International Conference on

Educational Science and Foreign Language Teaching, ICEF

Personalizing learning resource discovery through an AI chatbot-integrated

system

Ho Duc Loc1, Phan Minh Duy1, Le Van Nhuan1, Nguyen Thi Thanh Tu1*
1
Hanoi University of Science and Technology, Hanoi, Vietnam
*
Corresponding author email: [Link]@[Link]

Abstract
This study addresses the challenge of providing personalized access to digital educational resources by
developing Edvangarde, an AI chatbot-integrated online platform for educational content distribution. The
platform features an AI chatbot interface that leverages natural language processing to tailor content
recommendations to individual learners. A pilot evaluation of the prototype system was conducted with a group
of 75 students from Hanoi University of Science and Technology, aiming to assess the platform’s initial
effectiveness. The trial results demonstrated positive outcomes: faster content searches, more relevant
material recommendations, and higher user satisfaction compared to traditional resource platforms. These
findings suggest that integrating AI agents such as chatbots into online learning platforms can effectively
personalize learning materials and enhance the user experience. The study concludes that AI-driven
personalized learning platforms like Edvangarde have the potential to improve student engagement and
resource accessibility, ultimately contributing to better learning outcomes.
Keywords: AI Agent, AI Chatbot, Digital Library, Natural Language Processing, Personalized Learning.

1. Introduction1
Digital libraries have become more popular, but
most search interfaces are still based on keywords.
This system rarely understands the learners’ intent
when expressed in natural language, nor do they
consider the goals, proficiency levels, and preferred
learning style of each learner. As a result, learners
spend a lot of their time sifting through the long list of
search results, but still find it difficult to find suitable
material. To meet new expectations, we need systems
that can convert natural language questions into
personalized and pedagogically valuable resource Fig. 1. Forecast of AI Technology Applications in
recommendations, rather than merely keyword Education by 2032 (Source: Edtech Agency)
matching. In the context of digital libraries rapidly shifting
Artificial intelligence, especially conversation toward the application of smart technologies. AI
agent (AI agent), has rapidly developed and been Chatbot has emerged as an important tool to support
widely used in education for roles such as tutoring, and personalize users' experience. Zan, Zhao, and
regular assessment, grading, and classroom Mazumdar (2023) show that chatbots are primarily
management. Both academic research and industrial deployed on library websites or social platforms to
applications continue to grow [1, 2], and generative AI provide reference services, information retrieval, and
products are expected to significantly expand in the 24/7 support, while also reducing librarians’ workload
education market [3]. However, most current [4]. Similarly, Aboelmaged et al. (2024) emphasized
applications focus on integrated assistants in LMS, the role of chatbot in expanding accessibility,
foreign language tutoring, or Q&A in courses. Few improving conversation design, and usability, which
studies consider chatbots as an interface for accessing became practically prominent during the COVID-19
digital libraries or learning repositories to retrieve period. When libraries had to maintain online service
resources that match the goals and learning profiles of [5]. However, challenges regarding costs, security, and
learners. sustainability indicate the need for deeper research,
especially in developing personalized chatbot systems

1 [Link]
ISSN 2734-9381
Received: ......./..../......; accepted: ......./..../......

