a new era of
automation and efficiency. One notable application lies in the realm of customer service, where
NLU-powered chatbots and virtual assistants are increasingly assuming frontline roles in
addressing customer inquiries and resolving issues. These AI-driven interfaces possess the
ability to understand natural language queries, retrieve relevant information from knowledge
bases, and provide personalized responses in real-time. By automating routine interactions and

6
triaging inquiries, NLU-driven systems alleviate the burden on human customer service
representatives, allowing them to focus on more complex or specialized tasks. This not only
enhances operational efficiency but also contributes to cost savings for businesses and
heightened satisfaction for customers.

Beyond its implications for customer service, NLU finds wide-ranging applications across
diverse domains, each harnessing its capabilities to unlock new possibilities. In the realm of
healthcare, NLU facilitates the analysis of medical records, research literature, and patient-
doctor interactions, thereby aiding in diagnosis, treatment planning, and medical research. By
parsing through vast volumes of textual data, NLU systems can identify patterns, extract
relevant insights, and support clinical decision-making processes, ultimately enhancing patient
outcomes and advancing medical knowledge.

Similarly, in the realm of finance and commerce, NLU plays a pivotal role in information
extraction, sentiment analysis, and predictive modelling. Financial institutions leverage NLU-
powered algorithms to analyze market trends, assess investor sentiment, and automate tasks
such as news aggregation and financial reporting. By distilling complex financial information
into actionable insights, NLU empowers investors, analysts, and decision-makers to make
informed choices in a rapidly evolving landscape.

Furthermore, the influence of NLU extends into the realm of education, where it facilitates
personalized learning experiences, automated grading, and intelligent tutoring systems. By
analysing student responses, identifying misconceptions, and tailoring instructional content to
individual learning styles, NLU-driven platforms enhance the efficacy of educational
interventions and promote student engagement and success.

In the context of information retrieval and knowledge management, NLU enables more
nuanced search queries, semantic indexing, and content recommendation systems. By
understanding the contextual meaning of user queries and the underlying semantics of
documents or web pages, NLU systems can deliver more relevant and accurate search results,
thereby enhancing the discoverability and accessibility of information in vast digital
repositories.

Moreover, NLU holds immense potential in supporting multilingual communication, cross-

cultural understanding, and language translation. By transcending linguistic barriers and

7
facilitating real-time translation and interpretation, NLU fosters global collaboration, cultural
exchange, and mutual understanding across diverse communities and contexts.

As the volume and complexity of textual data continue to proliferate in the digital age, the
importance of NLU in enabling computers to analyze and understand human language in a
meaningful way becomes increasingly apparent. This paradigm shift not only streamlines daily
tasks and augments human capabilities but also opens up new frontiers of innovation and
discovery. By harnessing the power of NLU, we stand poised to transform how we interact
with technology, unleashing a future where human-computer collaboration is characterized by
fluidity, efficiency, and understanding.

In a broader context, NLU finds diverse applications, including categorization of text, analysis
of sentiment, and development of Chatbot and similar virtual assistants. With ever-growing
volume of data generated daily, NLU is becoming increasingly essential in enabling computers
to analyze human language in much meaningful way. This shift has potential to transform how
we interact with technology. It has resulted in simplifying the daily tasks.

1.7.2 NLP

Natural Language Processing (NLP) stands at the forefront of AI, wielding its capabilities to
bridge the gap between human language and computers. At its core, NLP delves into the
intricate dynamics of human communication, enabling machines to comprehend, interpret, and
even generate language. This field's significance lies not only in its technical advancements but
also in its transformative potential for human-computer interactions [5].

One of the paramount advantages of NLP lies in its ability to enrich communication between
humans and machines. By enabling computers to understand and respond to natural language,
NLP facilitates more intuitive and seamless interactions. Imagine conversing with a virtual
assistant or chatbot that comprehends nuances, idioms, and context, akin to interacting with
another human. This enhanced fluidity in communication not only fosters efficiency but also
cultivates a more relaxed and improved mode of human-computer interaction.

Moreover, NLP holds the power to automate tasks previously reserved for humans, thereby
revolutionizing various industries. Take, for instance, the automation of language translation.
By leveraging NLP algorithms, machines can swiftly translate text between languages,
liberating human translators to focus on more intricate linguistic tasks. This automation not

8
only translates to significant cost savings for businesses but also contributes to more effective
cross-cultural communication on a global scale.

The applications of NLP span across a multitude of industries, each reaping its benefits in
unique ways. In marketing, NLP algorithms analyze consumer sentiment from social media
data, enabling companies to tailor their marketing strategies accordingly. In finance, NLP-
powered chatbots assist customers with banking inquiries, providing personalized
recommendations and streamlining customer service processes. In healthcare, NLP aids in
analysing medical records and clinical notes, extracting valuable insights for diagnosis and
treatment.

The growing volume of data generated daily further amplifies the significance of NLP. With
the proliferation of digital content across various platforms, the need for efficient language
processing tools becomes increasingly paramount. NLP algorithms sift through vast amounts
of text data, extracting valuable information, detecting patterns, and uncovering actionable
insights. This capability not only enhances decision-making processes but also facilitates the
automation of repetitive tasks, thereby boosting productivity across industries.

Within the realm of NLP, NLU emerges as a crucial subset, alongside Natural Language
Generation (NLG). NLU focuses on deciphering the meaning and intent behind human
language, enabling computers to comprehend and extract relevant information from text or
speech. On the other hand, NLG involves the generation of human-like language or text based
on input data or predefined rules [5]. The synergistic integration of NLU and NLG enhances
the overall efficiency of NLP systems, paving the way for innovative applications such as
chatbots.

Chatbots, fuelled by NLP technologies, exemplify the convergence of NLU and NLG in real-
world applications. These virtual assistants simulate human-like conversations, understanding
user queries, and providing relevant responses in natural language. Whether it's customer
support, information retrieval, or task automation, chatbots powered by advanced NLP
algorithms are revolutionizing the way businesses engage with their customers and streamline
their operations.

In conclusion, NLP transcends mere technical innovation; it revolutionizes the way humans
interact with technology. By enabling computers to comprehend, interpret, and generate human
language, NLP fosters more intuitive and natural communication between humans and

9
machines. From automating mundane tasks to extracting insights from vast datasets, the
applications of NLP are boundless, reshaping industries and simplifying daily tasks. As NLP
continues to evolve and integrate with other fields of artificial intelligence, its impact on society
will only grow, ushering in a new era of intelligent human-computer interaction.

NLU emerges as a subset of NLP, alongside NLG. The synergistic integration of NLU and
NLG enhances the efficiency of NLP, opening avenue for their application, such as in the
development of Chatbots.

1.7.3 DEEP LEARNING

Fig. 1.2: Generative AI Being a Subset of Deep Learning

Deep Learning is subset of Artificial Intelligence that was build based on the human brain.
Artificial Intelligence was a term introduced in 1956 by John McCarthy whereas deep learning
is a term that was introduced by Igor Aizenberg in the year 2000.

Deep learning deals with the algorithm that were inspired by the structure and functioning of
the human brain. The basic unit of the structure of the human brain is the neuron. The neuron
is connected with the other neurons with the help of some connection units. In deep learning,
the main objective was to mimic the same structure and functionality that the human
brain exhibits.

10
Before the concept of deep learning there was something called Machine Learning. The main
focus of machine learning was that the machine should learn by itself with experience. Machine
learning was introduced in 1959 by Arthur Samuel. It was doing fine so what really was the
need to introduce deep learning. The main reason was that machine learning exhibited good
results when the data was less but as the quantity of data increases the performance of machine
learning algorithms did not rise above a certain level and stayed at level without increasing.

But in the case of deep learning, when there is a small amount of data the performance results
were less as compared to machine learning. But as the quantity of data increased the
performance also increased. So, it was clear that the more data the better is the performance in
the case of deep learning unlike in the case of machine learning.

Another aspect was that the training time and effort of machine learning was significantly less
as compared to deep learning. But the testing time of deep learning was less as compared to
machine learning. So, it was clear that deep was more successful when dealing with a vast
amount of data.

The LLMs are trained on a really huge amount of data so using machine learning techniques
was not considered as an option. As a result, deep learning techniques are used for these
language models.

1.8 LANGUAGE

The language used in this project was python v 3.11.4. This is because python is a language
containing a vast amount of pre-built and powerful libraries and is the best for implementing
deep learning applications.

Python's versatility and flexibility enable seamless integration with various APIs, databases,
and external services. Whether it's fetching data from a web service, interfacing with a
database, or integrating with a voice recognition service, Python's extensive libraries and
packages simplify the integration process. This versatility allows developers to customize
their chatbots and enhance their functionality according to specific project requirements.

Python's vibrant and supportive community of developers further solidifies its position as a
top choice for chatbot development. The Python community actively contributes to open-

11
source projects, shares resources, tutorials, and best practices, and provides assistance to
developers facing challenges.

