Integrating a LLaMa-based Chatbot with Augmented Retrieval

Generation as a Complementary Educational Tool for High School and

College Students

Dar

ıo S. Cabezas

1 a

, Rigoberto Fonseca-Delgado

1 b

, Iv

an Reyes-Chac

2 c

Paulina Vizcaino-Imaca

2 d

and Manuel Eugenio Morocho-Cayamcela

1,2 e

Yachay Tech University, School of Mathematical and Computational Sciences, DeepARC Research Group,

Hda. San Jos

e s/n y Proyecto Yachay, Urcuqu

ı, 100119, Ecuador

Universidad Internacional del Ecuador, Faculty of Technical Sciences, School of Computer Science,

Quito, 170411, Ecuador

Keywords:

Educational Chatbot, LLaMa-7B-chat, Artiﬁcial Intelligence in Education, Large Language Models, Vector

Databases.

Abstract:

In the current educational landscape, the transition from traditional paradigms to more interactive and per-

sonalized learning experiences has been accelerated by technological advancements, particularly in artiﬁcial

intelligence (AI). This paper explores integrating large language models (LLMs) with retrieval augmented

generation techniques (RAG) to develop an educational chatbot to enhance students’ learning experiences.

Leveraging the LLaMa 7B-chat model and RAG technique, our system incorporates a structured mathemat-

ical database supplemented with relevant audiovisual resources. Furthermore, leveraging the Pinecone API

enhances information retrieval efﬁciency through cosine similarity. This capability empowers the chatbot to

deliver precise and relevant responses to students’ inquiries by accessing documents from Pinecone. More-

over, incorporating system prompts and memory functionality contributes to a more personalized learning

experience, enriching student interaction with the educational assistant. The ﬁndings suggest these assistants

can enhance student engagement and facilitate knowledge acquisition.

1 INTRODUCTION

In the era of Education 1.0, students primarily ob-

tained information from books and web pages, yet

these traditional mediums lacked the essential ele-

ment of interactivity. The accurate perspective of-

fered by (Gerstein, 2014) on the 3 Rs—Receiving,

Responding, and Regurgitating—many educational

institutions remain entrenched in this outdated

paradigm. Education 1.0 is ﬁrmly grounded in the

passive acquisition of knowledge from instructors to

students (Gerstein, 2014; Songkram et al., 2021; Rane

et al., 2023). Consequently, Education 2.0 marked

a crucial turning point in the evolution of the educa-

tional process (Huk, 2021). Education in the ”Online

https://orcid.org/0000-0003-2668-3949

https://orcid.org/0000-0002-8890-3911

https://orcid.org/0009-0002-2731-5531

https://orcid.org/0000-0001-9575-3539

https://orcid.org/0000-0002-4705-7923

world” provided the opportunity for users to comment

and interact with content (Gerstein, 2014; Huk, 2021;

Songkram et al., 2021; Rane et al., 2023), reﬂecting

a signiﬁcant shift toward a more engaging and par-

ticipatory learning environment. Hence, Education

3.0 revolves around personalized learning tailored to

students’ interests, fostering innovation and creativity

(Huk, 2021; Songkram et al., 2021; Rane et al., 2023).

In this paradigm, the role of the teacher has trans-

formed, evolving into that of an organizer facilitating

an educational environment where students collabora-

tively cultivate their knowledge (Keats and Schmidt,

2007; Songkram et al., 2021). In this way, we are

moving from a passive to an active learning process

through technology and the internet.

As the educational landscape undergoes continu-

ous transformation, the rise of Education 4.0 marks

the forefront of technology integration in learning.

Central to the essence of Education 4.0 is the ongo-

ing enhancement of artiﬁcial intelligence (AI) (Huk,

2021; Rane et al., 2023). In (Rane et al., 2023), Rane

Cabezas, D., Fonseca-Delgado, R., Reyes-Chacón, I., Vizcaino-Imacaña, P. and Morocho-Cayamcela, M.

Integrating a LLaMa-based Chatbot with Augmented Retrieval Generation as a Complementary Educational Tool for High School and College Students.