1
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

to enhance the effectiveness of exploring and 2.1. Personalized Learning

accessing learning resources. This gap defines our
Personalization is a process of designing an
issue: personalizing the search and retrieval of learning
experience, product, or service to suit the needs, goals
materials on document sharing platform, based on
and reference frame of each individual. In education,
natural language requirements.
personalized learning is a teaching strategy in which
We conceptualize Edvangarde as an integrated content, process and strategies for learning are
digital library with a chatbot, based on an AI agent adjusted to suit the unique characteristic, interest, and
architecture linked to three research streams: NLP, AI capabilities of each learner. This concept stems from
agents, and digital libraries. The dialogue model based progressive educational theories, in which John
on LLM analyzes the learner's request, identifies Dewey emphasizes the learner’s active role and the
intentions and entities (e.g., course/topic, type of task, idea that teaching must go hand in hand with a
level, preferred learning method) and converts it into a constantly changing world [6].
structured query. The toolkit connects to a learning
Personalized learning often aims to help each
resource repository enriched with metadata (intended
learner reach their full potential by designing an
usage purpose, difficulty, comprehensibility, format,
optimal learning environment for each individual,
learning objectives). The Coordination layer plans
where every element from content to learning
retrieval and re-ranking, prioritizing resources that
method,… are flexibly adjusted to suit the level and
meet the learner's intentions and attributes, before
goal of each learner. Personalized learning not only
returning concise, actionable suggestions. This
helps to acquire knowledge but also encourages
mechanism forms a consistent theoretical framework
learners to be proactive, confident and responsible for
for personalized learning resource searching, rather
their own learning process. It also promotes self-
than just a general chatbot interaction.
learning and confidence in one’s abilities.
We posed a question: Can a digital library
To implement effectively, personalized learning
integrated with a chatbot enable faster, more accurate,
involves five elements: time, place, pace, learning path
and more personalized learning resources compared to
and learning objectives (Graham et al., 2019),
traditional keyword matching search? To evaluate
combined with four keys component: learner profiles,
Edvangarde, we designed a two-part study: (i) a task-
prior knowledge, personalized learning paths, and
based experiment, where students used the system to
flexible leaning environment based on data analysis
address specific information needs; and (ii) a post-task
(Shemshack et al., 2021). Technology plays an
survey measuring perceptions of search effectiveness,
important role in personalized learning, from learning
levels of personalization/accuracy, and
management systems to online learning tools and data
satisfaction/intention to continue use. The test set used
analysis, which allows the contents, methods, and pace
124 learning materials from 15 general education
of learning to be adjusted to suit the need of learners.
courses and involved 75 students at Hanoi University
However, for personalized learning to be truly
of Science and Technology.
effective, it must be combined with pedagogical
The contribution of the paper include three knowledge to help learners develop self-learning skills
points: and maintain long-term motivation for studying [7].
1) Problem framing & approached framework: 2.2. AI Agents and AI Chatbots
formalizing the problem of exploring
AI Agent, also known as an artificial intelligence,
personalized learning materials for digital
is an application capable of perceiving and interacting
libraries and proposing a general framework for
with surrounding environment through available tools
AI agents/NLP to realize personalization across
to process information and performs actions without
dimensions: content, level, learning method, and
direct human intervention, thereby finishing given
learning goals.
objectives. An AI Agent is made up from three main
2) System: implementing Edvangarde, an online components: Model, Tools and Orchestration Layer
platform embedded with a chatbot directly into [8]. Model is the heart of an AI Agent, usually powered
the document sharing interface. by large language models (LLMs) that allow the agent
to understand instruction, reason and act accordingly.
3) Empirical evidence: a pilot experiment with
For the model to operate effectively, it is necessary to
HUST students showed positive signs regarding
integrate the capability to track instructions, logical
search speed, relevance of suggestions, and user
reasoning, and the ability to self-learn from historical
experience compared to traditional keyword
data or fine-tuning. Tools are extensions of the
matching search methods.
model’s capabilities, allow AI Agent to interact with
the real world by retrieving data, calling API, or
updating database,… Tools allow AI Agent to pass the
2. Literature review limitations of normal LLMs. Orchestration layer is
responsible for action planning, determining the

