Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components

Ahmad Shabana; Ahmed Alhusayni; Ibrahim Aldailami; Israa Hamdine; Jumana Alsubhi; Mohammad D. Alahmadi; Suhayb Khayyat; Yazeed Iskandar

REVIEW 5 major objections 5 minor 21 references

Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A component-by-component test of Arabic retrieval-augmented generation finds that sentence-aware chunking, BGE-M3 embeddings, the bge-reranker-v2-m3, and Aya-8B together give the best measured RAG quality.

desk verdict Useful Arabic RAG benchmark, but the headline claims rely on an unnamed LLM judge and no significance tests. read the letter →

arxiv 2506.06339 v1 pith:X325OLDC submitted 2025-06-01 cs.IR cs.AIcs.CL

Jumana Alsubhi , Mohammad D. Alahmadi , Ahmed Alhusayni , Ibrahim Aldailami , Israa Hamdine , Ahmad Shabana , Yazeed Iskandar , Suhayb Khayyat This is my paper

classification cs.IRcs.AIcs.CL

keywords Retrieval-AugmentedGenerationArabicNLPchunkingstrategiesembeddingmodelsrerankingRAGASevaluationmultilingualembeddingsquestionanswering

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Retrieval-augmented generation (RAG) lets a language model answer questions by first pulling relevant passages from a document collection, but most component recommendations come from English-centric studies. This paper asks which off-the-shelf chunking, embedding, reranking, and generation choices work best when the documents and questions are in Arabic. Across six Arabic datasets and four RAGAS quality metrics, the authors report that sentence-aware chunking beats fixed-size, recursive, and semantic chunking; that BGE-M3 and Multilingual-E5-large outperform Arabic-specialized embedding models; and that the bge-reranker-v2-m3 plus Aya-8B further improve faithfulness and answer relevancy. If these rankings are right, they turn Arabic RAG from guesswork into a concrete default configuration.

What carries the argument

The load-bearing device is the controlled component-wise evaluation: keep every pipeline stage fixed except the one under test, vary it, and score the result with RAGAS's four metrics—context precision, context recall, answer faithfulness, and answer relevancy. Sentence-aware chunking is the segmentation choice that carries the core result because it preserves complete linguistic units, and the RAGAS metric suite is what makes the four component comparisons directly comparable.

What would settle it

Take a sample of outputs from the winning configuration (sentence-aware chunking, BGE-M3, reranker, Aya-8B) and have it scored by a different named judge model and by human Arabic-speaking annotators; if the new scores do not reproduce the reported rankings, the component choices are an artifact of the original automated evaluator.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is a practical ranking of Arabic RAG components. Sentence-aware chunking, which cuts at sentence boundaries, gives the highest average overall score (74.78) and leads in four of six datasets, while semantic chunking—despite its theoretical appeal—trails every other strategy. BGE-M3 (70.99) and Multilingual-E5-large (70.31) are the strongest embedders, ahead of both Arabic-specific models and other multilingual options. Adding the bge-reranker-v2-m3 raises the average score from 70.99 to 74.15, with the largest gains on ARCD and QA4MRE and a small decrease on SaudiWiki. Aya-8B outscores StableLM-1.6B on every dataset, by the widest margin on inference-heavy and semantically dense text such as QA4MRE and Quran Tafseer.

Load-bearing premise

The rankings assume the automated RAGAS scores are valid, unbiased measures of Arabic answer quality, even though the scoring model is not named and no human judgment was used for faithfulness or relevancy.

Editorial extensions

If this is right

Sentence-aware chunking should be the default segmentation method for Arabic RAG, since it won overall and led in four of six datasets.
Builders should prefer BGE-M3 or Multilingual-E5-large over Arabic-specific embedding models for general Arabic retrieval.
Adding the bge-reranker-v2-m3 reranker is worth it for noisy or inference-heavy Arabic corpora, though it can slightly hurt already-clean structured sources like SaudiWiki.
Aya-8B is the better generator of the two tested for Arabic, especially in reasoning-heavy and semantically dense text.
Together these choices form a concrete baseline configuration that future Arabic RAG work can compare against.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the implicit RAGAS judge is itself an LLM with weaker Arabic than the generator, the faithfulness and relevancy rankings could partly measure judge preference rather than true quality; a human-evaluation pass would settle this.
The pattern that multilingual contrastively trained embeddings beat Arabic-specialized small models suggests that scale and cross-lingual pretraining matter more than language-specific tuning at current model sizes, a hypothesis worth testing on other low-resource languages.
The failure of semantic chunking across Arabic datasets is surprising given its theoretical appeal; it may reflect noisy clustering on morphologically rich text and suggests a testable hybrid that constrains clusters to sentence boundaries.
The reranker results point to a domain-specific research direction: training or fine-tuning rerankers on legal, medical, or religious Arabic corpora could extend the gains seen in QA4MRE and Quran Tafseer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a solid, useful benchmark for Arabic RAG component selection, and the first real head-to-head of chunking strategies on Arabic that I know of. The descriptive results—sentence-aware chunking winning on average, bge-m3 and e5-large leading among embeddings, reranker helping on messy datasets—are plausible and consistent with what we see in English. If you build Arabic RAG systems, this gives you a sensible starting configuration.