DOI: 10.5220/0012763000003753

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Conference on Software Technologies (ICSOFT 2024), pages 395-402

ISBN: 978-989-758-706-1; ISSN: 2184-2833

395

speciﬁcally emphasizes how AI acts as a catalyst, pro-

pelling educational transformation. This paradigm

shift lays the foundation for exploring the key fea-

tures of AI in Education 5.0. These encompass per-

sonalized and adaptive learning, intelligent learning

assistants (ILA), predictive analytics for early inter-

vention, integration of augmented reality (AR) and

virtual reality (VR), gamiﬁcation and simulation, life-

long learning and skill development, and the ethical

considerations of AI in fostering inclusive education

(Rane et al., 2023).

The ﬁeld of education has undergone a remark-

able transformation, propelled by advancements in AI

technologies that have signiﬁcantly inﬂuenced per-

sonalized learning (Rane et al., 2023). Natural Lan-

guage Processing (NLP), a subset of AI, empow-

ers machines to comprehend, interpret, and gener-

ate human language. This technology offers numer-

ous educational advantages, particularly in facilitat-

ing personalized language-learning experiences (You-

nis et al., 2023). However, the ascendancy of deep

learning in NLP (Rawat et al., 2022), the availability

of vast public datasets (Lhoest et al., 2021), and the

capabilities of powerful computing devices (Sharir

et al., 2020) to handle substantial data through so-

phisticated algorithms have catalyzed the emergence

of Large Language Models (LLMs) like Large Lan-

guage Model Meta AI (LLaMa) (Touvron et al.,

2023) developed by Meta, GPT Family (Achiam

et al., 2023) developed by OpenAI, Gemini (Gemini

et al., 2023) by Google, among others. Renowned

for their proﬁciency in capturing semantic relation-

ships between words and phrases (Adnan and Akbar,

2019),these models, alongside chatbots, have become

a formidable force, ushering in an unprecedented rev-

olution in education (Younis et al., 2023; Rane, 2023).

The rapid advancement of technology, particularly

in the digital realm, is reshaping global education sig-

niﬁcantly. Despite the long-standing integration of

information and communication technology (ICT) in

education, the past four decades have shown the great-

est potential for transformative change (UNESCO,

2023). Simultaneously, an educational technology in-

dustry has emerged, dedicating its endeavors to the

development and dissemination of educational con-

tent, learning management systems, language appli-

cations, as well as augmented and virtual reality tools,

personalized tutoring, and assessment platforms (UN-

ESCO, 2023; Rane et al., 2023). However, this trans-

formation varies widely based on socioeconomic and

educational contexts (Arias Ortiz et al., 2024; UN-

ESCO, 2023). At the same time, in Latin America

and the Caribbean, the challenges persist. According

to the results of PISA 2022, the average rate of low

performance in the region was 75% in mathematics,

55% in reading, and 57% in science (Arias Ortiz et al.,

2024). The proposed solution involves the devel-

opment of an educational chatbot leveraging LLaMa

(Touvron et al., 2023), using retrieval augmented gen-

eration (RAG) (Lewis et al., 2020) techniques to en-

hance the accuracy and reliability of responses. This

enhancement is facilitated through the utilization of

a vector database(Pan et al., 2023; Pinecone, 2023),

which is queried to bolster the quality and reliability

of the model’s outputs.

2 METHODOLOGY

The methodology section outlines the steps taken to

develop the educational chatbot. We initiated the pro-

cess by integrating the LLaMa-7b-chat model into

our framework, covering its download, conversion,

and preparatory stages. Additionally, we explored

the role of vector databases, investigating their sig-

niﬁcance across various applications and examining

prominent options for accurate similarity search and

retrieval. Furthermore, we analyzed the visualiza-

tion of vector data, emphasizing the utility of three-

dimensional representations in clarifying data rela-

tionships. Lastly, we detailed the selection process

for tools and models used in embedding generation,

focusing on the Pinecone API and the all-MiniLM-

L6-v2 model from Hugging Face, highlighting their

functionalities and relevance to the study.

2.1 LLaMa-7b-chat Model

The LLaMa model, as introduced by Touvron et

al. (Touvron et al., 2023), served as the corner-

stone within our chatbot’s framework. Accessible for

download upon request via LLaMa Oﬁcial site, this

model played a pivotal role in our project. Speciﬁ-

cally, we utilized the 7B-chat model to enhance the

conversational abilities of our educational chatbot.

Upon approval of our request, we received a

signed URL via email to download the model. To ob-

tain it, we cloned the LLaMa repository from GitHub

and ran the download.sh script, initiating the down-

load process.