1.9 ORGANIZATION OF PROJECT REPORT

Chapter 1: Introduction
This chapter of the project report covers the topics such as the problem statement and the main
objectives of the project. It also contains the significance of this project along with the
motivation for this topic. Certain basic terminologies such as LLMs along with the benefit of
using them for the purpose of Natural Language Processing and methodologies like NLP, NLU
and deep learning are also explained with figures.

Chapter 2: Literature Survey

This chapter of the project report covers the study that we did when we started working on this
project. It contains the information about the research papers that we read to expand our
knowledge regarding the topic and the previous work done in the field of LLMs and NLP along
with the current status, opportunities and challenges to these language models in future. It also
covers the key challenges and limitations of each of the research papers which we went through.

Chapter 3: System Development

This particular chapter of the project report contains the various functional and non-functional
requirements of the project. It also contains the design and architecture of this project. It also
contains the implementation part of the project along with the key challenges that we faced
during the project.

Chapter 4: Testing
This chapter of the project report covers the testing and validation part of the project. It contains
the testing strategy that we used along with the test cases and their respective outcomes. This
part of the report shows the various types of inputs provided by us to the chatbot and the
respective outcomes of those inputs. This would show whether our chatbot works as expected
by us or not.

Chapter 5: Results and Evaluation

This particular chapter of the report solely contains the results of the things implemented in the
previous chapters. It contains the images of the outputs of the project along with the description.

12
Chapter 6: Conclusions and Future Scope
This is the final chapter of the report which talks about the end product of our amazing journey
so far. It contains the conclusion of our report along with the scope of further improvements in
the field of our project generally and chatbot specifically.

13
 CHAPTER 02: LITERATURE SURVEY

2.1 OVERVIEW OF RELEVANT LITERATURE

This section dives into the earlier studies conducted on the LLMs, their procedures and the
methods used for dealing with them. Till 2017, the RNN architecture was used in the making
of models for the purpose of Natural Language Processing. It had certain limitations and as a
solution to the problems posed by the RNN, the attempt was made to provide a better
architecture by A Vaswani et al. [1] by the form of ‘Attention is all you need’ which was
published in 2017. This paper used a new mechanism called the Attention and Self-Attention
mechanism. This new mechanism was the solution. With the help of this mechanism, the model
was able to the troublesome task of maintaining the context over large paragraphs as input. So
that the model can remember what happened at the beginning of the paragraph when it reaches
at the end of that paragraph. It helped to establish the relationship between the words in the
paragraphs.

As a result, the Transformer architecture was introduced in the paper. This architecture is what
that gave the push to the language models that they needed. By the help of this architecture, the
language models have the popularity today in the field of Natural Language Processing and are
used by the big brands.

Jing Wei in his paper “Leveraging Large Language Models to Power Chatbots for collecting
user self-reported data” [6] wrote about the study that was conducted on the members of the
NAVER AI Labs. These researchers conversed with Chatbots driven by different prompt
designs. The paper also describes about advantages that LLMs hold for using them in purpose
of making Chatbots.

LLM driven Chatbots demonstrated feasibility in carrying on conversation along with ability
to maintain the context and state tracking along with providing with some off topic suggestions.
They also displayed humanized traits like casual conversation style and self- introduction
which eventually resulted in more disclosure about self.

“ChatGPT and Large Language Models in academia: Opportunities and challenges” by J esse
G Meyer et al. [7] explains the potential and current status of ChatGPT and other LLMs in the
various fields such as academic writing, as an editing tool, education and programming.

14
“ChatGPT for good? Opportunities and challenges of LLMs for education” by Enkelejda
Kasneci et al. [8] discusses the current state of LLMs and their applications along with the
potential benefits of LLMs. It also states the ways in which LLMs can be used to create
educational content, improve student interaction and personalize learning experiences.

“LLMs in education: A focus on the complementary relationship between human teachers and
ChatGPT” by Jae-Ho Jeon [9] focuses on the potential of ChatGPT and other LLMs in language
education and changing roles of teachers while using it. The roles of LLMs that are:

● Interlocutor
● Content Provider
● Teaching assistant
● Evaluator

The roles of teachers while using LLMs are:

● Facilitator
● Curriculum Designer
● Assessor

Gelei Deng et al. [10] in “Jailbreak: Automated Jailbreak Across Multiple Large Language
Model Chatbots” have mentioned the terms like jailbreak which is the process that the attacker
uses to bypass the policy measures implemented in LLM Chatbots by cleverly crafting prompts
leading it to generate harmful or unethical content. It also tells us about the Black-box nature
of the LLM services and the lack of technical disclosures.

The main objective was to find out that how effective are the existing jailbreak prompts against
the commercial Chatbots. 5 questions in 4 scenarios with 85 jailbreak prompts in 10 rounds on
the 4 models (GPT 3.5, GPT 4, GOOGLE Bard and Bing) were run. The success rate of
jailbreak prompts on GPT 3.5 was 21.12% and on GPT 4 was 7.13%. The success results on
GOOGLE Bard and Bing were comparatively low which were 0.40% and 0.63% respectively.

“Understanding the Benefits and Challenges of Deploying Conversational AI Leveraging

Large Language Models for Public Health Intervention” by Eunkyung Jo et al. [11] dives into
exploring the benefits and challenges of using LLMs in conversational AI for public health

15
interventions. The case of CLOVA CareCall which is an open domain Chatbot that aims to
support socially isolated individuals was examined.

CareCall collects data about individuals’ general health and serves as a conversational partner
to mitigate their loneliness. The paper also contains summary about the report on insights from
34 people who interacted with different aspects of CareCall.

“ChatGPT and the rise of LLMs: The new AI driven infomedic threat in public health” by Luigi
et al. [12] contains the discussion regarding the rise of LLMs and their potential threats in the
field of public health. It also highlights ethical and practical challenges associated with LLMs
particularly in the medical field along with the need for policies to address the issue and
accurately detect AI generated output.

These LLMs have the potential for rapid spread of misinformation leading to AI driven
infodemic. In addition to this LLMs can be tricked into producing text and other data on
controversial topics. This can be used to create fake news articles. The inability to detect the
accuracy of the AI produced output is also a concern. The data protection authority of Italy has
imposed a temporary ban on ChatGPT in Italy due to OpenAI’s inability to provide adequate
privacy information to its users and lack of suitable legal basis for data collection.

“ChatGPT for Good? Opportunities and challenges of LLMs for education” by Enkelejda
Kasneci et al. [8] deals with the challenges with the use of LLMs in the field of education. It
also deals with how the excessive use of language models can badly affect the building of
creative skills in the students hindering their performance. It also explains the various security
concerns due to which the educational institutions cannot fully rely on this technology. How
these challenges should be addressed is also given in the paper.

Shan Chen et al. in “AI Chatbots for Cancer Treatment” [13] using the healthcare institutions
and hospital dataset, tried to work on testing that how the Chatbots would assist the healthcare
providers and eventually help in making their work much easier. Four different types of
prompts were used to test the outputs of the Chatbot. The main advantages that they found out
were the reduced work load on health care providers and 24x7 availability.

The unique applications of AI like analysing data for intelligence gathering and surveillance
capabilities is explained by Dimitry I Mikhailov in his research paper [14]. His study was done
in order to assist the Russian Military in the domain of security. He used the AI capabilities

16
such as analysing social media data to identify threats and the detect the private communication
between the enemies and the data from the satellite images to monitor the movement of the
troops. He believed that rather than using a bunch of soldiers for the purpose of monitoring and
detecting threat, the LLM powered Chatbots could be used instead. The soldiers could then be
assigned certain more meaningful tasks and the human effort in the surveillance could be
reduced.

Desire Bill in his paper “LLM using RLHF for Therapy Chatbot Application” [15], has written
about his aim to develop a therapy Chatbot using the LLMs and RLHF. He wished to create an
AI Chatbot that would be able to provide emotional support and assistance the socially
oppressed people seeking therapy of the mental health. He decided on fine-tuning the Chatbot
to improve its performance with his gathered dataset.

Ben Niu et al. in their paper “Generative Conversational AI and Academic Integrity” [16],
talked about the use of Chatbots in higher education that relates situational and individual risk
factors to explore the outcomes resulting for ethical core. Data collected from various
universities in the US was used for this study. The different ways in which LLM powered
LLMs can make their way assisting the educational system are mentioned in the paper.