What's new and good: six public Arabic datasets, four chunking strategies, six embeddings, one reranker, two generators, all measured with RAGAS. The tables are readable and per-dataset breaks are useful. The authors also acknowledge in Section 4 that faithfulness and answer relevancy are automated metrics without human judgment, which is more honest than most benchmarks of this type.

The soft spots are real, though. The biggest one: the RAGAS judge model is never named. Faithfulness and answer relevancy are LLM-based metrics, and if the judge is Aya-8B (the same model being tested in RQ4), that comparison is self-preferential. If the judge is a model with weak Arabic ability, every ranking is suspect. Section 3.6 mentions 'manually curated ground-truth answers' and 'semi-automated evaluation scripts with expert validation,' but that doesn't fix the LLM-judge problem. The stress-test note is on point.

Second, there are no error bars, confidence intervals, or significance tests. The differences between chunking strategies at the top end are small—74.78 vs 69.41 vs 69.13—and could easily be noise. The word 'significantly' is used in the abstract and conclusions without any statistical support. For an empirical paper with practical recommendations, that's a load-bearing gap.

Third, the post-hoc selection of sentence-aware chunking for all embedding, reranker, and generator experiments means those results are conditional on that choice. It's not circular, but it's a selection bias, and the paper doesn't discuss how robust the component rankings are to that choice.

Minor points: the related work citations are a bit light and some references look misformatted, but that doesn't affect the results.

Bottom line: this paper deserves peer review. It's a genuinely new empirical dataset for a low-resource language area, and practitioners will get value from it. But the revision needs to name the RAGAS judge, add a human evaluation or at least report results with multiple judges, and either compute variances or soften the causal language. Without that, the headline claims remain unverified.

Referee Report

5 major / 5 minor

Summary. The manuscript reports an empirical comparison of RAG pipeline components for Arabic: four chunking strategies, six embedding models, one reranker, and two generation LLMs, evaluated with the RAGAS metric suite on six Arabic datasets (ARCD, ArSQUAD, SaudiWiki, QA4MRE, Quran Tafseer, Hindawi Books). The authors claim that sentence-aware chunking gives the best overall RAG performance; that BGE-M3 and Multilingual-E5-large are the most effective embedding models; that adding the bge-reranker-v2-m3 improves faithfulness and overall scores, especially on complex datasets; and that Aya-8B outperforms StableLM-1.6B for Arabic generation. All component comparisons are made through a single equal-weight average of four RAGAS metrics (context precision, context recall, answer faithfulness, answer relevancy). The paper frames the contribution as a practical default configuration for Arabic RAG systems and a reproducible evaluation framework.

Significance. If the findings are robust, the paper provides a useful empirical baseline for Arabic RAG component selection, a relatively underexplored area. The study covers multiple public Arabic datasets and includes both Arabic-specific and multilingual models, which is valuable for practitioners. The paper's descriptive results—e.g., the reported average scores in Tables 1–4—are directly supported by the tables, and the use of standardized tools (LangChain, RAGAS, FAISS) is a strength. However, the paper's broader language of 'outperforms', 'significantly boosts', and 'consistently outperforms' is not supported by any uncertainty quantification or significance testing. In addition, the RAGAS faithfulness and answer-relevancy metrics depend on an unnamed LLM judge, and Section 4 explicitly states that these metrics were estimated without human judgment, which raises a correctness risk for the central rankings. The paper is a reasonable empirical study in need of stronger statistical grounding and more transparent evaluation methodology before the practical guidelines can be accepted.

major comments (5)