After downloading the weights of the 7B-Chat

model, the next step involves converting them into

a format that we can readily utilize. This conver-

sion process, known as quantization in the context of

LLMs, entails converting the model’s weights from

higher-precision data types to lower-precision ones,

aiming to reduce the computational resources needed

(Nagel et al., 2021; Kuzmin et al., 2024). To accom-

ICSOFT 2024 - 19th International Conference on Software Technologies

396

Figure 1: Word embedding map words in a corpus of text to vector space (Rozado, 2020).

plish this task, we utilize llamacpp, available from the

following link:llamacpp Github Repository. We then

proceeded to convert the model to FP16 format and

subsequently to ggml format, the format chosen for

the study. Speciﬁcally, we adopted the q4 0 quantiza-

tion method as outlined in Table 1.

Table 1: Available quantization types for the 7B-

Chat model. More information about each quantiza-

tion can be found at https://huggingface.co/TheBloke/

Llama-2-7B-Chat-GGML.

llama-2-7b-chat

Name Quant method Bits

.ggmlv3.q2 K.bin q2 K 2

.ggmlv3.q3 K S.bin q3 K S 3

.ggmlv3.q3 K M.bin q3 K M 3

.ggmlv3.q3 K L.bin q3 K L 3

.ggmlv3.q4 0.bin q4 0 4

.ggmlv3.q4 K S.bin q4 K S 4

.ggmlv3.q4 K M.bin q4 K M 4

.ggmlv3.q4 1.bin q4 1 4

.ggmlv3.q5 0.bin q5 0 5

.ggmlv3.q5 K S.bin q5 K S 5

.ggmlv3.q5 K M.bin q5 K M 5

.ggmlv3.q5 1.bin q5 1 5

.ggmlv3.q6 K.bin q6 K 6

.ggmlv3.q8 0.bin q8 0 8

2.2 Vector Database

Vector databases are crucial in various applications,

from information retrieval to recommendation sys-

tems and natural language processing (Han et al.,

2023; Pan et al., 2023). To explore their signiﬁcance,

we investigated popular vector database systems such

as (ChromaDB, ), Faiss (Douze et al., 2024), and

Pinecone (Pinecone, 2023) that stores a set of vectors

called embeddings, and provides a function to search

in them.

Figure 2: Visualization of Euclidean and Manhattan dis-

tances, and Cosine similarity.

These systems offer various methods for accurate

similarity search and retrieval, including cosine sim-

ilarity, Euclidean distance, Manhattan distance, and

dot product (Han et al., 2023; Douze et al., 2024; Pan

et al., 2023). For visual representations of Euclidean,

Manhattan, and Cosine distances, refer to Figure 2.

In Figure 3, we observe a visualization of embed-

dings in three dimensions generated using an embed-

ding projector. To fully comprehend how embeddings

can be visualized, Figure 1 presents a visualization

of various words processed through an embedding

Integrating a LLaMa-based Chatbot with Augmented Retrieval Generation as a Complementary Educational Tool for High School and

College Students

397

model, aiding in semantically translating their mean-

ings alongside a depiction of the vectors involved.

Figure 3: Representation of 150 points in 4 dimensions re-

duced to 3 dimensions using the PCA method of the Iris

dataset and colored by classes in Embedding Projector.

Finally, we employed the Pinecone API

(Pinecone, 2023) for its versatile functionalities,

alongside leveraging the all-MiniLM-L6-v2 model

from Hugging Face for embedding generation. This

model offers a 384-dimensional dense vector space,

facilitating comprehensive analysis and representa-

tion of textual data. Additionally, we integrated the

Langchain framework (Chase, 2022) to seamlessly

merge all components.

3 RESULTS

In our endeavor to construct a comprehensive dataset,

we adopted a systematic approach by drawing upon

the structured content of the widely utilized textbook

Prec

alculo: Matem

aticas para el C

alculo (Stewart

et al., 2010). Appreciated for its wide adoption in

Latin American academic institutions, this textbook

was the foundation for our database construction. By

aligning our dataset with the topics covered in the

book, we ensured a coherent and thorough represen-

tation of essential mathematical concepts.