2
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

execution sequence, and managing the process of aiming to make the dream of universal access to human
perception - reasoning - action. The more complex this language a reality, According to Fox et al. (1995), a
layer is, the more flexible and efficient the Agent. digital library is not merely a digitized version of
traditional library, but rather a distributed information
system that integrates technologies such as electronic
publishing, database management, search and retrieval
systems, allowing storage and sharing of digital
resources in the form of text, images, sound, video, and
scientific data [11].
Digital resources are organized systematically
and stored in the central storage or distributed across
networking platforms, allowing users to access
Fig. 2. General agent architecture and components information anytime and anywhere. These systems
(Source: Google) have contributed to long term knowledge preservation
An AI Chatbot is a specialized form of AI Agent, and ensured sustainable data retrieval. Digital
focusing on the ability to communicate with humans resources have become an irreplaceable part of modern
using natural language. Within the boarder studying and researching, especially in the growing
classification of AI Agent, AI Chatbot is considered trends of online learning and open education [12].
conversation-oriented agents, whose primary goal is to The development of technology has driven the
receive input in the form of text or speech, perform restructuring of the functions and roles of traditional
analysis, and generate appropriate response in real libraries, transforming them into digital information
time. According to Davar et al. (2025), AI Chatbots centers that integrate data, technology and users’
function as conversation-oriented agent, utilizing interaction. Today, the digital library is not only a
LLMs and natural language processing (NLP) to knowledge storage, but also a global cooperation
understand context, determine user intent, and platform, contributing to promoting pedagogical
generate meaningful response [9]. Similar to other AI innovation, widely shared knowledge and ensuring
Agents, AI Chatbots follow the perception - reasoning fairness in accessing scientific information [13].
- action process; however, chatbots’ actions are limited
to generation text or speech, rather than taking 3. Methodology
physical actions in the real world. 3.1. Survey and quantitative methods
2.3. Natural Language Processing Some recent research studies show that students
Natural Language Processing (NPL) is a core have many limitations in searching and evaluating
field of AI that allow computers to understand, online learning materials. In the study of Morrison
interpret, and respond to human language. NPL (2024) mentioned that despite having 20 minutes to
combines computer science, linguistics, and machine search, students only spend more than 4 minutes on
learning to analyze natural language in a contextually each topic, some groups offer 11 different queries, or
and semantically meaningful way. NLP tasks include only 2 queries, but they show the concept of online
tokenization, stop word removal, parts of speech search as a fast process, only to take "answers";
tagging, text normalization, syntactic parsing. More Students interpreting the search is "fast and easy", in
advanced techniques include feature extraction (TF- fact nearly half of the search for the search and many
IDF, word embedding) and deep learning models cases stop after a few dozen seconds [14]. Oger &
(LSTM, BERT, GPT) allow machines to gain more Sanchiz (2025) also discovered the average time for
contextual understanding. simple tasks fluctuating 269–308, while with complex
tasks increasing to 536–604 units of time, showing the
NLP is the foundation for many AI systems such level of investment depends on the context and
as chatbots, virtual assistants, machine translation, difficulty [15]. Add to it, McGrath et al. (2025) Prove
sentiment analysis. NLP plays the roles of a that reading skills are regulating factors: Students who
“processing engine” that helps chatbots understand read well the source assess the source is not reliable is
and hold conversation with users. NLP enables less useful and less reliable, while the weak reader is
chatbots identify intents, extract important entities, and based on personal beliefs or keyword joints [16].
generate relevant, coherent response. As a result,
chatbots can go beyond simple keyword matching to On this basis, the authors conducted a survey to
fully understand user question based on semantics and evaluate the effectiveness of the document sharing
structures [10]. platform integrated with AI chatbot. Participants were
informed about the purpose, benefit, and voluntary
2.4. Digital Library nature of the study. The questionnaire consisted of
Digital libraries emerged from the integration of three scales corresponding to three criteria: Document
the internet, hypertext, and multimedia technologies,

