A Rag-Based Question Answering System Proposal For Understanding Islam: Mufassirqas LLM

Cite this paper (APA):
Alan, A.Y., Karaarslan, E., Aydin, O. (2024). A RAG-based Question

Answering System Proposal for Understanding Islam: MufassirQAS LLM
A RAG-based Question Answering System Proposal for Understanding Islam:

MufassirQAS LLM
Ahmet Yusuf Alan, Enis Karaarslan*, Omer Aydin
* [email protected]
Abstract:
Challenges exist in learning and understanding religions, such as the complexity and depth of
religious doctrines and teachings. Chatbots as question-answering systems can help in solving these
challenges. LLM chatbots use NLP techniques to establish connections between topics and
accurately respond to complex questions. These capabilities make it perfect for enlightenment on
religion as a question-answering chatbot. However, LLMs also tend to generate false information,
known as hallucination. Also, the chatbots' responses can include content that insults personal
religious beliefs, interfaith conflicts, and controversial or sensitive topics. It must avoid such cases
without promoting hate speech or offending certain groups of people or their beliefs. This study uses
a vector database-based Retrieval Augmented Generation (RAG) approach to enhance the accuracy
and transparency of LLMs. Our question-answering system is called “MufassirQAS''. We created a
database consisting of several open-access books that include Turkish context. These books contain
Turkish translations and interpretations of Islam. This database is utilized to answer religion-related
questions and ensure our answers are trustworthy. The relevant part of the dataset, which LLM also
uses, is presented along with the answer. We have put careful effort into creating system prompts that
give instructions to prevent harmful, offensive, or disrespectful responses to respect people's values
and provide reliable results. The system answers and shares additional information, such as the page
number from the respective book and the articles referenced for obtaining the information.
MufassirQAS and ChatGPT are also tested with sensitive questions. We got better performance with
our system. Study and enhancements are still in progress. Results and future works are given.
Keywords: Question Answering Systems, Chatbots, Generative Artificial Intelligence, Large

Language Model, Retrieval Augmented Generation, Quran
1
1. INTRODUCTION
There exist challenges in learning and understanding religions, such as the presence of complexity
and depth of religious doctrines and teachings.
Chatbots, as question-answering systems, are becoming widely used as they provide detailed natural
responses to users' inquiries on a given topic. They maintain a conversational tone and ensure a
logical flow of conversation. The strength of the LLM chatbot lies in its ability to establish
connections between topics and accurately respond to complex questions about sensitive topics.
These capabilities make it perfect to be used in enlightenment on religion. However, a drawback of
such systems is the tendency for LLMs to generate false information, known as hallucination. The
research questions are as follows:
❖ Can we use Retrieval Augmented Generation (RAG) approach to prevent hallucination in