To further enhance the depth and breadth of our

dataset, we engaged mathematics educators through

a targeted survey. This initiative solicited feed-

back on additional mathematical content and refer-

ences to audiovisual resources that complement our

dataset. Having assimilated the input and audiovisual

resources gathered, we structured our dataset as out-

lined in the following subsection. Each entry in this

table represents metadata essential for the retrieval

process utilized by Pinecone, enabling the retrieval of

the most relevant information. The metadata is then

transformed into embeddings using the all-MiniLM-

L6-v2 model showcased in Figure 4.

3.1 Math Database

In our endeavor to construct a comprehensive dataset,

we adopted a systematic approach by drawing upon

the structured content of the widely utilized textbook

Prec

alculo: Matem

aticas para el C

alculo (Stewart

et al., 2010). Appreciated for its wide adoption in

Latin American academic institutions, this textbook

was the foundation for our database construction. By

aligning our dataset with the topics covered in the

book, we ensured a coherent and thorough represen-

tation of essential mathematical concepts.

Table 2: Structure of the Math Database.

Field Description

Area

The speciﬁc mathematical area

covered by the content

Author Author(s) of the material

Chapter

Chapter number within the

textbook

Description Brief description of the content

Exercise

Mathematical exercise or

problem provided

Textbook Name of the textbook

Audiovisual

Material

Supplementary audiovisual

material related to the content

Subchapter

Subchapter or section within the

chapter

Subtopic

Subtopic or speciﬁc theme

addressed within the subchapter

Topic

Topic or broader category to

which the content belongs

Text

Text that help us to create

embeddings

To further enhance the depth and breadth of our

dataset, we engaged mathematics educators through

a targeted survey. This initiative solicited feedback

on additional mathematical content and references to

audiovisual resources that complement our dataset.

Through collaborative efforts, we integrated various

pedagogical materials and perspectives. This collab-

orative approach ensured the relevance and compre-

hensiveness of our dataset for educational purposes.

Having assimilated the input and audiovisual re-

sources gathered, we structured our dataset as out-

lined in Table 2. Each entry in this table represents

metadata essential for the retrieval process utilized

by Pinecone, enabling the retrieval of the most rele-

vant information. The metadata is then transformed

into embeddings using the all-MiniLM-L6-v2 model

showcased in Figure 4.

ICSOFT 2024 - 19th International Conference on Software Technologies

398

Figure 4: Visual representation of 1362 points reduced to 3

dimensions using PCA from our database, visualized using

Embedding Projector.

In Figure 4, each topic from our metadata is visu-

alized, allowing us to observe how the clusters of top-

ics differentiate through distinct colors. Furthermore,

the visualization reveals instances where clusters in-

tersect, indicative of shared topics. For instance, clus-

ters such as Analytical Trigonometry, Trigonometric

Functions - Unit Circle, and Trigonometric Functions

- Triangle Method intersect colored in yellow, green,

and brown, respectively, highlighting commonalities

across these topics. Conversely, certain sections ap-

pear distinctly isolated, such as Complex Numbers

colored in pink, primarily due to their limited over-

lap with other topics and subtopics. This transfor-

mation facilitated efﬁcient retrieval and integration

within our chatbot, providing contextual information

for enhanced interaction.

3.2 Pinecone Retrieval Responses

In this study, we utilized the Pinecone API to stream-

line our database management, enabling efﬁcient and

scalable integration. Leveraging the all-MiniLM-L6-

v2 model, we generated embeddings for our data and

conducted retrieval testing by converting our queries

into embeddings, facilitating search in the vector-

ized database. The Pinecone platform facilitated the

upload and management of our database, allowing

for seamless integration and scalability. Through

Langchain components (Chase, 2022), we created a

vector store using the text key, formed by concatenat-

ing all keys from each row in our database. Finally,

leveraging this vector store as a retriever, we con-

ducted queries and retrieved each query’s top k = 3

results.

The retrieval mechanism demonstrated through

the queries “I do not understand real numbers” and

“How to evaluate a function” shows the effectiveness

of Pinecone in retrieving relevant documents. In table

3, the top three retrieved documents from Pinecone

show cosine similarity scores of 0.6141, 0.5809, and

0.5714, while in table 4, the top three retrieved doc-

uments from Pinecone have scores of 0.5534, 0.5117

and 0.4967 respectively. These results suggest high

relevance between documents and queries, underscor-

ing the retrieval system’s effectiveness in providing

context.