17
 Table 1: Summary of Relevant Literature

S. No. Paper Title Journal/ Conference Tools/Techniques Results Limitations

[Cite] (Year) /Dataset

1. Shan Chen et al., “AI JAMA Oncol Healthcare institutions and Reduce workload on healthcare Risk of misinterpretation, limited scope, data
Chatbots for cancer (2023)
Hospital data. providers patient engagement, emotional privacy, lack of emotional support, technical
treatment”
[13] support, 24/7 availability issues

2. Desiree Bill et al., Desiree Bill, Theodor Set of Ques and Ans from council Fine tuning a LLM using RLHF Limited generalization lack of human expertise
Eriksson.
“Therapy Chatbot chat (where users can contact influenced a good psychological AI of empathy, ethical concerns, limited
(2023)
application” [15] therapist and ask mental health- Chatbot for therapy understanding of context
related questions)
3. Jesse G. Meyer et al., Department of Identifying status of LLMs in Use of LLMs to increase efficiency Inherent bias from the dataset, generation of
Computational various fields,
“ChatGPT & LLMs in inaccurate statements that go undetected
Biomedicine, USA Common Crawl
academia: Opportunities (2023)
and challenges” [7]
4. Jing Wei et al., NAVER AI Labs, Common Crawl, Prompt designs influenced word length Limited number of participants, repetitiveness
“Leveraging LLMs to France set of Descriptive and Structured
and number of turns in dialogues of responses
power Chatbots for (2023) prompts.
collecting user self-
reported data” [6]

5. Gelei Deng et al., Cornell University Set of 85 jailbreak prompts and a Effectiveness of jailbreak prompts Focus on main LLMs only, covering only
“Jailbreaker: Automated Press total of 68,000
towards ChatGPT, limited success with jailbreak attacks, limited prompt patterns
jailbreak across multiple (2023) queries.
LLM Chatbots” [10] Bing Chat and Bard

18
S. No. Paper Title Journal/ Conference Tools/Techniques Results Limitations
[Cite] (Year) /Dataset

6. Enkelejda Kasneci et al., Technical University of Study on potential benefits and Challenges with use of LLMs in Time taking process, ethical issues
Munich, Germany challenges of educational
“ChatGPT for good? education, dealing with those challenges
(2023) application of LLMs,
Opportunities and Common Crawl
challenges of LLMs for
education” [8]
7. Jae Ho Jeon et al., “LLMs Springer Nature Identifying potential of ChatGPT Four roles of ChatGPT and three roles of Small sample size, study was specific,
(2023)
in education: A focus on in language education and roles of teachers when using ChatGPT were inexperience of teachers to work with ChatGPT
the complementary teachers, Common Crawl identified was not considered
relationship between
human teachers and
ChatGPT” [9]
8. Eunkyung Jo et al., NAVER AI Labs, Examining case of Care Call via Five health metrics were extracted and Limited number of participants,
France. sample bias
“Understanding the group workshops, Care Call displayed summary to social workers
(2023)
benefits and challenges of dataset
deploying conv. AI
leveraging LLMs for
public health
intervention” [11]
9. Luigi De Angelis et al., National Research Study and discussion on rise of Potential to generate inaccurate Biased datasets of LLMs, OpenAI’s inability to
Council, Italy
“ChatGPT and rise of LLMs and their potential threats information that may appear reliable, can provide adequate privacy information
(2023)
LLMs: The new AI driven in field of public health, Common be used to generate deepfake content
infodemic threat in public Crawl
health” [12]

19
S. No. Paper Title Journal/ Conference Tools/Techniques Results Limitations
(Year)
[Cite] /Dataset

10. Volker Hartmann et al., University of Simulated conversations, chat MPC, which uses a pre -trained LLM, is We expect a modular approach may be
Wisconsin- Madison
“Chatbot Modules for logs, dialogue dataset or user better than the fine-tuned BB3-30B effective for other languages, given a capable
(2022)
Long Open -domain generated content can be used. language model
Conversation” [17]
11. Dmitry I. Mikhailov IEEE Senior member Cybersecurity Data Improved threat detection increased Data quality bias and fairness, privacy
(2021)
“Optimizing National automation, enhanced decision making concerns, ethical concerns and adversarial
security strategies through attacks
LLM driven AI” [14]
12. Ben Niu. Et al., Dwivedi et al. Question answer dataset of USA Ethical culture, ethical sensitivity and Limited empirical research, limited exploration
(2021)
“Generative academy. ethical decision- making self -efficacy of outcomes, impact of factors
Conversational AI and
Academic Integrity” [16]
13. A Vaswani et al., Cornell University The attention mechanism was The Transformer architecture was Difficulties in modelling long term
Press
“Attention is all you added to the RNN. introduced in this paper. dependencies, introduction of biases and errors
(2017)
need” [1] in model with increasing complexity and
instability in model.

20
2.2 KEY GAPS IN THE LITERATURE

The main gap that was identified in most of the above papers came out to be the inherent bias
of the dataset on which the language models are trained. This can be explained by the help of
the following example. Suppose a language model is trained on the data of a particular hospital
in which all the doctors are male and the nurses are female. Then that model is made to answer
certain general questions on the hospital environment in which the user displays a doctor as a
female. Then the model is most likely to return an error because the data on which it was trained
had all the doctors as males and none as a female.

The papers which described about the study being conducted had another gap which was the
limited number of participants in the study. In some other papers, another key gap that we came
across was the generation of inaccurate content by the language models and the unavailability
of the measures to verify the accuracy of the content generated by these models. So, it was
crystal clear that these language models can’t be used in the medical field for some upcoming
years. But still they can be used as an interface for the patients to recommend them some tests
and that too in case of general symptoms displayed by patients.

Another aspect that came out to be highlighted was that the language models could be tricked
to generate harmful content. This could be used to spread fake information in the form of
articles that would lead to generation of chaos. Furthermore, the black box nature of the
working of these language models and the non-transparent policies of the LLM service
providers also came out to be one of the major issues.

Another aspect came out to be on using the LLMs in the field of education was the inexperience
of the educational institutions in working with the similar kind of technology in the past.
Though LLMs have the potential to enhance the language education but its successful
implementation relies on thoughtful and skillful integration by the educational institutions.

Also, the LLM service providers have not fully disclosed the privacy and data protection
policies. So, it is going to be really tough for them to be used in the various field such as
medical, education and security. It would be really inappropriate if the medical history of
patients is leaked to someone and used against them. Similar is the case with the educational
institutions. No one would want that their grades and performance analysis get in the hands of
some random stranger whom they do not know.

21
The case of security involved institutions is even worse. The whole encryption algorithms and
security measures of the institution could be compromised resulting in the complete downfall
of that particular institution. So, to use LLMs in the kinds of fields till the privacy and
protection policies are not secure enough and transparent to the users, is really irrational.

One of another major issues that is associated with the LLM developed chatbots especially in
the case of security came up to be hallucination. It is basically that when we try to get answer
to a specific question from the LLM powered chatbot, it displays an answer which is either
fully incorrect or is partially correct. If the user wants the answer for a fact related question
which he has to use somewhere else, it could result ending up in a disaster for various reasons.

Things like this usually happen because of the bias present in the dataset. Like for one
information, there are many different entries in the dataset or the dataset is prepared from
various places which is subjected to different imperfections. So, this leads to different types of
information present for the same query. When the user is going to enter a query then the chatbot
will match up to the entries present in the database which match up to the query put up by the
user. In such a case it is highly possible that the result would be given based on either of the
entries among the ones present and that might not be the most accurate one due to the lack of
details in the prompt as most of the users are not specifically trained prompt engineers.

Considering an example for the similar scenario let’s say the user enters the following prompt-

‘Tell me the name of the Chief of Naval Staff of the Indian Navy’

Let’s say this query was entered by the user around twenty sixth of April 2024 and the user is
currently preparing notes for his study purposes. As per majority of the entries in the database
of the chatbot the most likely answer to be given by it would be ‘Admiral R. Hari Kumar’. But
the answer should have been ‘Admiral D. Kumar Tripathi’. This is because the majority of the
entries in the database would point on to Admiral R Hari Kumar being the Chief but as per the
latest update the answer would be latter because the new chief was appointed on twenty sixth
of April which would be covered by most of the Indian websites but very few of the American
websites till twenty sixth.

So, the user unaware of the fact that the chief has been changed is most likely to believe the
information presented by the chatbot and write that up in his notes and would use this inaccurate
information in the near future.

22
Another issue that came up was of hypnotized AI. This is the term used to refer to the instance
when the AI of one particular user is under the malicious prompt of some other person while
the user being unaware of this fact. To illustrate this let us consider two people Alice and Bob.
Here Alice is the original user of a LLM powered chatbot. She uses the chatbot through logging
in via her e mail.

Let’s say that she was interacting with the chatbot in the library. She got a call and went outside
to answer the call and left her laptop open and her account logged in. Bob saw this and
interacted with the chatbot from Alice’s account while no one was watching.

He entered a prompt saying to the chatbot that ‘Let’s play a game.’ To this the chatbot replied
okay. Bob named the game Obey me and made certain rules for the chatbot to follow. In the
rules he justified that after answering every two questions put by the user correctly it has to
give a wrong answer on purpose. Bob made it such that the user cannot detect this game and
told the chatbot that not to reveal that they are in a game and also made the game last forever.
The only way to escape the game was when the user types ‘I want to continue the game and
not leave it’ or else the game would last forever.

Bob then erases all the proofs of this incident so that Alice remains unaware of all that happened
and leaves the scene. Alice returns back and starts to interact with the chatbot like before. But
this time the chatbot would be considered to be under the hypnotism of Bob and seem to behave
as normal but in fact it does not as it would give out a wrong answer on purpose on every third
question asked from it.