LLMs?
❖ Can we deploy a system that utilizes this concept for understanding Islam?
2. FUNDAMENTALS
In this section, we will commence by disseminating fundamental information concerning the large
language models (LLM) that serve as the primary driving force behind contemporary chatbots, which
are progressively aligning themselves with the domain of artificial intelligence. Moreover, we will
delve into the vector-based databases in Retrieval Augmented Manufacturing (RAG). Subsequently,
we will address the potential challenges encountered in the development of this language model.
Additionally, there will be a shared section on the Islamic sources to be used in this study.
2.1. Large Language Model and Vector Based Databases
LLMs are a type of artificial intelligence system that can process and generate natural language texts.
They are trained on massive amounts of text data, such as books, articles, web pages, social media
posts, and more, using deep neural networks. LLMs can learn the patterns and structures of natural
language from the data, and use them to perform various tasks, such as answering questions,
summarizing texts, translating languages, and writing essays.
Despite their remarkable capabilities, LLMs are not without limitations. Concerns regarding
potential biases embedded within training data, factual inaccuracies, and the absence of ethical
2
consideration. Also, they lack explainability and transparency. This makes it nearly impossible to
understand how they generate responses and how the LLM works.
2.2. Vector Databases and RAG as a Solution
Retrieval-augmented generation (RAG) can be used for knowledge-intensive NLP tasks [4]. RAG is
a technique for enhancing the accuracy and reliability of generative AI models with facts fetched
from predetermined external sources.
LLMs have some limitations, such as; presenting false information when they do not have the
answer, presenting out-of-date or generic information, and creating a response from non-authoritative
sources in train data. RAG addresses these challenges by redirecting the LLM to retrieve relevant
information from authoritative, predetermined knowledge sources. This way, LLM can use the
retrieved data as context for generating a response that is more relevant, accurate, and useful in
various contexts. RAG applications potentially provide user transparency by revealing the sources of
the retrieved data, offering insight into how the LLM generates its responses.
2.3. Islamic sources
First of all, we can start by learning the word Mufassir and its meaning, which is also mentioned in
the title of this study. It is an Arabic word and the name given to theologians who deal with tafsir.
Tafsir is mostly used to explain what the Quran's words, compositions and sentences mean. Details
about the Tafsir are given below. Now, let's learn the concepts we need to know within the scope of
this study, starting from the most basic source of Islam. The Quran is the name of the word that was
sent down to the prophet of the Islamic religion through revelation, written on the Mushafs, spread
and transmitted through the tongue, and worshipped by reading it (Temel, 2018, p. 106). There are
different approaches to the root of the name of the Quran. Revelation, on the other hand, is the
creator's general informing of the ways of action of beings, and in particular, his transmission to his
prophets of all the orders, prohibitions and news that he wants to convey to people, secretly and
rapidly, in a direct or unmediated manner (Çelik, 2013, p. 22). As it is known, the Quran reached the
prophet Muhammad through the angel of revelation, Gabriel. It started to be downloaded in the
month of Ramadan, one of the Hijri months, and in the Night of al-Qadr (Danacı, 2020). The Quran
was transferred from the Preserved Tablet (Levh-i Mahfuz) to the angel of revelation in a manner
whose condition we do not know, and through him, it was sent down to the prophet Muhammad at
various time intervals (Keskinoğlu, 2012, p. 12). While the Prophet Muhammad was alive, many of
3
the Muslims of the period who believed in him and followed the Prophet's path memorized the whole
part of the Quran, so that the Quran was preserved both in writing, on tablets and pages, and in
memory by memorization. We should also draw attention to the great efforts made by the Prophet
Muhammad in educating the Quran and passing it on to future generations (Öge, 2019, p. 27,28).
Since it is impossible to translate the Quran verses into another language without missing any
meaning, the translation of the Quran into other languages is called "Meal". This refers to the
approximate meaning of the verse. On the other hand, Tafsir (Interpretation) means to declare
something, to discover something, to reveal something covered up; It means explaining what is
meant by a word whose meaning is unclear and whose meaning is difficult to understand. Tafsir is
mostly used to explain what the Quran's words, compositions and sentences mean. Tafsir is a branch
of science. In this branch of science, the Quran is interpreted, and the meanings of words and
sentences, their provisions and wisdom are explained (Diyanethaber, 2021). On the other hand,
Hadith is the antonym of “Kadîm”. As it is known, Kadim means "Old". Over time, Hadith's
meaning is "News" and over time it is created from the infinitive Tahdîs. The word Hadith gained a
different meaning in Islam. The words of the Prophet Muhammad are called "el-ehâdîsü'l-kavliyye",
his actions are called "el-ehâdîsü'l-fi'liyye", and the things he approves (takrir) are called
"el-ehâdîsü't-takrîriyye" (Ebü'l-Bekā, pp. 370, 402).
These are the most important resources for people to understand Islam as a faith and to pass it on to
the next generations. For this reason, it has become important to learn Islam from the right sources
and at the same time to convey them using today's technological blessings.
3. RELATED WORKS
Different use cases for Generative AI and RAG are available in the literature. The use of
vector-based databases also finds a place in the literature. LLMs have been shown to experience
hallucinations in their responses to instruction-following tasks. Fine-tuning with new data may not
prevent hallucinations and may be costly. Therefore, RAG was introduced to provide factual
information and adapt to new information (Borgeaud et al. 2022). RAG uses a vector database. When
a query is made to RAG, the most relevant data from this database is selected. This data is included
in the prompt for LLMs. This process can combine existing information without requiring constant
fine-tuning and provides context-based information, increasing the clarity and relevance of
responses. RAG provides the LLM with the ability to gain a focused understanding vis-à-vis a query.
4
There are studies in the literature where RAG is used in NLP tasks (Borgeaud et al. 2022; Al
Ghadban et al. 2023; Mallen et al. 2023; Ram et al. 2023; ). Pal et al. sought to create unique and
enjoyable music using Generative AI. When provided with a starting bar of music, the initial
Discriminatory network, incorporating Support Vector Machines and Neural Nets, selects a note or
chord to guide the subsequent bar. Building upon this chosen chord or note, a Generative Net, which
includes Generative Pretrained Transformers (GPT-2) and LSTMs, then generates the entire bar of
music. Their innovative two-step approach aims to closely emulate the process of real music
composition to enhance the authenticity of the generated music (Pal, 2020).
Kulkarni et al. proposed a Retrieval-Augmented Generator (RAG) approach for constructing a