Table 3: Top documents retrieved from Pinecone based on

cosine similarity to the query and semantic relevance.

query = “I don’t understand real numbers”

Document Document Document

id = 1 id = 0 id = 5

score = 0.614 score = 0.581 score = 0.571

Table 4: Top documents retrieved from Pinecone based on

cosine similarity to the query and semantic relevance.

query = “How to evaluate a function”

Document Document Document

id = 191 id = 196 id = 387

score = 0.553 score = 0.512 score = 0.497

3.3 Responses Comparison

Table 5 illustrates a comparative analysis of responses

generated by different chatbots, including ChatGPT

(Achiam et al., 2023), Gemini (Gemini et al., 2023),

and LLaMa (Touvron et al., 2023) version 7b-chat, in

response to a simple ”I don’t understand real num-

bers. Do you have videos?” query. This query serves

as an initial test to evaluate the behavior of the mod-

els when no additional context is provided. In such

instances, the responses from the base models tend

to resemble those of generic chat assistants, lack-

ing a clear purpose. However, when prompted with

additional context, such as information about their

name, mission, and developers, the models demon-

strate improved understanding and relevance in their

responses. In our proposed approach, the model is

trained to understand its purpose explicitly and high-

lights its capability to recommend relevant audiovi-

sual material to address user queries effectively.

Finally, our database is seamlessly integrated into

our retriever system via the Pinecone API. Utiliz-

ing the ConversationalRetrievalChain (CRC) from

Langchain (Chase, 2022), we execute our ﬁnal

queries, incorporating documents from the database

for contextual information. It leverages a language

model to rephrase the conversation into a query sent

to the retriever if available. Otherwise, it directly uti-

lizes the newest user input. Subsequently, it retrieves

relevant documents and incorporates them, along with

the ongoing conversation, to generate a response us-

ing another language model. Thus, the workﬂow of

our application, as illustrated in Figure 5, demon-

strates this process. Additionally, we have introduced

Integrating a LLaMa-based Chatbot with Augmented Retrieval Generation as a Complementary Educational Tool for High School and

College Students

399

Table 5: Comparison of Responses for Different Chatbots.

Query ChatGPT Gemini LLaMa

I don’t un-

derstand real

numbers.

Do you have

videos?

Understanding real num-

bers can be a bit tricky at

ﬁrst, but there are plenty

of resources available to

help you grasp the con-

cept! Videos can indeed

be a great way to learn.

There are many educa-

tional platforms like Khan

Academy, YouTube chan-

nels like Numberphile, and

various online courses that

offer video lessons on real

numbers and related math-

ematical topics.

Absolutely, there are many

videos that can help you

understand real numbers.

Here are a couple ways to

search for them:

What are Real Num-

bers? YouTube WHAT

IS A REAL NUMBER?

WHOLE NUMBER?

NATURAL NUMBER?

YouTube Real Numbers -

YouTube I hope this helps!

Hello! I’m here to help you

understand real numbers.

However, I must inform

you that I cannot provide

videos or any other con-

tent that promotes harmful

or false information. Real

numbers are a fundamen-

tal concept in mathematics,

and it’s important to under-

stand them in a safe and re-

sponsible manner.

Table 6: Comparison of Responses for Different LLaMa-7b-chat Conﬁgurations.

Query

LLaMa 7b-chat (Base

Model)

LLaMa 7b-chat +

PROMPT (Base Model +

System Prompt)

LLaMa 7b-chat +

PROMPT + Retrieval

(Our Proposal)

“I don’t un-

derstand real

numbers.

Do you have

videos?”

Of course! I’d be happy

to help you understand real

numbers. Here’s a video

that may help:

[Insert video link here]

I hope this video helps

you understand real num-

bers better. If you have any

questions or need further

clariﬁcation, please don’t

hesitate to ask. I’m here to

help!

Hello! So you’re looking

for help in mathematics,

perfect! As a virtual assis-

tant for mathematics, I’m

here to help you with any

questions you may have.

Whether you’re in high

school or university, I’m

here to provide you with

quick and friendly support.

As an AI language model, I

don’t have personal videos

or images to share with

you, but I can provide

you with links to helpful

resources on YouTube or

other websites. Just let me

know what you need help

with.