3
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

searching experience, Personalization and accuracy, The documents were stored on Google Drive,
User satisfaction and continuous usage intention. each documents have data such as title, course,
knowledge unit, format, proficiency level,
The survey link is sent to over 100 students who
understandability, and leaning objective, which are
expressed interest in testing the system over an eight
stored in a Google Sheets.
week period from May 1, 2025 to June 19, 2025. A
total of 86 responses were received, of which 75 valid Next, we design a chatbot model using the n8n
responses were retained after removing incomplete or automation platform. The model operates based on the
invalid entries. GPT-4o-mini language model. The input data for the
chatbot is the repository of 124 learning materials for
We uploaded the data from 75 responses onto the
15 courses. The model is connected to the Google
SPSS27 software to perform quantitative data analysis.
Sheets file containing the data of the materials. Figure
Only observed variables with Cronbach’s alpha
3 illustrates the architecture of AI Chatbot on the n8n
coefficient greater than 0.7 (Cortina, 1993) were
platform:
included in the analysis [17]. The questionnaire used a
5-point Likert scale, ranging from 1 represented
“Strongly disagree” to 5 represented “Strongly agree”.
3.2. Experimental method
We conducted an experiment with the student
participants, with detailed instructions in the survey.
To prepare for the experiment, we designed
Edvangarde, which is a document sharing website
designed to search for documents for courses taught at
Hanoi University of Science and Technology (HUST).
The website had a built in chatbot that support natural
language document search.
Fig. 3. Architecture of the Edvangarde AI Chatbot
We prepared a repository of 124 learning (Source: Authors’ own work)
materials, covering 15 fundamental courses taught to
Finally, we develop the user interface for the
student at HUST, and uploaded to the system. The
Edvangarde website, where all the learning materials
included materials served different academic purposes
are made accessible. We integrated the AI Model into
and all content was in Vietnamese, with the main file
the website as an interactive button, which, when
format being PDF, a small number are audio, video,
clicked will active a chat window to communicate with
image,… The list of included courses is shown in
the chatbot (Figure 4).
Table 1:
The operation of chatbot on the website
Table 1. Summery of documents used
Edvangarde when the user is input required to be
described in Figure 5. This process consists of four
Course name Quantity
stages: extracting keywords, database query, context
Calculus I 8 supplements and answering answers, integrated
Calculus III 10 through Openai API to create accurate and
personalized feedback for users.
Algebra 14
Probability and Statistics 13 The prototype will be implemented with some
sample questions such as:
Physics I 13
Physics II 17 - “Tìm cho tôi TẤT CẢ đề thi cuối kỳ môn Đại số phù
Introduction to Computer Science 4 hợp với NHÓM NGÀNH 2” (Find me ALL final exam
papers for the algebra course suitable for GROUP 2)
Introduction to the Legal Environment 16
Philosophy of Marxism and Leninism 2 - “Tôi muốn tìm tài liệu Giải tích 3 phù hợp với việc
tôi hay học qua hình thức xem VIDEO bài giảng” (I
Political Economics of Marxism and
5 want to find Calculus 3 materials suitable for my
Leninism
learning through VIDEO lectures)
Scientific Socialism 4
History of Vietnamese Communist Party 2 - “Tôi gặp khó khăn trong việc giải các dạng bài tập
môn Kinh tế chính trị Mác-Lênin, tìm cho tôi hướng
Ho Chi Minh Ideology 3
dẫn giải bài tập!” (I struggle with solving Political
Soft Skills 3 Economics of Marxism and Leninism problems. Find
Applied Psychology 10 me some solution guides.)

4
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

- “Tôi gặp khó khăn trong việc chuẩn bị thí nghiệm - “Tôi cần tài liệu TRỰC QUAN giúp tôi hệ thống hoá
VLĐC1, hãy cho tôi tài liệu dễ hiểu nhất!” (I struggle môn Chủ nghĩa Xã hội khoa học” (I need a VISUAL
to prepare for experiment in PHYSICS I. Find me the material to help me systematize review the Scientific
easiest to understand material) Socialism course)
- “Tôi cần tìm công cụ hay bảng tính giúp tôi XỬ LÝ - “Tôi cần tìm tài liệu Tư tưởng Hồ Chí Minh phù hợp
SỐ LIỆU thí nghiệm Vật lý đại cương nhanh hơn” (I với cách học tập qua việc lắng nghe” (I need to find
need a tool or spreadsheet to help me PROCESS materials for Ho Chi Minh Ideology course that
EXPERIEMNT DATA for Physic more efficiently) suitable for auditory learning)

Fig. 4. Edvangarde website interface integrate with AI Chatbot

Fig. 5. User Request Processing Flowchart (Source: Proposed by Authors)