[§3.6 and §4] The RAGAS judge LLM is never identified. Faithfulness and answer relevancy are LLM-based metrics, and Section 4 concedes that they 'are estimated without human judgment'. If the judge model has weak Arabic ability, or if the judge is the same model used as the generator (Aya-8B in the RQ4 comparison), the reported rankings—including the central claim that Aya-8B surpasses StableLM—could be artifacts of self-preference or judge bias. The statement in Section 3.6 about 'manually curated ground-truth answers' does not resolve this, because Section 4 restricts that claim to the retrieval metrics. Please name the judge model, report its configuration, and provide at least a small human-evaluation sample (or a judge-robustness analysis) for the generation-quality metrics.
[§3.8, Table 1] The claim that 'sentence-aware chunking outperforms all other segmentation methods' is not supported by the reported numbers. Table 1 shows that sentence-aware is not the best on SaudiWiki (fixed-size scores 86.69 vs. 82.76) and is essentially tied with recursive chunking on ArSQUAD (52.21 vs. 52.35) and with semantic chunking on Quran Tafseer (81.63 vs. 82.45). The paper reports only single average values with no error bars, confidence intervals, or significance tests, so the differences that drive the headline conclusion may be within sampling noise. Please provide per-sample variance or bootstrap confidence intervals, or temper the 'outperforms all' language to reflect the actual margin.
[§3.3] All subsequent component experiments (embedding models, reranker, LLMs) use sentence-aware chunking because it 'demonstrated superior performance in our earlier experiments'. This post-hoc selection biases the later comparisons: the reported ranking of embedding models, for example, is conditional on one chunking strategy, and there is no sensitivity analysis showing whether the embedding ranking holds under other chunking methods. Because the paper's contribution is a set of 'optimal components', the lack of interaction analysis is a load-bearing limitation. Please either test at least one alternative chunking strategy for the embedding comparison or explicitly restrict the conclusions to the sentence-aware configuration.
[§3.9, Table 2] The claim that BGE-M3 and Multilingual-E5-large 'are the most effective embedding models' rests on an average difference of 0.68 points (70.99 vs. 70.31). Table 2 shows that Multilingual-E5-large is not the top model on four of six datasets (ARCD, QA4MRE, Quran Tafseer, Hindawi), and BGE-M3 is not top on ArSQUAD or Hindawi. Without statistical testing or a clear statement of the tie-breaking rule, this ranking is fragile. Please provide a more careful analysis of the per-dataset variability and justify why a single overall average is the appropriate decision criterion.
[Abstract and §3.10] The abstract says the reranker 'significantly boosts faithfulness in complex datasets', but the paper reports only overall mean scores in Table 3, not the per-metric faithfulness scores for each dataset, and performs no significance test. The narrative in Section 3.10 mentions a '15-point' faithfulness gain on ARCD, but the corresponding table does not report faithfulness separately, so the reader cannot verify this claim. Please include the per-metric results (at least for faithfulness and answer relevancy) and a significance or effect-size measure for the reranking comparison.

minor comments (5)

[§3.7] Section 3.7 is titled 'Results' but is empty; the actual results begin in Section 3.8. Either remove the empty heading or move the results under a numbered 'Results' section.
[§3.1, Ref. [16]] The ArSQUAD dataset is described as 'an Arabic adaptation of the popular English SQuAD dataset', but reference [16] is the original SQuAD paper, not an Arabic adaptation. Please cite the correct source for ArSQUAD or clarify the relationship to reference [16].
[Declarations] The Declarations section contains a typo: 'Acknoweldment' should be 'Acknowledgement'. Additionally, the acknowledgment statement duplicates the conflict-of-interest statement; please separate them or remove the duplicate.
[Tables 1–4] The table captions contain formatting artifacts ('T able', 'F aithfulness', 'Re-rank') that should be corrected. The bold highlighting in Table 2 does not appear consistent with the text description (e.g., ArSQUAD's best score is 56.55 for E5-Large, which is not bolded).
[§2.4] The related work section describes NoMIRACL and Futurepedia but does not draw explicit connections to the methodological choices made in this paper (e.g., why these benchmarks were not used for evaluation). Adding one sentence connecting these benchmarks to the experimental design would improve the framing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the component rankings are direct RAGAS measurements on public datasets, with no fitted-parameter prediction loop or load-bearing self-citation.

full rationale

The paper's central claims are empirical comparisons of RAG components (chunking, embedding, reranking, generation) measured with the RAGAS framework over six public Arabic datasets. There is no fitted parameter that is later renamed as a prediction, and no derived quantity that is equivalent to an input by construction. The only structural choice that could raise a concern is the post-hoc adoption of sentence-aware chunking for all embedding, reranking, and generation experiments, as stated in Section 3.3: 'For consistency, all embeddings were generated using sentence-aware chunking, as this strategy demonstrated superior performance in our earlier experiments.' This is an adaptive experimental design choice, disclosed in the paper, and it does not make the reported component scores circular: each component is still independently measured under the stated configuration, and the chunking comparison itself does not depend on the later experiments. The reliance on RAGAS automated metrics without an identified LLM judge, acknowledged in Section 4 ('faithfulness and answer relevancy are estimated without human judgment'), is a validity or reproducibility limitation, not a circular derivation. There is also no load-bearing self-citation: the RAGAS framework is cited as an external tool, and the paper's own contributions are presented as new experiments. Therefore the derivation chain is self-contained and no circular step is present.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims rest on measurement choices that are not fully reported, especially the RAGAS judge model and the exact retrieval configuration. No new entities are introduced.