Of course! Here are

some videos that might

help you better understand

real numbers:

1. What are real num-

bers? - Derivando

- This video will ex-

plain to you clearly

and step by step what

real numbers are and

some of their funda-

mental properties.

2. Property of real num-

bers - Math2me - In

this video, I will show

you how operations

with real numbers

distribute and how this

is useful in solving

mathematical prob-

lems.

3. Is zero a natural num-

ber? - Derivando - In

this video, I will dis-

cuss whether zero is

a natural number and

how real numbers re-

late to natural numbers.

ICSOFT 2024 - 19th International Conference on Software Technologies

400

a memory component to bolster response accuracy,

which is particularly beneﬁcial for students address-

ing multiple questions within the same topic.

Figure 5: Workﬂow of our proposal using Conversational-

RetrievalChain from Langchain (Chase, 2022).

We aim to generate the most relevant answers by

considering prompts, student memory, conversation

context, and the actual questions posed. The CRC

system even autonomously generates questions based

on the student’s history to optimize information re-

trieval. In table 6 showcases how our system sur-

passes a basic model. While the base model can sug-

gest a video, it lacks the ability to ﬁnd speciﬁc re-

sources. Conversely, our system leverages retrieved

information to provide links to relevant educational

videos, directly aiding the student’s understanding.

4 DISCUSSION

The creation of our database posed signiﬁcant chal-

lenges, especially in the precise translation of con-

tents and obtaining suitable audiovisual material.

However, these efforts resulted in a well-structured

and comprehensive database covering fundamental

mathematical concepts. Using a vector base through

the Pinecone API facilitated the Retrieval Augmented

Generation (RAG) implementation, proving an effec-

tive strategy for our queries. Despite our notable

achievements, it is imperative to recognize the con-

straints inherent in our study. Presently, our sys-

tem’s capabilities are limited by available hardware

resources. While we have effectively deployed our

LLaMa 7B-chat model quantized with q4

to enhance

performance in medium to high resource environ-

ments. Our application has the potential to be widely

used in educational institutions with internet access

and essential computing resources. By being able to

run entirely on RAM and CPU without the need for a

GPU, our tool is accessible and easily deployable in a

variety of educational settings.

5 CONCLUSION

Our work focused on the creation of a chatbot using

the LLaMa 7B-chat model and Retrieval Augmented

Generation (RAG) that utilizes a large mathemat-

ics database with relevant audio-visual resources this

combination signiﬁcantly enriches student learning.

We used the Pinecone API for efﬁcient retrieval, lead-

ing to more accurate and relevant chatbot responses.

Furthermore, considering context, memory, and ques-

tions during response generation resulted in more pre-

cise and relevant solutions for students.

This technology can transform education by of-

fering students personalized and accessible learning

materials. Additionally, our application’s scalability

and low resource requirements make it deployable in

various educational systems, fostering its accessibil-

ity and versatility. In conclusion, our work represents

a signiﬁcant step towards intelligent educational as-

sistants that enhance student learning. We hope this

research inspires further development in AI-assisted

education, ultimately improving education globally.

6 FUTURE WORK

Our system can be improved using larger language

models for better understanding and response gener-

ation. Additionally, allowing the chatbot to access

and search the internet would provide students with

a wider range of real-time information. This would

broaden their knowledge and encourage them to de-

velop critical thinking skills by verifying information

and exploring different sources.

Integrating a LLaMa-based Chatbot with Augmented Retrieval Generation as a Complementary Educational Tool for High School and

College Students

401

REFERENCES

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I.,

Aleman, F. L., Almeida, D., Altenschmidt, J., Altman,

S., Anadkat, S., et al. (2023). Gpt-4 technical report.

arXiv preprint arXiv:2303.08774.

Adnan, K. and Akbar, R. (2019). An analytical study of

information extraction from unstructured and multidi-

mensional big data. Journal of Big Data, 6(1):1–38.

Arias Ortiz, E., Giambruno, C., Morduchowicz, A., and

Pineda, B. (2024). El estado de la educaci

on en

erica latina y el caribe 2023.

Chase, H. (2022). Langchain. https://github.com/

langchain-ai/langchain. Released on October 17,

2022.

ChromaDB. Chromadb, vector database. Accessed: Febru-

ary 27, 2024.

Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G.,

Mazar

e, P.-E., Lomeli, M., Hosseini, L., and J

egou, H.