5
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

4. Results 5 minutes remains a challenge for many, possibly due

to differences in operation speed or familiarity with the
4.1. Participants
system, as well as the time users spend reflecting and
Through data processing, we compiled statistical formulating comprehensible requests to the chatbot.
information from survey participants. The participants
Additionally, the variable DSE3 get a mean score
comprised 38 females (50.7%) and 37 males (49.3%).
of 4.13, but with the highest standard deviation (0.949)
The 18 students (24%) were enrolled in second
among all items. This reflects a relatively large
university year, 21 students (28%) were enrolled in
difference between the respondents - while some users
third university year, 16 students (21.3%) were
find it very easy to locate suitable materials, others
enrolled in third university year, and 20 students were
may still have some difficulties in identifying the
in the fourth university year (26.7%).
suitable learning materials.
4.2. Reliability of the Scale
Overall, the survey results indicate that students
The reliability of the scale was graded using have a positive view of the search experience on the
Cronbach’s Alpha to determine whether the observed system, particularly regarding the speed and
variables had high internal consistency and provided a convenience enabled by the chatbot. The time users
general explanation for the underlying construct being spend on a single document retrieval task has been
measured. The analysis showed that all factors had a reduced to some extent compared with prior studies on
Cronbach’s Alpha above 0.8, indicating high search behaviour by Morrison, Oger & Sanchiz
reliability: Document searching experience (α = discussed above. However, there remains a
0,883), Personalization and accuracy (α = 0,844), User considerable difference in users’ ability to quickly
satisfaction and continuous usage intention (α = retrieve suitable documents, suggesting a need for
0,831). This result indicated that the collected data further improvements in the interface,
show strong internal consistency, both for the overall recommendation algorithms, or guidance on how to
scale and each individual component. formulate effective prompts.
4.3. Description of Components 4.3.2. Personalization and accuracy
4.3.1. Document searching experience In the next measurement scale regarding
Personalization and accuracy, Table 3 shows that mean
In the first scale regarding the Document
score of the observed variables about personalization
searching experience, Table 2 shows that the mean
and accuracy of the system range from 3.99 to 4.07.
scores for document searching experience of the
This is a relatively high score range reflects a positive
observed variable range from 3.88 to 4.48. This shows
attitude of students toward the system’s ability to
a relatively high level of user satisfaction with the AI suggest suitable learning materials. This suggests that
integrated system, indicating that the product has the integrated chatbot has, to some extent, met
effectively met expectations in supporting convenient learners’ needs for contextual understanding and
and fast retrieval of study materials. relevant recommendations.
Table 2. Survey for Document searching experience Table 3. Survey for Personalization and accuracy
Observed
Mean Score Std. Deviation Observed
Variable Mean Score Std. Deviation
Variable
DSE1 4.15 0.730 PA1 4.01 0.797
DSE2 4.25 0.790
PA2 3.99 0.951
DSE3 4.13 0.949 PA3 4.07 0.935
DSE4 3.88 0.929
DSE5 4.48 0.860 The observed variables assessing the
Personalization and accuracy of the system show a
Overall, two variables DSE2 and DSE1 received general trend of trust and high evaluation for the
high mean score of 4.25 (SD = 0.790) and 4.15 (SD = suggested materials to students. Among these, the
0.730) respectively. These result shows that students variable PA3 had the highest mean score (M = 4.07;
value the role of chatbot in increasing the speed of SD = 0,935), indicating that most learners agreed the
academic materials retrieval. Notably, when compare system can adapt to individual need. However, the
the to criteria related to retrieval time, the variable relatively high standard deviation reflects variability in
DSE 5 had a significantly higher mean score (4.48) user experiences. The variable PA1 has mean score of
than the variable DSE 4 (3.88). This difference 4.01 (SD = 0,797), showing the system’s ability to
indicates that learners easily achieve an efficient suggest suitable materials to be fairly consistent. On
search experience with the time spent of around 10 the other hand, the variable PA2 received the lowest
minutes, while locating the desired materials in under mean score of 3.99 and the highest standard deviation