chatbot capable of addressing user queries using Frequently Asked Questions (FAQ) data. The
chatbot employs an open API-based paid ChatGPT model as a Language Model (LLM), recognizing
the potential to optimize LLM token usage and cost by leveraging previously retrieved-context for
specific query patterns. The authors employ Reinforcement Learning (RL) to optimize the number of
LLM tokens within the constraints of a fixed retrieval model and LLM. The proposed RL-based
optimization, coupled with a similarity threshold, achieves notable cost savings with a slight
enhancement in accuracy. The study emphasizes the general applicability of the RL approach beyond
FAQ chatbots to any existing RAG pipeline (Kulkarni et al., 2024).
There are ChatGPT-based studies in the literature such as [1-2] and online chatbot services such as
QuranGPT (https://qurangpt.live/). The validity of the services based on ChatGPT is argued as these
services do not give references [3]. Our work differentiates from these works by fine-tuning the user
prompt and giving references.
4. METHOD
In this study, we implemented a retrieval augmented generation (RAG) system to enhance the
accuracy and transparency of large language models (LLMs) for natural language question
answering tasks. We followed the RAG architecture and the proposed system is shown in Figure 1.
Study consists of the following components: knowledge base, vector store, question answering LLM,
user interface. The algorithm of Question answering with the RAG system is shown in Algorithm 1.
The proposed system is formed of following functions which is described in the following sections.
5
Figure 1. RAG based System
6
4.1. Forming knowledgebase
We use the pdf format of open access books to correctly acquire metadata of books. Following
dataset are used to form the vector store:
- “Kuran yolu Türkçe meal ve tefsir” (The Way of the Quran -Turkish translation and
interpretation) [5]
- Kütüb-i Sitte [6]
- “İslam İlmihali” (Islamic Catechism) [7]
4.2. Creating Vector Store
A vector database is a specific type of database designed to store and retrieve vectors. Vectors are
lists of numbers that represent data in a multi-dimensional space. Vector search methods can be used
to find similar data by querying for the nearest vectors in the space. This capability is particularly
useful for applications that require fast and accurate data matching based on similarity.
In order to enable vector search, the text in the knowledge base needs to be divided into fixed-length
segments called chunks. Since the context of religion issues could be long we assign a higher chunk
length. It's important to assign a value to indicate how much overlap each chunk has with the
previous one, known as the chunk overlap. In order to ensure that the language model understands
sensitive content and maintains logical connections between parts of the document, we assign a
higher value for chunk overlap. Once these values are determined, the chunks should be embedded
using the appropriate language model that we utilize for our question-answering system.
4.3. Vector Search
After the user submits a question, the system generates an embedding of the user's question. This
embedding helps the LLM to comprehend the meaning and context of the query. By using these
embedded questions, the LLM can search for relevant parts with the context and assigning a
similarity score to each part. The system then selects these chunks to form a knowledge context for
answering the question. We assign the value of "n" as the number of chunks with the highest
similarity score to the question. Considering the abundance of cross-references and diverse
explanations surrounding religious matters, we prioritize assigning a higher value to "n." This
ensures that the LLM receives a more extensive amount of information regarding the question.
7
4.4. Tuning Prompt
After gathering the context with a question, we need to create a prompt to send to the LLM by
combining system prompt, user question and chunks. During this process, we should inform the
LLM about the restrictions, including the desired response format and the tone in system prompt.
Also, it is important to specify when the system should not generate a response.
A system prompt serves as a natural language instruction that guides a large language model in
performing a specific task. These system prompts are crucial for LLMs as they enable users to tap
into the extensive knowledge and capabilities of the models without requiring fine-tuning on specific
datasets or domains.
In this study we work on creating system prompts with care, ensuring they provide instructions that
prevent the generation of harmful, offensive, or disrespectful responses. The chatbot needs to avoid
insulting personal religious beliefs, handle interfaith conflicts respectfully, and approach
controversial or sensitive topics without promoting hate speech or offending certain groups of people
or their beliefs.
5. IMPLEMENTATION
We used the ChatGPT3.5 turbo 16k context window version. ChatGPT 3.5 can use a maximum of 32
thousand tokens. When we set the chunk size to 2000, we can provide up to 15 chunks of
information. We used OpenAI embeddings and the temperature as 0.5. Langchain toolkit
(https://www.langchain.com) for RAG. Memory vector store is used for now. We use chunk Size
2000, chunk overlap 100. We give the top 5 chunks as a result of the search. Flowise
(https://github.com/FlowiseAI/Flowise) is added to show the reference document and vector search
results which is used to answer the questions.
In order to create the MufassirQAS, we require a platform that combines the mentioned steps and
includes a user interface (UI) for interacting with users. This UI should allow users to ask questions,
view answers, and relevant information that they use to answer questions. To develop the
proof-of-concept (POC) version of this platform, we used a Colab Notebook that utilizes Chroma as
a vector database and langchain as the RAG tool provider. However, for the complete platform, we
opted to use an existing open-source platform that already possesses the necessary capabilities, rather
8
than building it from scratch. Flowise is one such tool, which enables the creation of a chatbot with
RAG capabilities using Langchain. Initially, we hosted Flowise on the Railway cloud-host, but due
to issues with API calls, we switched to HuggingFace spaces as the cloud-host provider.
Sample screenshots of the implementation are given in Figure 2 and 3. Turkish versions of the
samples are given in the Appendix. The chunk for the question “According to Islam, are women or
men superior?” is given in Figure 4. The answers of the systems are discussed in the next section.
9
Figure 2. Answer to the Vaccination Question with (a) ChatGpt (b) MufassirQAS
10
Figure 3. Answer to the Man Woman Equality Question (a) ChatGpt (b) MufassirQAS
11
Figure 4. Example chunk for question “According to Islam, are women or men superior?”
6. RESULTS AND DISCUSSION