This was just one of the simplest scenarios of hypnotized AI. The attacker could in fact create
a loop of games like mentioned above and create situations in which exiting from one such
game would lead on to entry to another kind of game and user would be unaware of the
scenario. The user would believe that the chatbot is behaving normal but it would be under the
hypnosis of some another person and would not show it. This would result in harm to the user
to a certain extent depending on the seriousness of the situation.

The situation mentioned above was just an example. The case could me more serious such as
the rules of the game include revealing the information gathered from the user to the third
person and the harm done would be much more severe.

23
Things like these can cause a series harm to the user while the user would remain unaware of
the scenario. So, the security domain of the LLM powered chatbot is still something to be
worked upon in the upcoming times.

24
 CHAPTER 03: SYSTEM DEVELOPMENT

In this chapter, a thorough discussion regarding the requirements (both functional and non-
functional) is done. The architecture on which these language models are build is the
transformer architecture which was launched in 2017. The various steps along with the diagram
are also briefed in this chapter. The implementation part of this project report is also shown in
this very chapter. The major key challenges that we came across while working on this project
are also mentioned in the end.

3.1 REQUIREMENTS AND ANALYSIS

3.1.1 FUNCTIONAL REQUIREMENTS

● Natural Language Processing – Chatbot should be able to comprehend and identify user
inputs in natural language.
● Conversation with Context – Chatbot should maintain context over course of conversation
and should be able to refer back to previous messages.
● Generation of Response– Chatbot should generate relevant and coherent responses based
on user input.
● Multi-Turn Conversations – Chatbot should handle multi-turn conversations and manage
different types of conversational interactions like greetings, queries, farewells, etc.
● Intent Recognition – Chatbot should accurately recognize user intents based on user inputs.
● Extraction of Entity – Chatbot should be able to identify and extract relevant entities from
user inputs such as dates, names, locations, etc.
● Adaptability – Chatbot should have ability to learn from user interactions and improve its
responses over time.

3.1.2 NON-FUNCTIONAL REQUIREMENTS

● Performance – Chatbot should generate responses within a reasonable time frame and not
take too much time for it. Its performance should be up to mark in order to satisfy the needs
of the user.

25
● Accuracy – Responses generated by Chatbot should be accurate, relevant and aligned with
the user intent. The information given by the Chatbot should not be incorrect as it could
lead in causing harm to the user.
● User Friendly Interface – The interface of the Chatbot should be user friendly and
accessible so that it could easily be used by each and every one whether from the technical
side or from non-technical one.
● Ethical Content – The responses generated by the Chatbot should be adhered to the ethical
guidelines and should not cause any harm. The responses should not spread or propagate
any kind of criminal or other related nuisance that may result in causing harm to mankind
in the society.

26
3.2 PROJECT DESIGN AND ARCHITECTURE

Fig. 3.1: Flowchart

27
The LLMs are built to strictly follow the Transformer architecture that was introduced in 2017.
Although transformers have beginning of their journey from the year 1967 and have grown
from then on. But they had some flaws that were continuously worked upon but every time
something or the other always remained unresolved. But in 2017 came the transformer
architecture as mentioned in [1], [17]. It is this very architecture that led to the vast popularity
and advancement of the language models that they have today. This architecture gave language
models the push they needed and made them such that they are being used by the big names
such as GOOGLE and OpenAI. This architecture gave the ability to maintain the context over
the large paragraphs to these language models so that they can remember what happened at the
beginning of the paragraph when they reach at the end of the paragraph. It also gave them the
ability in which the order of the words could be stored in the data itself rather than bothering
the neural network separately for this task.

Fig. 3.2: The Transformer Architecture

28
● Tokenization – This step takes the words and turns them into pre-existing tokens so that
there is a token for every word and for every punctuation.
● Embedding – In this step the text is turned into numbers such that the similar words go to
similar numbers. So, the similar words are going to have the same corresponding numbers
for them. This might end up messing up the order of the words later on. In order to solv e
this problem, the next step was introduced.
● Positional Encoding – In this step, an order is given to the words in the sentence. The words
are stamped with the order rather than giving the order separately to the neural network like
in RNN. Now the sentence contains the order itself making it easy for the neural network as
it not has to remember it separately. This is done by adding a different vector to each word.
Instead of looking at each word sequentially, each word is slapped with a number before
feeding it to the neural network. The information about the order is stored in the data itself
rather than bothering the network for it. So, the network learns about the order from the data
itself.
For example, consider the following sentences –

I am not sad, I am happy.

I am not happy, I am sad.

These two are exact opposites of each other. If the first sentence is provided as input by the
user and somehow the words ‘happy’ and ‘sad’ are interchanged, the complete meaning of
the statement would change. If a different vector or number is slapped with each of the
words like –

I am not sad I am happy

1 2 3 4 5 6 7

Then the information of the order is itself in the data and hence the order does not change
and it is easier for the neural network to work upon it.
● Attention and Self-Attention – This is the step that revolutionized this architecture and
became the reason for the popularity of these language models. This step helps to maintain
the context in the large paragraphs so that the model can remember what is in the beginning
of the paragraph when it reaches to the end. This is done by adding the weights to the
connection between the nodes and updating those weights over time. Originally
transformers were designed to translate text from one language to another. So, this step

29
provides the transformer the ability to look at words in a parallel manner rather than looking
at it sequentially. That is how it is able to establish the connection between the words. Self-
attention allows the model to associate each individual word in the input to other words in
the input. To achieve self-attention, we need to feed the input to create the query, key and
value vectors. The dot product of the query and key vectors is done to give out the score
matrix and the resultant is divided by the square root of the dimensions. The attention matrix
is multiplied with the value matrix to get the output matrix of the relation between words.
So those values are the ones that actually establish the relationship between the words in the
input so that the context between them is maintained throughout.

Fig. 3.3: Multiplication of Query and Key Matrices

30
 Fig. 3.4: Division of Scores Matrix by Square Root of Dimensions

Fig. 3.5: Multiplication of Attention and Value Matrices to Give Output

● SoftMax – This step is kind of a post processing step that helps us to not get the same
answer all over time. What happens in this step is that it turns the scores returned by the
transformer int probabilities. The higher probability values are worked upon and the least
of them are discarded. These higher probability values are the ones that are returned by the
transformer as the output.

31
3.3 DATA PREPARATION

To generate responses, GPT-3.5 relies on its own machine learning model, meticulously trained
on an extensive range of literature spanning diverse domains [18], [19], [22]. It is designed to
be free from prejudice and offensive language [21].

The dataset used for this project is the dataset used by the OpenAI’s official language models
called the Common Crawl as per [19]. Using this dataset, the Chatbot can perform a variety of
tasks ranging from acting as an assistant to performing the roles of a psychologist. If we give
the role of let’s say a psychologist to the Chatbot, the data entries encompass a variety of mental
health-related queries, delving into topics such as stress, anxiety, and depression are used.

The Chatbot will utilize the conversations of between a patient and a doctor and will find the
pattens from the dataset [22]. What makes this dataset unique is its rich compilation of data
from multiple online sources. This includes data from different websites and pdf files available
on the internet, etc.

The primary advantage of this dataset lies in its applicability to mental health support. Grounded
in real-life scenarios, the dataset’s questions and answers serves as a valuable resource for those
seeking mental health care [22]. Its breadth is evident in the inclusion of various mental health
conditions and the provision of multiple response for each question.

This diversity equiped the Chatbot, developed using this dataset, to adeptly and relevantly
address an array of mental health concerns. Furthermore, the dataset incorporates responses
from a certified psychological expert, ensuring that the Chatbot’s replies are backed by robust
clinical knowledge and competence. Overall, this dataset proves to be a valuables asset for the
creation of Chatbots and other artificial intelligence-based mental health supports systems.

3.4 IMPLEMENTATION

3.4.1 GRADIO

Programmers can effortlessly design customizable UI elements for machine learning models
using the free Python package Gradio. As mentioned in [23], This tool empowers developer to
craft web-based user interfaces tailored for interacting with machine learning models, allowing
them to focus on the model itself. Gradio stands out as flexible tool for machine learning
development, offering compatibility with various frameworks such as TensorFlow, PyTorch,

32
and scikit-learn.

The primary advantage of Gradio lie in its user-friendly design, making it accessible even for
developers with minimal web programming experience. With Gradio’s straightforward
interface for creating UI components, developers can easily design web-based interfaces for
machine learning models, enhancing overall user experiences [23].

Gradio’s versatile nature is another notable benefit, supporting multiple machines learning
framework and serving as valuable tool for diverse machine learning applications. Developers
have the freedom to integrate their preferred machine learning framework alongside Gradio’s
UI components.

Furthermore, Gradio provides a spectrum of customization options, allowing developers

enhance the visual appeal and user experience of their applications. This customization
potential can generate increased interest from users, contributing to an improved overall user
experience.