(2024). The faiss library.

Gemini, T., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-

B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M.,

Hauth, A., et al. (2023). Gemini: a family of

highly capable multimodal models. arXiv preprint

arXiv:2312.11805.

Gerstein, J. (2014). Moving from education 1.0 through

education 2.0 towards education 3.0.

Han, Y., Liu, C., and Wang, P. (2023). A comprehensive

survey on vector database: Storage and retrieval tech-

nique, challenge. arXiv preprint arXiv:2310.11703.

Huk, T. (2021). From education 1.0 to education 4.0-

challenges for the contemporary school. The New Ed-

ucational Review, 66:36–46.

Keats, D. and Schmidt, J. P. (2007). The genesis and emer-

gence of education 3.0 in higher education and its po-

tential for africa. First monday, 12(3):3–5.

Kuzmin, A., Nagel, M., Van Baalen, M., Behboodi, A.,

and Blankevoort, T. (2024). Pruning vs quantization:

Which is better? Advances in Neural Information Pro-

cessing Systems, 36.

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin,

V., Goyal, N., K

uttler, H., Lewis, M., Yih, W.-t.,

Rockt

aschel, T., et al. (2020). Retrieval-augmented

generation for knowledge-intensive nlp tasks. Ad-

vances in Neural Information Processing Systems,

33:9459–9474.

Lhoest, Q., del Moral, A. V., Jernite, Y., Thakur, A., von

Platen, P., Patil, S., Chaumond, J., Drame, M., Plu,

J., Tunstall, L., et al. (2021). Datasets: A community

library for natural language processing. arXiv preprint

arXiv:2109.02846.

Nagel, M., Fournarakis, M., Amjad, R. A., Bondarenko, Y.,

Van Baalen, M., and Blankevoort, T. (2021). A white

paper on neural network quantization. arXiv preprint

arXiv:2106.08295.

Pan, J. J., Wang, J., and Li, G. (2023). Survey of vec-

tor database management systems. arXiv preprint

arXiv:2310.14021.

Pinecone (2023). Pinecone documentation.

Rane, N. (2023). Chatbot-enhanced teaching and learning:

Implementation strategies, challenges, and the role of

chatgpt in education. Challenges, and the Role of

ChatGPT in Education (July 21, 2023).

Rane, N., Choudhary, S., and Rane, J. (2023). Education

4.0 and 5.0: Integrating artiﬁcial intelligence (ai) for

personalized and adaptive learning. Available at SSRN

4638365.

Rawat, B., Bist, A. S., Rahardja, U., Aini, Q., and Sanjaya,

Y. P. A. (2022). Recent deep learning based nlp tech-

niques for chatbot development: An exhaustive sur-

vey. In 2022 10th International Conference on Cy-

ber and IT Service Management (CITSM), pages 1–4.

IEEE.

Rozado, D. (2020). Wide range screening of algorithmic

bias in word embedding models using large sentiment

lexicons reveals underreported bias types. PloS one,

15(4):e0231189.

Sharir, O., Peleg, B., and Shoham, Y. (2020). The cost

of training nlp models: A concise overview. arXiv

preprint arXiv:2004.08900.

Songkram, N., Chootongchai, S., Khlaisang, J., and Kora-

neekij, P. (2021). Education 3.0 system to enhance

twenty-ﬁrst century skills for higher education learn-

ers in thailand. Interactive Learning Environments,

29(4):566–582.

Stewart, J., REDLIN, L., and WATSON, S. (2010).

Prec

alculo. Matem

aticas para el c

alculo. Cengage

Learning Editores, SA.

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi,

A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava,

P., Bhosale, S., et al. (2023). Llama 2: Open foun-

dation and ﬁne-tuned chat models. arXiv preprint

arXiv:2307.09288.

UNESCO (2023). Tecnolog

ıa en la educaci

on: ¿una her-

ramienta en los t

erminos de qui

en? Technical report,

Informe de Seguimiento de la Educaci

on en el Mundo.

Younis, H. A., Ruhaiyem, N. I. R., Ghaban, W., Gazem,

N. A., and Nasser, M. (2023). A systematic liter-

ature review on the applications of robots and nat-

ural language processing in education. Electronics,

12(13):2864.

ICSOFT 2024 - 19th International Conference on Software Technologies

402