6
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

if 0.951, show that a portion of students felt the chat Students gave positive feedback, especially about
bot didn’t fully understand their search intents. the speed and usefulness of the search results. The
chatbot was able to understand questions in
In summary, the results indicate that the system
Vietnamese and provide answers right away.
is on the right track in terms of personalization and
However, the way the system personalizes results is
increasing accuracy, but still need to improve the
still simple, as it mainly uses basic filters like
ability to understand users’ search intents to deliver a
keywords or formats that need to be labeled by
more consistent experience for all users.
administrators. The system also has trouble with
4.3.3. Satisfaction and continuous usage intention unclear or complicated questions, so the quality of
results still depends a lot on how clearly students ask.
In the final measurement scale regarding
Satisfaction and continuous usage intention, Table 4 Since the system is still experimental, it is not yet
shows that the mean score of the observed variable feasible to expand its scope considerably. More
related to satisfaction and continuous usage intention documents and courses could be added in the future,
range from 4.07 to 4.16. This is a relatively high score, but for now the main focus should be on checking its
reflecting the positive attitudes and trust of students in performance and stability before any expansion.
the system’s effectiveness and value that the system Although the GPT-4o-mini model is a more cost-
brings in learning process. effective option compared with larger models, large-
scale implementation may still face challenges related
Table 4. Survey on Satisfaction and continuous usage
to expenses and technical limitations. The system has
intention demonstrated that it can support fast, accurate, and
Observed relevant resource discovery - an important step toward
Mean Score Std. Deviation more flexible, student-centered learning. Combining
Variable
the chatbot with the document sharing platform also
SC1 4.07 0.664 enhances students’ self-learning ability and shows
SC2 4.15 0.911 strong potential for future development. Moving
SC3 4.16 0.839 forward, we need to continue to improve context
understanding, refine the recommendation method,
The variables that evaluate satisfaction and future and provide clearer guidance to users, so that learners
usage intentions show a consistent and positive trend of all levels can get the most out of the system.
among students. Notably, the variable SC3 received
the highest mean score of 4.16 and a standard deviation 5. Discussion
of 0.839, showing that the system not only meets Based on the pilot results, the team recognizes
individual learning needs but also trustworthy enough that the AI-powered learning material sharing platform
for users to recommend to others. Similarly, the has proven promising preliminary efficacy in assisting
variable SC2 had the mean score of 4.15, although students in locating relevant content promptly,
slightly higher standard deviation of 0.911 shows some accurately, and individually. To expand its usage and
variation in long term usage across students. improve the overall learning process, certain strategic
Interestingly, the variable SC1 had mean score of 4.07 and long-term suggestions are advanced.
and the lowest standard deviation, indicates strong
agreement and positive feedback from the majority of First, schools must consider integrating AI
learners regarding their experience in using the Chatbots into their holistic digital innovation strategies
platforms. These findings shows that students not only as key components of intelligent learning. Not only
satisfied with the system quality but also have an can the chatbots serve as knowledge centers for
inclination to continue using it and share their learning content, but they can also evolve into virtual
experience in the learning community. teaching assistants that give students personalized
guidance, offering tailored learning tracks based on
4.4. Findings learners' competency levels and goals. It is critical to
have policies and short-term training in the fields of
The preliminary survey results and feedback
deploying AI chatbots and prompting abilities, given
from pilot users show that the Edvangarde system is
starting to work well in helping students find learning the rapid evolution of the technology. The system also
materials that match their needs. The chatbot goes needs to be supported by regular guidelines and user
beyond simple keyword searches by understanding the manuals to make it accessible and enhance interaction
abilities, making the technology accessible with ease
meaning of queries and suggesting resources relevant
to students.
to the course, topic, or format. In the test setting with
a digital library of 124 documents from 15 general Secondly, to ensure system utility and cost-
education modules, students usually spent 5 to 10 efficient performance, especially with the resource
minutes on a search task, while the chatbot gave database increasing, it is advisable to adopt open-
answers within 10 to 20 seconds. source, internally controlled models like Ollama or
Continue. These are adaptable and help in obtaining