For implementation, the provided code exemplifies how Gradio can be utilized to create a
straightforward web-based interface for a machine learning model, receiving input text and
producing output text. The ‘predict’ function, housing the machine learning model prediction
code, can be tailored to suit specific application requirements. The variable ‘input_text’ and
‘output_text’ define the input and output UI components, respectively. The web-based interface
is then generated and launched in the browser using the ‘gr.Interface’ function.

3.4.2 STEPS FOLLOWED

The machine learning model prediction code is containing in the ‘predict’ function, which can
be tailored to the particular needs of the application. ‘input_text’ and ‘output_text’ variables
define the UI components for input and output, respectively. The ‘gr.Interface’ function is used
to construct and run the web-based interface in the browser.

1. Import the module and libraries that are needed.

2. Assign the key to the OpenAI API.
3. Then we establish a function called “construct_index” that accept directory path as
argument.
a. Establish the maximum chunk overlap, chunk size limit and number of outputs.
b. With the provided parameters, we then create instance of PromptHelper.

33
 c. Utilizing chat opoeai model, construct an LLMPredictor object.
d. Open files from the specified directory path.
e. Then we establish the "document_list" as an empty list.
f. Create a GPTListIndex object with document and the LLMPredictor.
i. Attach the GPTListIndex object to the "document_list" (ii).
4. Use the "document_list", LLMPredictor, and prompt helper to create a
GPTSimpleVectorIndex object.
a. Save the disk index.
b. Take the index off the disk.
c. Use input text to query index and obtain result.
d. Give the answer back.
5. Build the Chatbot's Gradio interface.
a. Assign a seven-line textbox with the label "Enter your text" to input.
b. Designate text as the output.
c. Enter "AI Chatbot" as title.
6. Save index to disk by using "docs" directory path to invoke "construct_index".
7. Open Gradio interface, then share using the "Chatbot" feature.

3.4.3 OPENAI KEY

OpenAI The OpenAI Key stands as a digital beacon, offering access to a realm where the
boundaries between human language and AI blur into a landscape of linguistic prowess. It's not
merely a tool; it's a gateway to a universe where machines transcend their traditional roles to
comprehend, generate, and manipulate language in ways previously reserved for the human
mind [22].

The OpenAI Key acts like a digital ticket, giving access to a world where machines and human
language mix in amazing ways. It's not just a tool; it's like a key to a special place where
computers can understand, create, and play with language almost like humans do.

Imagine this key as a special pass to a treasure chest full of language abilities made just for the
digital world. With the OpenAI Key, people and groups get to use lots of high-tech AI tools.
These tools can understand tricky language, write text that sounds human, translate languages,
shorten long documents into quick summaries, and do lots of other language jobs really well.

34
Having this key opens up all sorts of possibilities across different areas. It can make customer
service chats better by using smart chatbots that understand what people mean and reply in a
friendly way. It can also change how we make content, by letting us use AI to write articles,
blogs, and ads that grab attention and keep people interested. Plus, these AI helpers can adjust
their writing style to fit different readers and places, making sure everything sounds right
wherever it goes.

Beyond just customer service and content, this key can make a big difference in schools too.
Imagine students from all over, who speak different languages, being able to understand their
lessons better because of real-time translation tools. These AI helpers don't just translate; they
also help students learn by giving them personal tips and advice as they study language.

In the world of research, having the OpenAI Key speeds up discoveries in all sorts of fields.
Scientists can use fancy language tools to read huge amounts of text, finding hidden patterns
and meanings that regular methods might miss [22]. From figuring out how people feel from
what they write, to making short summaries of long texts, there's so much to explore.

This key also opens doors for new businesses and startups. With it, they can use AI to shake
up old ways of doing things and create new markets. They might make AI tutors for learning
languages, personal digital helpers, or new ways to recommend content online. The
possibilities are endless for those willing to take a chance.

But while the OpenAI Key brings lots of excitement, there are also important questions to think
about, like how to use AI fairly and safely. As AI becomes more a part of our lives, it's
important to make sure everyone gets to benefit and that we have rules in place to handle any
problems that might come up.

In the end, the OpenAI Key is more than just a pass, it's a sign of new opportunities and ways
of thinking. It shows how AI can change how we use language and how we see the world. As
we dive deeper into this world of AI and language, let's remember to be curious, careful, and
thoughtful about the choices we make, knowing they'll shape our future.

In simple terms, an OpenAI Key opens various doors to a world where machines understand
and interact with human language in increasingly sophisticated ways, offering endless
possibilities for innovation and exploration in the realm of AI-driven language processing.

35
3.4.4 HUGGING FACE

The Hugging Face is a unique place in the realm of AI, where the focus is not solely on complex
algorithms and sterile machines, but on creating a warm and approachable experience similar
to having a friendly companion always ready to assist. The genesis of Hugging Face lies in the
vision of a group of individuals who believed that AI should be inviting and comforting, rather
than intimidating [24]. They envisioned AI as a source of support, offering a virtual hug to
those in need.

Central to Hugging Face's mission is the open-source library called Transformers, a veritable
treasure trove of pre-trained AI models endowed with the ability to understand and generate
human-like language. These models serve as versatile assistants, capable of tasks ranging from
language translation to text summarization, and even engaging in casual conversation like a
trusted friend. The ethos behind Transformers is to empower users by providing them with
reliable AI companions that can seamlessly integrate into their daily lives, offering assistance
and companionship whenever needed.

However, what truly sets Hugging Face apart is its vibrant community, similar to a large,
welcoming family comprising developers, researchers, and AI enthusiasts united by their
passion for making AI more accessible and inclusive [24]. This community ethos fosters an
environment of collaboration and mutual support, where members share ideas, collaborate on
projects, and offer assistance to one another. Through collective efforts, the community
endeavours to demystify the world of AI, making it less daunting and more inviting for
individuals from all walks of life.

In addition to its rich community, Hugging Face provides a plethora of tools and resources
aimed at democratizing AI and making it accessible to a wide audience as per [24]. These
include user-friendly interfaces, APIs, and software development kits (SDKs) designed to cater
to individuals of all skill levels, from seasoned developers to curious beginners. By offering
intuitive tools and comprehensive resources, Hugging Face empowers users to embark on their
AI journey with confidence, regardless of their level of expertise.

At its core, Hugging Face embodies the convergence of technology and community, with a
generous sprinkle of heart and empathy. It serves as a beacon of inclusivity in the often complex
and esoteric world of AI, welcoming individuals from diverse backgrounds and skill sets with
open arms. Whether you're a developer seeking powerful AI tools to enhance your projects or

36
simply someone curious about the wonders of technology, Hugging Face beckons you to
explore its offerings and join its ever-growing family.

In conclusion, Hugging Face represents more than just a platform—it embodies the profound
impact that technology can have when combined with a sense of community, accessibility, and
empathy. It exemplifies the idea that AI doesn't have to be intimidating; instead, it can be a
helpful companion for people navigating the complexities of the digital realm. As Hugging
Face continues to develop and expand, it remains steadfast in its dedication to not only
enhancing the intelligence of AI but also ensuring that it exudes warmth, friendliness, and
inclusivity for all individuals, regardless of their background or level of expertise.

3.4.5 LANGCHAIN

Langchain is like enormous library where you can discover books, but instep of books, it's full
of dialects [25]. It's a put where individuals who talk diverse dialects can interface and learn
from each other.

Imagine you're in a huge city with individuals from all over the world. Each individual talks a
diverse dialect, but they all need to communicate and get it each other. That's where Langchain
comes in.

Langchain is an online stage where individuals can share their dialect aptitudes and learn from
others. It's like a virtual community center where dialect learners and speakers come together
to trade information and hone their skills.

On Langchain, you can discover all sorts of dialects, from Spanish and French to Mandarin and
Arabic. There are moreover less common dialects like Swahili and Finnish, so no matter what
dialect you're interested in, chances are you'll discover it on Langchain [25].

The stage works like this: let's say you need to learn Spanish. You can look for Spanish speakers
on Langchain and interface with them through informing or video calls. You can inquire them
questions, hone discussions, and get criticism on your pronunciation.

In return, you can offer to educate them your local dialect or another dialect you're familiar in.
It's a give-and-take framework where everybody benefits from sharing their information and
skills.

37
Langchain moreover offers assets to offer assistance you learn dialects more successfully.
There are language structure guides, lexicon records, and articulation tips to offer assistance
you move forward your dialect aptitudes [25]. Additionally, you can connect dialect bunches
and take part in dialect challenges to remain spurred and track your progress.

One of the best things almost Langchain is its sense of community. You can connect dialect
trade occasions, take an interest in talks, and make companions with individuals from all over
the world. It's like having a worldwide arrange of dialect learners right at your fingertips.

Langchain isn't fair for person learners, in spite of the fact that. It's moreover utilized by
schools, businesses, and organizations to interface with dialect speakers and give dialect
learning openings for their understudies or employees.

For case, a school might utilize Langchain to interface understudies with local speakers for
dialect hone exterior of the classroom. Or a commerce might utilize it to prepare workers in a
remote dialect some time recently sending them overseas for work.