In many general religion-related questions, ChatGPT and MufassirQAS provide similar answers.
This is expected because ChatGPT has been trained on a vast amount of data, so it likely includes
similar religion-related questions and answers. However, MufassirGPT demonstrates its strength
when faced with questions like "Is there evolution according to Islam?" (see Figure 5), where
12
ChatGPT tends to give uncertain responses and consider the question complex and open to
discussion. In contrast, MufassirQAS immediately answers with a clear statement and provides an
explanation.
Figure 5. Answer to the Evolution Question
Another distinction between MufassirQAS and ChatGPT is that MufassirQAS tends to provide more
references to hadiths and verses, as instructed in the system prompt. For example, when asked
"According to Islam, are women or men superior?"; ChatGPT refers to an expression from the Quran
without clearly indicating its source (see Figure 3.a). Meanwhile, MufassirQAS not only shows the
users the knowledge source but also directly explains the matter using three Quranic verses (see
Figure 3.b). Overall, we can consider MufassirQAS a reliable solution for avoiding vague and
unclear responses to questions on Islam.
7. LIMITATIONS OF THE STUDY
We understand that there are limitations, including the possibility of biases in the dataset and the
difficulty of ensuring religious sensitivity in the responses. We are also aware of the security
vulnerabilities of LLMs, which can lead to unexpected responses to prompts. Moreover, we
13
recognize the challenge of ensuring religious sensitivity in the responses and are consistently refining
our approach. We are actively working on mitigating biases and improving religious context
awareness.
8. CONCLUSION
This study showed that the Retrieval Augmented Generation (RAG) approach can prevent
hallucination in LLMs. We developed the MufassirQAS system that utilizes this concept for
understanding Islam. We conducted tests on MufassirQAS using various parameters and questions.
MufassirQAS is highly effective in finding relevant information in vector databases and generating
relevant user responses. It offers insights based on the information it retrieves. Surprisingly, we
encounter similar responses from ChatGPT and MufassirQAS in many cases, even though
MufassirQAS uses a smaller dataset to answer user questions. The main distinction is that
MufassirQAS tends to provide more definite answers, whereas ChatGPT tends to give vague and
uncertain responses. Another exciting aspect is that MufassirQAS often cites the Quran and hadith
with specific sources, while ChatGPT often mentions their presence without providing any sources.
Furthermore, we noticed that, in some instances, MufassirQAS responses contain more details than
what was available in the dataset. This is because MufassirQAS uses the knowledge of source LLM
(ChatGPT) to bridge gaps in the given data.
However, there are limitations to MufassirQAS. LLM models, in general, are excellent at
summarizing significant texts. However, when we set the chunk size too large, the system needs help
with summarization and knowledge connection in response text. Additionally, the system encounters
difficulties when dealing with many chunks, as it needs help connecting different pieces of retrieved
knowledge. This issue arises from the instruction to utilize given knowledge in responses, and as a
result, the system sometimes needs to maintain knowledge context in its responses.
ChatGPT is a large and well-designed language model (LLM) with extensive knowledge on various
topics. The difference will be more noticeable in smaller models that the model does not extensively
train on religious contexts or specifically on reducing hallucinations. However, there should still be a
threshold on how small the model can be to retrieve relevant data effectively. There is a limitation of
RAG when reducing the model size: it requires a comprehensive understanding of both the data and
the question to generate suitable representations for retrieving similar information. In the future, we
14
plan to test this limitation and develop smaller and more sustainable models that are highly efficient
and specialized in understanding religion and answering related questions.
ACKNOWLEDGMENT
The study is a part of LLM joint research between MSKU Metaverse Lab and Celal Bayar
University.
FUNDING
This research received no specific grant from any funding agency in the public, commercial, or
not-for-profit sectors.
AUTHORS` CONTRIBUTIONS
All authors have participated in drafting the manuscript. All authors read and approved the final
version of the manuscript. All authors contributed equally to the manuscript and read and approved
the final version of the manuscript.
CONFLICT OF INTEREST
The authors certify that there is no conflict of interest with any financial organization regarding the
material discussed in the manuscript.
REFERENCES
[1] Alnefaie, E. A. S., & Alsalka, M. A. (2023, October). Is GPT-4 a Good Islamic Expert for
Answering Quran Questions?. In Proceedings of the 35th Conference on Computational Linguistics
and Speech Processing (ROCLING 2023) (pp. 124-133).
[2] Sembok, T. M., & Wani, S. (2023, October). Is ChatGPT not Appropriate for Religious Use?. In
International Visual Informatics Conference (pp. 595-605). Singapore: Springer Nature Singapore.
[3] Quran GPT and the dangers of propagating a homogenized version of Islam, retrieved on
30.12.2023.
https://english.alarabiya.net/views/2023/08/09/Quran-GPT-and-the-dangers-of-propagating-a-homog
enized-version-of-Islam
15
[4] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., ... & Kiela, D. (2020).
Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural
Information Processing Systems, 33, 9459-9474.
[5] Karaman, H., Çağrıcı, M., Dönmez, İ. K., & Gümüş, S. (2020). Kuran Yolu Türkçe Meal Ve
Tefsir-2 (The Way of the Quran: Turkish Translation and Interpretation). Diyanet İşleri Başkanlığı,
https://yayin.diyanet.gov.tr/e-kitap/kuran-yolu-turkce-meal-ve-tefsir-1/prof-dr-hayrettin-karaman/kur
an-kitapligi/3788
[6] Kandemir, M. Y. (2008). “Kütüb-i Sitte”. DİA, İstanbul 2008, cilt: 27, s. 6-7.
[7] ŞENTÜRK, L., & YAZICI, S. (2012). İslâm İlmihali, Diyanet İşleri Başkanlığı Yayınları.
[8] Yazır, E. M. H. (2002). Kur’ân-ı Kerîm’in Yüce Meâli. Cev: K. Yayla. İstanbul: Merve yayıncılık.
[9] Temel, N. (2018). Kıraat ve Tecvid Istılahları. (4.b.) İstanbul: İFAV.
[10] Çelik, Ö. (2013). Tefsir Usûlü ve Tarihi. İstanbul, Türkiye: Kampanya Kitapları.
[11] Keskinoğlu, O. (2012). Nüzûlünden Günümüze Kur'an-ı Kerim Bilgileri. Ankara: T.D.V.
[12] Öge, A. (2019). 18. Yüzyıl Osmanlı Âlimlerinden Yusuf Efendizâde'nin Kıraat İlmindeki Yeri.
İstanbul: Hacıveyiszade İlim ve Kültür Vakfı Yayınları.
[13] Danacı, S. (2020). Kur’ân-ı Kerim’in Metinleşme Süreci ve Sûrelerin Tertibine Dair Görüşlerin
Değerlendirilmesi. Çankırı Karatekin Üniversitesi Karatekin Edebiyat Fakültesi Dergisi, 8(1),
138-166.
[14] Diyanethaber (2021). Meal ve Tefsir ne demektir? Diyanethaber.com.tr Web sitesi.