7
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

data security. Additionally, the system must be hosted theories, chatbot technology, natural language
on a stable, dedicated server to ensure uninterrupted processing, and personalized learning. Aware of the
operation. common issue students face while looking for
appropriate materials, we created Edvangarde, which
Thirdly, in order for this technology to become
contains 124 general documents, allowing students to
most effective, it is essential to continuously enhance
communicate with the system in natural language
the contextual recognition capability of the chatbot.
questions based on the GPT-4o-mini language model.
These involve improving suggestion algorithms,
The chatbot allows natural language questions and
improving comprehension of the purpose of a query,
provides descriptions and summaries of the documents
and incorporating intelligent analyses of feedback
to allow the user to make quick, appropriate material
from user behavior. The system can also be augmented
choices. User experience survey responses indicate the
by advanced analysis capabilities such as deep filtering
system offers enhanced search effectiveness, an
of content, learning objectives alignment, response in
enhanced overall learning experience, and indicates
terms of importance, or even comparison operations
promise for adaptive learning. Nevertheless, the
across different documents. These improvements
product remains short of limitations, such as the
would enable more thorough, individualized responses
relatively small base of uploaded documents that
based on differentiated learning styles. Furthermore,
constrains its completeness, and the capabilities of the
integrating automatic labeling, classification, and
model as of now are not more than simple semantic
metadata updates upon document uploads, coupled
understanding. Additionally, the cost of operations in
with copyright detection, would enhance efficiency
the form of using GPT-4o-mini as a barrier to mass
and precision in content management.
deployment. In the future, we recommend developing
Fourth, broadening interaction modalities from comprehensive user guides for the chatbot, enhancing
text-based interactions to others could significantly the system capabilities, implementing further
expand the user experience. Such technologies as interaction features, constructing a larger document
image recognition or voice interaction can be store, and migrating to open-source models like
integrated into the system and enhance its flexibility Ollama. Through this approach, expenses are reduced,
and usability. Maintenance of the learner resource and customization is improved, offering a sustainable
database from time to time and ongoing maintenance foundation for deployment in higher education.
of the system should be ensured to prevent the chatbot
References
from becoming outdated or providing wrong
information. Automated checks and manual [1] L. Chen, P. Chen, and Z. Lin, Artificial intelligence in
verification can enhance the content consistency and education: A review. IEEE Access, vol. 8, pp. 75264–
reliability. Rights management of access should also 75278, 2020,
be ensured to protect sensitive data, so that security is [Link]
ensured while optimizing retrieval processes for data. [2] C. W. Okonkwo and A.-I. Abejide, Chatbots
applications in education: A systematic review.
Finally, as AI becomes increasingly prevalent in Computers and Education: Artificial Intelligence, vol.
the learning environment, safeguarding user data 2, 2021, Art. no. 100033,
privacy and data security should be of prime [Link]
importance. The chatbot must adhere strictly to legal
[3] Edtech Agency, Vietnam’s Edtech White Paper 2024.
regulations for gathering, processing, and preserving Edtech Agency, Hanoi, Vietnam, 2024.
personal information. Ethical practices for leveraging
AI must be clearly defined to ensure transparency in [4] R. Yan, X. Zhao, and S. Mazumdar, Chatbots in
responding and advising and prevent any negative libraries: A systematic literature review. Education for
impact on the learning process. In summary, deploying Information, vol. 39, pp. 431–449, 2023,
[Link]
AI Chatbots in schools is not exactly a technical
solution; it is a strategic, long-term decision to [5] M. Aboelmaged, S. Bani-Melhem, M. A. Al-Hawari,
optimize the whole education system. The and I. Ahmad, Conversational AI Chatbots in library
recommendations herein rely on striking a balance of research: An integrative review and future research
leveraging the advantage of AI while addressing its agenda. Journal of Librarianship and Information
Science, vol. 57, no. 2, pp. 331–347, 2025,
risks. Having an orchestrated, continuously analyzed,
[Link]
and dynamic roadmap will be the catalyst to drive
further sustainable, effective, and future-oriented [6] T. T. T. Hương, Học tập cá nhân hóa: Các nhân tố cần
educational reform in keeping with the prevailing thiết và lưu ý triển khai. Tạp chí Khoa học Giáo dục
digital transformation currents. Việt Nam, vol. 18, 2022.

6. Conclusion [7] C. R. Short and A. Shemshack, Personalized Learning.

EdTechnica: The Open Encyclopedia of Educational
As such, based on this research, our group was Technology, 2023, [Link]
able to develop and apply a resource-sharing platform
in tandem with an AI Chatbot, guided by AI agent

8
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

[8] J. Wiesinger, P. Marlow, and V. Vuskovic, Agents. access. Information Research, vol. 30, no. 2, 2025,
Google, Mountain View, CA, USA, White Paper, Sep. [Link]
2024.
[14] R. Morrison, Learning to search and learning to slow
[9] N. F. Davar, M. A. A. Dewan, and X. Zhang, AI down or “The quick and the dead”. Journal of
Chatbots in Education: Challenges and Opportunities. Documentation, vol. 80, no. 6, pp. 1475–1493, 2024,
Information, vol. 16, no. 5, p. 235, May 2025, [Link]
[Link]
[15] D. Oger and M. Sanchiz, Influence of task context and
[10] N. Bhardwaj, Natural Language Processing in AI task complexity on students’ task model quality, search
Chatbots. 2025, behavior and search outcome. Reading and Writing,
[Link] Jan. 2025, [Link]
7.
[11] E. A. Fox, R. M. Akscyn, R. K. Furuta, and J. J. Leggett,
Digital libraries. Communications of the ACM, vol. 38, [16] Catherine A. McGrath, Jason L.G. Braasch, Laura K.
no. 4, pp. 22–28, Apr. 1995, Allen, and Erica D. Kessler, How do topic beliefs and
[Link] reading skill influence college students’ evaluations of
search engine results for usefulness and
[12] J. Raju, Knowledge and skills for the digital era trustworthiness?. Learning and Individual Differences,
academic library. The Journal of Academic 122, 102751, 2025,
Librarianship, vol. 40, no. 2, pp. 163–170, Mar. 2014, [Link]
[Link]
[17] J. M. Cortina, What Is Coefficient Alpha? An
[13] S. Hawamdeh, D. Alemneh, and J. Allen, The digital Examination of Theory and Applications. Journal of
transformation of the information professions: Applied Psychology, vol. 78, no. 1, pp. 98–104, 1993,
Exploring the impact of digital publishing and open [Link]