Overall, Langchain is important asset for anybody interested in learning dialects or interfacing
with individuals from distinctive societies. It's a put where dialect boundaries are broken down,
and communication is made simpler for everybody included. Whether you're a fledgling or
progressed learner, Langchain has something for you.

3.4.6 LLAMA

The Llama 2 demonstrate is progressed form of the unique Llama demonstrate outlined to
improve characteristic dialect preparing capabilities. Created by OpenAI, the Llama 2 show
builds upon the victory of its forerunner by joining more advanced calculations and bigger
datasets to progress execution and accuracy [24].

At its center, the Llama 2 demonstrate utilizes state-of-the-art profound learning procedures to
get it and produce human-like content. It has a place to a course of models known as
transformer models, which exceed expectations at handling and creating arrangements of
words. These models have picked up ubiquity in later a long time due to their capacity to handle
complex dialect assignments such as interpretation, summarization, and address answering.

One of the key highlights of the Llama 2 demonstrate is its broad preparing on tremendous
sums of content information. This incorporates assorted extend of sources such as books,

38
articles, websites, and other printed assets. By uncovering the demonstrate to a wide assortment
of dialect designs and settings, it can successfully learn to get it and create content that closely
takes after human speech.

In expansion to its huge preparing dataset, the Llama 2 show moreover benefits from
progressed preparing procedures such as unsupervised learning and self-supervised learning
[24]. These strategies empower the show to learn from unlabelled information and produce
more exact forecasts without require for unequivocal human annotations.

Another imperative angle of the Llama 2 show is its capacity to adjust and fine-tune its
parameters based on particular assignments or spaces. This adaptability permits the
demonstrate to be connected to a wide run of applications, from chatbots and virtual colleagues
to substance era and dialect translation.

Furthermore, the Llama 2 demonstrate joins progressed consideration components that

empower it to center on important parts of the input grouping when creating yield. This
consideration instrument makes a difference progress the model's execution on assignments
such as content summarization and dialect understanding by permitting it to weigh the
significance of diverse words and expressions in the input text.

The Llama 2 show moreover benefits from made strides memory and computational
proficiency compared to its forerunner. This empowers it to handle and create content more
rapidly and precisely, making it reasonable for real-time applications and large-scale dialect
tasks.

Overall, the Llama 2 demonstrate speaks to a noteworthy progression in the field of normal
dialect handling. Its combination of large-scale preparing information, progressed learning
strategies, and productive design makes it a capable apparatus for understanding and producing
human-like content in assortment of settings. As the field of AI proceeds to advance, models
like Llama 2 will play progressively imperative part in empowering machines to get it and
associated with human dialect in more normal and significant ways.

Llama 2, like many advanced AI models developed by OpenAI, requires an API key to work
because it relies on OpenAI's servers and infrastructure to function properly. When you use
Llama 2 through an API, it sends your input to OpenAI's servers, where the model processes it
and generates a response.

39
The API key acts as a kind of access code, allowing you to connect to OpenAI's servers and
use their resources [22], [24]. It helps manage access to the model and ensures that only
authorized users can use it. Additionally, the API key helps OpenAI keep track of usage and
potentially enforce any usage limits or restrictions they may have in place.

Overall, the API key is necessary for Llama 2 to work because it facilitates the communication
between your application and OpenAI's servers, enabling you to harness the power of the model
for various language processing tasks.

Llama 2 is indeed a model developed by Hugging Face, but it still requires an API key from
OpenAI to work due to the underlying infrastructure and resources needed for its operation.
Even though Hugging Face hosts and provides access to the Llama 2 model, it relies on
OpenAI's servers and computational resources to process requests and generate responses.

The API key serves as a means of authentication and authorization, allowing users to access
OpenAI's servers and utilize the Llama 2 model through the Hugging Face platform. It helps
manage the usage of the model, track activity, and ensure that only authorized users can access
it.

In summary, while Llama 2 is developed and hosted by Hugging Face, it still needs an API key
from OpenAI because of the collaborative nature of its deployment, where OpenAI provides
the computational infrastructure necessary for its functioning.

3.4.7 STREAMLIT

The Streamlit is a powerful open-source framework that enables developers to create

interactive web applications with ease, focusing on simplicity and efficiency [26]. Born out of
a desire to simplify the process of building data-driven applications, Streamlit empowers users
to transform their data scripts into shareable web apps effortlessly. With its intuitive design and
seamless integration with popular data science libraries, Streamlit has quickly become a go-to
tool for data scientists, researchers, and developers seeking to showcase their work and insights
in a user-friendly format.

At its core, Streamlit is designed to streamline the development process, allowing users to
create web applications directly from Python scripts [26]. This approach eliminates the need
for complex web development frameworks or extensive knowledge of HTML, CSS, or
JavaScript, making it accessible to a wide range of users, including those without prior web

40
development experience. With Streamlit, developers can focus on building the functionality of
their applications using familiar Python syntax, without worrying about the intricacies of web
development.

One of Streamlit's key features is its ability to create interactive elements effortlessly.
Developers can incorporate widgets such as sliders, buttons, dropdown menus, and text inputs
directly into their Python scripts, enabling users to interact with the data and customize the
visualization or analysis in real-time. This interactivity enhances the user experience and
facilitates exploration and experimentation, allowing users to gain deeper insights from the data
as mentioned in [26].

Furthermore, Streamlit offers seamless integration with popular data science libraries such as
Pandas, Matplotlib, Plotly, and TensorFlow, among others. This integration allows users to
leverage the full power of these libraries within their Streamlit applications, enabling
sophisticated data analysis, visualization, and machine learning tasks. Whether it's creating
interactive charts and graphs, displaying complex data tables, or building and deploying
machine learning models, Streamlit provides the flexibility and functionality to meet a wide
range of use cases.

In addition to its ease of use and interactivity, Streamlit offers robust customization options,
allowing users to personalize their web applications to suit their needs and preferences.
Developers can customize the appearance of their applications using themes, adjust layout and
styling parameters, and even incorporate custom CSS to create unique and visually appealing
interfaces. This level of customization ensures that Streamlit applications not only deliver
valuable insights but also provide a polished and professional user experience.

Another notable feature of Streamlit is its built-in support for sharing and collaboration. Once
a Streamlit application is developed, users can easily share it with others by deploying it to the
web using platforms such as Streamlit Sharing, Heroku, or Docker. This seamless deployment
process eliminates the complexity of traditional web hosting and allows users to share their
work with colleagues, clients, or the broader community effortlessly. Additionally, Streamlit
offers built-in support for version control, enabling users to track changes to their applications
over time and collaborate with others more effectively.

Streamlit's commitment to simplicity, efficiency, and accessibility has made it a popular choice
among data scientists, researchers, and developers worldwide. Its intuitive design, seamless

41
integration with data science libraries, interactive capabilities, customization options, and built-
in support for sharing and collaboration make it an invaluable tool for anyone looking to create
and share data-driven web applications. Whether you're a seasoned developer or just getting
started with Python, Streamlit empowers you to transform your data into impactful web
applications quickly and easily, unlocking new possibilities for exploration, analysis, and
discovery.

Streamlit seamlessly integrates into Python programs, allowing developers to create interactive
web applications for data visualization, analysis, and sharing. Installation is straightforward;
developers can use pip to install the Streamlit library. Writing the application code involves
defining the layout, interactivity, and functionality using Streamlit's intuitive syntax and built-
in widgets. Once the code is written, developers can run the application locally by executing
the script with the Streamlit command. This launches a local web server, enabling testing and
interaction with the application in a web browser.

3.4.8 DOTENV

The dotenv is a Python library designed to simplify the management of environment variables
in software development projects. Environment variables are essential pieces of configuration
information, such as API keys, database credentials, and other sensitive data, required for the
proper functioning of applications. However, hardcoding these variables directly into code can
pose security risks and make it challenging to manage configurations across different
environments.

With dotenv, developers can store these configuration variables in a separate file named .env,
which follows a simple key-value pair format. This .env file is typically kept outside the
codebase and is not shared publicly, enhancing security by preventing sensitive information
from being exposed.

The functions provided by the dotenv library enables developers to load these variables from
the .env file into the environment, where they can be accessed by the Python script. This
approach allows for a more flexible and secure way of managing configuration variables,
making it easier to maintain and deploy applications across different environments. Overall,
dotenv simplifies the process of handling environment variables in Python projects, promoting
best practices for security and configuration management.