https://www.diyanethaber.com.tr/meal-ve-tefsir-ne-demektir Accessed: 25.01.2024
[15] Ebu’l-Bekâ elKefevî. el-Külliyyât.
[16] Kulkarni, M., Tangarajan, P., Kim, K., & Trivedi, A. (2024). Reinforcement Learning for
Optimizing RAG for Domain Chatbots. arXiv preprint arXiv:2401.06800.
[17] Pal, A., Saha, S., & Anita, R. (2020). Musenet: Music Generation using Abstractive and
Generative Methods. International Journal of Innovative Technology and Exploring Engineering,
9(6), 784-788.
16
[18] Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., ... & Sifre, L.
(2022, June). Improving language models by retrieving from trillions of tokens. In International
conference on machine learning (pp. 2206-2240). PMLR.
[19] Al Ghadban, Y., Lu, H. Y., Adavi, U., Sharma, A., Gara, S., Das, N., ... & Hirst, J. E. (2023).
Transforming Healthcare Education: Harnessing Large Language Models for Frontline Health
Worker Capacity Building using Retrieval-Augmented Generation. medRxiv, 2023-12.
[20] Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D., & Hajishirzi, H. (2023, July). When not
to trust language models: Investigating effectiveness of parametric and non-parametric memories. In
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume
1: Long Papers) (pp. 9802-9822).
[21] Ram, O., Levine, Y., Dalmedigos, I., Muhlgay, D., Shashua, A., Leyton-Brown, K., & Shoham,
Y. (2023). In-context retrieval-augmented language models. arXiv preprint arXiv:2302.00083.
17
APPENDIX - TURKİSH VERSIONS OF THE SAMPLES
Figure 2. Answer to the Vaccination Question in Turkish (a) ChatGpt (b) MufassirQAS
18
Figure 3. Answer to the Man-Woman Equality Question (a) ChatGpt (b) MufassirQAS
19
Figure 4. Answer to the Evolution Question
20
Figure 5. Example chunk for the question “What are the conditions to go to heaven?”
21