Appendix: Scale and Coding

Scale Observed Variable Coding
Document searching I feels that the documents searching process on the system is
DSE1
experience fast.
Chatbot helps me cut down on time on searching for
DSE2
documents compared to the traditional way.
I can easily find the materials that fit my learning needs. DSE3
The average time it takes me to find a suitable material is
DSE4
under 5 minutes.
The average time it takes me to find a suitable material is
DSE5
under 10 minutes.
Personalization and The materials recommended by the chatbot match the
PA1
accuracy subject and topic I am looking for.
Chatbot understand my search intent through my questions. PA2
I feel that the system provides personalized
PA3
recommendations for learning materials.
Satisfaction and I’m satisfied with the overall experienced on the website. SC1
continuous usage
intention I will continue to use the system in the future. SC2
I’m willing to recommend this system to my friends. SC3

9
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

Reviewer Communication Form

(Authors should include this part in revised submission)

Reviewers Comments to Author(s) Author Response and Revision

1. First Review
Reviewer 1
Once your manuscript has come back from reviewers The following is an example as to how to respond to
you may be given the opportunity to revise it in a reviewer comment:
accordance with the reviewer comments. You will
usually receive a letter from the editor who handled
your manuscript outlining the changes they would like
you to make and links to the reviewer reports. This
letter usually contains information on how to return
your revised manuscript including instructions on how
to highlight the changes made and when you need to
return the revised version.
Reviewer comment: “In your analysis of the data you Response in agreement with the reviewer: “We
have chosen to use a somewhat obscure fitting agree with the reviewer’s assessment of the analysis.
function (regression). In my opinion, a simple Our tailored function does make it impossible to fully
Gaussian function would have sufficed. Moreover, the interpret the data in terms of the prevailing theories.
results would be more instructive and easier to In addition, in its current form, we agree it would be
compare to previous results”. difficult to tell that this measurement constitutes a
significant improvement over previously reported
values. We have therefore re-analyzed the data using
a Gaussian fitting function”.
Response disagreeing with the reviewer: “We agree
with the reviewer that a simple Gaussian fit would
facilitate comparison with the results of other studies.
However, our tailored function allows for the analysis
of the data in terms of the Smith model [Smith et al,
1998]. We have added two sentences to the paper
(page 3, paragraph 2) to explain the use of this
function and Smith’s model”
Recommendation: Major revision
Reviewer 2
When revising your manuscript and responding to Address all points raised by the editor and reviewers.
peer review comments you must:
Describe the major revisions to your manuscript in
your response letter followed by point-by-point
responses to the comments raised.
Perform any additional experiments or analyses the
reviewers recommend (unless you feel that they
would not make your paper better; if so, please
provide sufficient explanation as to why you believe
this to be the case in your response letter)
Provide a polite and scientific rebuttal to any points
or comments you disagree with. Remember if your
manuscript is sent for a second round of peer review
the reviewers will see this letter too.

10
 International Conference on
Educational Science and Foreign Language Teaching, ICEF

Clearly show the major revisions in the text, either

with a different color text, by highlighting the
changes, or with Microsoft Word’s Track Changes
feature. This is in addition to describing the changes
in your point by point cover letter.
Return the revised manuscript and response letter
within the time period allotted by the editor.
Recommendation: Minor revision
2. Second Review
Reviewer 1

Reviewer 2
Recommendation: Suitable for publication in its
current form.
or
Recommendation: Not suitable for publication.

11