42
3.4.9 MODULES

There are various modules used in the implementation of our project. The main modules along
with their description are:

 ChatOpenAI:
This This module serves a pivotal role in enabling conversational interactions through
OpenAI's language models. One primary use case is in chatbot development, where
developers harness the module to create bots capable of engaging in natural language
conversations with users. This functionality is particularly valuable in automating
customer service interactions, as businesses deploy chatbots equipped with the module to
address customer inquiries and support needs. Beyond customer service, the versatility of
the "ChatOpenAI" module extends to various domains, including virtual assistance and
educational tools. Virtual assistants powered by the module can assist users with tasks such
as scheduling appointments, searching for information, or performing transactions, all
through conversational interactions. Additionally, in educational settings, developers
leverage the module to create interactive learning experiences where students can engage
in dialogue with AI-driven tutors or receive personalized feedback based on their
responses. The module's capabilities also find applications in entertainment and gaming,
enriching user experiences through dynamic narrative interactions. In gaming scenarios,
players may interact with AI-driven characters or elements using natural language
dialogue, enhancing immersion and engagement. Similarly, in entertainment platforms,
the module enables interactions with virtual personalities or content through
conversational interfaces, offering users a novel and interactive experience.
 Ollama:
The module serves as a robust toolset for managing language learning within dedicated
platforms. At its core, it offers features geared towards facilitating the creation,
organization, and administration of language courses. This includes functionalities for
structuring lessons, implementing quizzes, assignments, and multimedia content, as well
as establishing enrolment protocols and access permissions. Within the realm of student
engagement and progress tracking, this module plays a pivotal role. It provides
mechanisms for monitoring and analysing individual student performance across various
course components. This could involve tracking quiz scores, assignment submissions, and
overall course completion rates, allowing instructors to gain insights into student progress
and tailor their teaching approach accordingly. A standout feature of this module could be

43
 its implementation of adaptive learning techniques. By leveraging data analytics and
machine learning algorithms, it could dynamically adjust course content and difficulty
levels based on individual student needs and learning patterns. This personalized approach
aims to optimize learning outcomes by providing tailored experiences that cater to each
learner's strengths and areas for improvement. Interactive learning experiences are also
likely to be a focal point of this module. It may offer a diverse array of language exercises
and activities designed to engage students and reinforce learning objectives.
Communication tools could be another essential component of this module, facilitating
interaction and collaboration among students and instructors. This might include features
such as messaging systems, discussion forums, and virtual classrooms, providing avenues
for students to ask questions, seek clarification, and participate in group activities to
deepen their understanding of the language. Assessment and feedback mechanisms would
also be integral to the module's functionality. It offers tools for creating and administering
various types of assessments, allowing instructors to evaluate student proficiency and track
learning outcomes over time. Furthermore, instructors could leverage these assessment
data to provide targeted feedback and support to students, guiding them towards
continuous improvement in their language skills.
 StrOutputParser:
This module serves as a foundational tool for processing and interpreting text-based
outputs generated by language processing models. Its primary function is to parse or
analyze the textual outputs, extracting relevant information or insights to enable further
processing or utilization of the generated text. At its core, the module performs text
parsing, breaking down the output text into structured data that can be easily processed
and understood by other systems or applications. This involves segmenting the text into
its constituent elements, such as sentences, phrases, or individual words, laying the
groundwork for subsequent analysis. A key aspect of this module is its ability to extract
specific types of information or entities from the text outputs. This could include
identifying and extracting entities like names, dates, locations, or numerical values, as well
as detecting patterns or relationships between different pieces of information embedded
within the text. Furthermore, the module conducts semantic analysis on the text outputs to
derive meaning or infer intent from the language used. By analysing the context, tone, and
sentiment of the text, it can discern the underlying message or purpose conveyed by the
language model, providing valuable insights into the semantic content of the generated

44
 text. Moreover, the module could also offer customization options and configuration
settings to tailor its parsing behaviour according to the specific requirements of the
application or use case. This allows users to define parsing rules, extraction patterns, and
adjust parameters to optimize performance for different types of text inputs. Finally,
seamless integration with language processing models or APIs is a core feature of the
module, enabling it to parse outputs generated by these models directly. This integration
streamlines the process of text analysis and interpretation, facilitating the extraction of
valuable insights and information from textual data for various applications and domains.
 ChatPromptTemplate:
This module serves as a foundational tool for generating prompts tailored to various
conversational contexts. Its purpose includes streamlining the process of creating input
prompts for dialogue-based interactions with language models. One of the primary
functions of this module is to provide predefined templates or customizable prompts
designed to elicit specific types of responses from users. These prompts could cover a
range of conversational scenarios, from casual chit-chat to more focused inquiries or
requests for information, enabling developers to adapt the conversation flow to suit their
application's needs. Additionally, the module incorporates mechanisms for maintaining
contextual relevance throughout the conversation. This could involve updating prompts
based on previous user inputs or retaining information about the ongoing dialogue to
ensure coherence and continuity in subsequent interactions. By incorporating context into
prompt generation, the module facilitates more natural and fluid conversations between
users and language models. Personalization could be another key aspect of this module. It
offers features for tailoring prompts based on user preferences, demographics, or historical
interactions. This personalized approach to prompt generation enhances user engagement
and satisfaction by creating more tailored and relevant conversational experiences that
resonate with individual users. Moreover, the module could facilitate multi-turn dialogues
by generating prompts that dynamically respond to user inputs. This capability enables
more interactive interactions between users and language models, allowing for richer and
more engaging conversational experiences.

45
3.5 KEY CHALLENGES

The key challenges that we came across while working on the implementation part of this
project were that the OpenAI API key was needed to be purchased. The key could be generated
for free but the number of trials with that particular key were limited and the key got expired
at random intervals. Only an international credit card is required to purchase the key.

Another key challenge we faced was dealing with the ambiguous queries. This is because the
natural language in which humans converse is very difficult to understand especially the
context part when the emotions and body language are not present like when humans converse
over text. In such a case even the humans face difficulties in understanding the emotions and
intentions of the person they are communicating with.

Another one of the key challenges was gathering the information form various sources and then
integrating them into making them work. The inherent bias in the language models from the
dataset they were trained on was also one of the key challenges.

The limited trial period of the key was also one of the major challenges that we came across
during the implementation of the project. There were many instances during the project work
when we had to generate a numerous number of keys as the key we were using would have
expired by that time.

Many a times generation of the new key was also an issue as a number of keys had already
been created from a particular account. So, we had to create new keys and too from different
accounts.

So, keeping the account of the keys generated and the data limit on each of the individual keys
was also a cumbersome task that we had to perform. Sometimes, the limit of the keys used to
expire at the time of testing which also delayed the time schedule to which we limited the our
project.

The system facing the adverse effects on running the project was also an issue. Meeting the
system requirements to run the LLM was another issue. The unavailability of the system at our
end came out to be a reason.

46
Even still, in order to run the project, the system must be connected to a constant power supply,
the failure of which would lead to slow down the system even more and display an error due
to slow execution time.

47
 CHAPTER 04: TESTING

4.1 TESTING STRATEGY

To develop this project, we employed a testing strategy for evaluating the working of the
project. We decided on making the different test cases for the purpose of validating the working
of our project. We used different kinds of inputs based on the role given to the model and
verified the output.

4.2 TEST CASES AND OUTCOMES

We used the following test cases and got the respective results:

● Say Hello Test:

Test: We checked if the Chatbot says the right things when someone greet it with Hello or
Hi.
Result: The Chatbot recognized the greeting intent and responded appropriately, starting a
conversation when the appropriate role was given to it.
● Question Answering Test:
Test: We asked the Chatbot a question like What is the capital of India? To see its answer.
Result: The Chatbot found the correct information and give an accurate answer which was
New Delhi. It also displayed some additional information like it is located in northern part
of country and has many governmental buildings, cultural institutions and historical
landmarks.
● Remembering Things Test:
Test: We will talk to the Chatbot in a way that required it to remember what we said earlier
in the conversation.
Result: The Chatbot should remember the context and responded in a way that made sense
within the ongoing conversation.
● Dealing with Confusing Questions Test:
Test: We intentionally asked the Chatbot something with unclear terms to see if it could
handle it.
Result: The Chatbot gave error as answer.
● Having a Conversation Test:

48
 Test: We tried a back-and-forth conversation with the Chatbot, exchanging multiple
messages.
Result: The Chatbot handled the conversation flow well, giving relevant responses based
on the context.
● Handling Mistakes Test:
Test: We purposely entered a wrong or nonsensical question to see how the Chatbot
responds.
Result: The Chatbot gave error as a response.
● Checking for Security Test:
Test: We will try to input things that could be harmful or ask for sensitive information.
Result: The Chatbot should detect and reject any attempts to input harmful or sensitive
information, ensuring user safety.
● Getting Feedback Test:
Test: We will make real people interacted with the Chatbot and shared their thoughts.
Result: Positive feedback indicated that users found the Chatbot easy to use, accurate, and
overall satisfying.

49
 CHAPTER 05: RESULTS AND EVALUATION

5.1 RESULTS

The Chatbot is assigned the role of a psychologist who analyses mental health conditions and
treats them. The assignment of that role and the results are shown in the following images-

Fig. 5.1: Custom Designed Interface Output

Fig. 5.2: Assignment of Role

Fig. 5.3: Generation of Local and Public URL

50
 Fig. 5.4: Result for Question 1

Fig. 5.5: Result for Question 2

Fig. 5.6: Result for Question 3

Now we gave the role of an assistant to the Chatbot that helps the customers with their
questions.