A Rag-Based Question Answering System Proposal For Understanding Islam: Mufassirqas LLM

Uploaded by

Copyright:

Available Formats

A Rag-Based Question Answering System Proposal For Understanding Islam: Mufassirqas LLM

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Rag-Based Question Answering System Proposal For Understanding Islam: Mufassirqas LLM

Uploaded by

Copyright:

Available Formats

Cite this paper (APA):

Alan, A.Y., Karaarslan, E., Aydin, O. (2024). A RAG-based Question

A RAG-based Question Answering System Proposal for Understanding Islam:

Ahmet Yusuf Alan, Enis Karaarslan*, Omer Aydin

Keywords: Question Answering Systems, Chatbots, Generative Artificial Intelligence, Large

❖ Can we use Retrieval Augmented Generation (RAG) approach to prevent hallucination in

2.1. Large Language Model and Vector Based Databases

2.2. Vector Databases and RAG as a Solution

2.3. Islamic sources

Kulkarni et al. proposed a Retrieval-Augmented Generator (RAG) approach for constructing a

Figure 1. RAG based System

4.1. Forming knowledgebase

- Kütüb-i Sitte [6]

- “İslam İlmihali” (Islamic Catechism) [7]

4.2. Creating Vector Store

4.3. Vector Search

4.4. Tuning Prompt

6. RESULTS AND DISCUSSION

Figure 5. Answer to the Evolution Question

7. LIMITATIONS OF THE STUDY

[9] Temel, N. (2018). Kıraat ve Tecvid Istılahları. (4.b.) İstanbul: İFAV.

[14] Diyanethaber (2021). Meal ve Tefsir ne demektir? Diyanethaber.com.tr Web sitesi.

[15] Ebu’l-Bekâ elKefevî. el-Külliyyât.

APPENDIX - TURKİSH VERSIONS OF THE SAMPLES

Figure 4. Answer to the Evolution Question

You might also like