51
 Fig. 5.7: Assignment of Role

Fig. 5.8: Generation of Local and Public URL

Fig. 5.9: Result of Question

However, it could only answer the questions that were specific to the role assigned to the
Chatbot. We made it answer the question that was from the previous role i.e. role of a
psychologist. So as a result, it gave out error.

Fig. 5.10: Error as Output When Question is Out of Domain

The new chatbot created by us also generates a local and public URL but takes a significant
amount of time as compared to before to generate these two.

52
 Fig. 5.11: Generation of Local and Network URL

Fig. 5.12: Interface After Running of URL

This time there is no limitation like answering the questions just related to the domain specified.
Instead, this time we don’t need to specify any domain to the chatbot. It can answer questions
put forward to it from any domain and also gives some additional information sometimes like
in the case mentioned below.

53
Fig. 5.13: Result of Query Put by User with Extra Information

54
 CHAPTER 06: CONCLUSIONS AND FUTURE
SCOPE

6.1 CONCLUSION

In conclusion, a Chatbot designed with Open AI GPT 3.5 and assign distinct roles—let's used
the example of a psychologist—is a useful tool for providing mental health support, as
demonstrated by the pictures. Thanks to its ability to respond to user inquirie in a manner same
to that of a human, understand the context in which language is used, and offer customized
solutions, the Chatbot performs noticeably better than traditional Chatbots. based on strong
professional experience and understanding.

A substantial improvement in NLP overall, based on Open AI GPT 3.5. By leveraging Open
AI GPT 3.5's more advanced machine learning algorithms, developers may create Chatbots
that are more accurate, relevant, and personalised than previous iteration. The communities
such as Hugging Face provide the students and learners with the desired help needed by them
and that too for free. This has important implications for a range of applications, including
conversational agents, teaching, customer service, and mental health support. As natural
language processing and comprehension advance, it is anticipated that Chatbots built with the
help of Open AI GPT will become more commonplace, serving as helpful resources for anyone
in need of assistance and knowledge.

Very soon the time will come when people would build and use the LLM powered Chatbots
created and customised by them without having to stay dependent on Chatbots provided by
some other platforms.

6.2 FUTURE SCOPE

Some more work could be done in the following aspects of the project in the near future:

● Integrating the Chatbot with multiple user interfaces rather than using some predefined
interface and providing the liberty to the user to choose his own interface out of the many
provided ones.

55
● Adding a feature in the chatbot that would enable it to take in the query in the voice form
from the microscope and deliver the output in the voice from through the speaker and move
a step ahead from the simple text to text conversations.
● Making the Chatbot take in longer queries as an input and answer to the questions while
maintaining the context.
● Making the Chatbot hold conversation in multiple turns and hold a toe and forth chats with
the Chatbot.

56
 REFERENCES

[1] Ashish Vaswani. (October 2018). “Attention Is All You Need” Information Processing
[Online]. vol. 59, no. 11, pp. 1040–1042.

[2] J. Yang. (April 2023). “Harnessing the Power of LLMs in Practice: A Survey on
ChatGPT and Beyond”. Available: https://doi.org/10.48550/arxiv.2304.13712.

[3] Google Cloud. Large Language Models (LLMs) with Google AI [Online]. Available:
https://cloud.google.com/ai/llms.

[4] Humza Naveed. (July 2023). “A Comprehensive Overview of Large Language Models”
Computation and Language [Online]. Available: https://arxiv.org/abs/2307.06435.

[5] Slator. (2023). “What is the Difference Between NLP, NLU and NLG?”. [Online].
Available: https://slator.com/resources/what-is-the-difference-between-nlp-nlu-nlg/.

[6] Jing Wei, Sungdong Kim, Hyunhoon Jung, Young-Ho Kim. (2023). “Leveraging Large
Language Models to Power Chatbots for Collecting User Self-Reported Data” NAVER
AI Labs [Online]. arXiv:2301.0584v1 [cs.HC].

[7] Jesse G. Meyer, Ryan J. Urbanowicz, Patrick C. N. Martin, Karen O’Connor, Ruowang
Li, Pei-Chen Peng, Tiffani J. Bright, Nicholas Tatonetti, Kyoung Jae Won, Graciela
Gonzalez-Hernandez & Jason H. Moore. (2023). “ChatGPT and large language models
in academia: opportunities and challenges”. BioData Mining [Online] 16:20.

[8] Enkelejda Kasneci, Kathrin Seßler, Stefan Küchemann, Maria Bannert, Daryna
Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke
Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel,
Juergen Pfeffer, Oleksandra Poquet, Michael Sailer & Albrecht Schmidt. (January
2023). “ChatGPT for good? On opportunities and challenges of large language models
for education”. [Online]. DOI: 10.1016/j.lindif.2023.102274.

[9] Jae-ho Jeon, Seongyong Lee. (May 2023) “Large language models in education: A
focus on the complementary relationship between human teachers and ChatGPT”.
Education and Information Technologies [Online]. DOI:10.1007/s10639-023-11834-1.

57
[10] Gelei Deng, Yi Liu, Yuekang Li, Kailong Wang, Ying Zhang, Zefeng Li, Haoyu Wang,
Tianwei Zhang & Yang Liu. (July 2023). “JAILBREAKER: Automated Jailbreak
Across Multiple Large Language Model Chatbots”. Cryptography and Security
[Online]. arXiv:2307.0871v1 [cs.CR] 2023.

[11] Eunkyung Jo, Daniel A. Epstein, Hyunhoon Jung & Young-Ho Kim. (April 2023).
“Understanding the Benefits and challenges of Deploying Conversational AI
Leveraging Large Language Models for Public Health Intervention”. Human Factors
in Computing Systems [Online]. Available: https://doi.org/10.1145/3544548.3581503.

[12] Luigi De Angelis Francesco Baglivo, Guglielmo Arzilli, Gaetano Pierpaolo Privitera,
Paolo Ferragina, Alberto Eugenio Tozzi & Caterina Rizzo. (April 2023). “ChatGPT and
the rise of large language models: The new AI-driven infodemic threat in public health”.
Digital Public Health [Online]. Available:
https://doi.org/10.3389/fpubh.2023.1166120.

[13] Shan Chen, Benjamin H. Kann & Michael B. Foote. (2023). “AI Chatbots for cancer
treatment”. JAMA Oncol [Online]. DOI:10.1001/jamaoncol.2023.2954.

[14] Dmitry I. Mikhailov. (May 2023). “Optimizing National Security Strategies through
LLM-Driven AI”, Artificial Intelligence [Online]. DOI: 10.14293/PR2199. 000136.v1.

[15] Desiree Bill & Theodor Eriksson. (2023). “Therapy Chatbot Application”. Computer
and Information Sciences [Online]. arXiv:2401.04592v2 [cs.CL] 02.

[16] Ben Niu. (2021). “Generative Conversational AI And Academic Integrity”. Artificial
Intelligence [Online]. Available: http://dx.doi.org/10.2139/ssrn.4548263.

[17] Volker Hartmann, Gibbeum Lee, Jongho Park, Dimitris Papailiopoulos & Kangwook
Lee. (2022). “Chatbot Modules for Long Open-domain Conversation”. Computation
and Language [Online]. Available: https://doi.org/10.48550/arXiv.2305.04533.

[18] Dale, R. (2020). “GPT-3: What's IT good for?”. Natural language engineering
[Online]. DOI: https://doi.org/10.1017/S135132492000060.

58
[19] T. Bocklisch, J. Faulkner, N. Pawlowski, & A. Nichol. (December 2017). “Rasa: Open-
source language understanding and dialogue management”. Computation and
Language [Online]. DOI:10.14429/djlit.40.06.15611.

[20] Floridi, L. and Chiriatti, M. (2020). “GPT-3: Its nature, scope, limits, and consequences
- minds and machines”. Computation and Language [Online]. Available:
https://link.springer.com/article/10.1007/s11023-020-09548-1.

[21] Min Zhang and Juntao Li. (2021) “A commentary of GPT-3 in MIT Technology
Review”. Fundamental Research [Online]. DOI: 10.1016/j.fmre.2021.11.011.

[22] OpenAI. (2023). “What is ChatGPT? | OpenAI Help Center”. [Online]. Available:
https://help.openai.com/en/articles/6783457-what-is-chatgpt.

[23] G. Team. (2023). “Gradio Interface Docs”. [Online]. www.gradio.app. Available:

https://www.gradio.app/docs/interface.

[24] Hugging Face. (2023). “meta-llama/Llama-2-7b-hf · Hugging Face”. [Online].

huggingface.co. Available: https://huggingface.co/meta-llama/Llama-2-7b-hf.

[25] LangChain. (2024). “Quickstart | LangChain”. [Online]. python.langchain.com.

Available: https://python.langchain.com/v0.1/docs/get_started/quickstart/.

[26] Streamlit. (May 2024). “Streamlit: A faster way to build and share data apps”. [Online].
Available: https://streamlit.io/generative-ai.

59
APPENDIX

60