free parameters (4)

Chunk size and overlap
Not reported in the paper; the performance of all four chunking strategies directly depends on these values, and their absence prevents replication.
Top-k retrieved chunks
Not reported; this value controls how many contexts are passed to the reranker and generator, affecting context precision and recall.
RAGAS judge LLM
The LLM used by RAGAS to compute faithfulness and answer relevancy is unspecified; this choice can strongly affect the reported scores, especially for Arabic.
Equal-weight overall score
The 'overall score' appears to be an unweighted average of four RAGAS metrics; this aggregation choice affects which component is declared best.

assumptions (3)

domain assumption RAGAS automated metrics provide valid measures of Arabic RAG quality.
Invoked throughout Section 3.6; no human validation is performed, and the paper acknowledges this in the threats to validity section.
domain assumption The six selected datasets are representative of Arabic document types and tasks.
Stated in Section 3.1 as providing 'a robust foundation'; no sampling details or coverage analysis are given.
standard math Cosine similarity in FAISS is an appropriate retrieval mechanism for the tested embeddings.
Used in Section 3.3; this is standard practice and not directly tested in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components." pith.science (2026). https://pith.science/paper/X325OLDC

@misc{pith2026250606339,
  author       = {Pith},
  title        = {Pith review of: Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X325OLDC}},
  note         = {Machine review of arXiv:2506.06339}
}

read the original abstract

Retrieval-Augmented Generation (RAG) has emerged as a powerful architecture for combining the precision of retrieval systems with the fluency of large language models. While several studies have investigated RAG pipelines for high-resource languages, the optimization of RAG components for Arabic remains underexplored. This study presents a comprehensive empirical evaluation of state-of-the-art RAG components-including chunking strategies, embedding models, rerankers, and language models-across a diverse set of Arabic datasets. Using the RAGAS framework, we systematically compare performance across four core metrics: context precision, context recall, answer faithfulness, and answer relevancy. Our experiments demonstrate that sentence-aware chunking outperforms all other segmentation methods, while BGE-M3 and Multilingual-E5-large emerge as the most effective embedding models. The inclusion of a reranker (bge-reranker-v2-m3) significantly boosts faithfulness in complex datasets, and Aya-8B surpasses StableLM in generation quality. These findings provide critical insights for building high-quality Arabic RAG pipelines and offer practical guidelines for selecting optimal components across different document types.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 11 canonical work pages

[1]

arXiv preprint arXiv:2312.109972, 1 (2023)

Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, H., Wang, H.: Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.109972, 1 (2023)

arXiv 2023
[2]

Advances in neural information processing systems33, 9459–9474 (2020)

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.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33, 9459–9474 (2020)

work page 2020
[3]

Procedia Computer Science244, 296–307 (2024)

El-Beltagy, S.R., Abdallah, M.A.: Explor- ing retrieval augmented generation in arabic. Procedia Computer Science244, 296–307 (2024)

work page 2024
[4]

arXiv preprint arXiv:2410.21970 (2024)

Wu, S., Tang, J., Yang, B., Wang, A., Jia, K., Yu, J., Yao, J., Su, J.: Not all languages are equal: Insights into multilin- gual retrieval-augmented generation. arXiv preprint arXiv:2410.21970 (2024)

arXiv 2024
[5]

Communications of the ACM64(4), 72–81 (2021)

Darwish, K., Habash, N., Abbas, M., Al- Khalifa, H., Al-Natsheh, H.T., Bouamor, H., Bouzoubaa, K., Cavalli-Sforza, V., El- Beltagy, S.R., El-Hajj, W.,et al.: A panoramic survey of natural language pro- cessing in the arab world. Communications of the ACM64(4), 72–81 (2021)

work page 2021
[6]

Morgan & Claypool Publishers, ??? (2010)

Habash, N.Y.: Introduction to Arabic Natu- ral Language Processing. Morgan & Claypool Publishers, ??? (2010)

work page 2010
[7]

IEEE Transactions on Learning Technologies16(1), 112–125 (2023)

Chen, J., Wang, Z., Zhao, T., Yin, H.: Retrieval-based question answering for e- learning. IEEE Transactions on Learning Technologies16(1), 112–125 (2023)

work page 2023
[8]

arXiv preprint arXiv:2112.07899 (2022)

Ni, J., Bian, Y., Qu, W., Cheng, H.: Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899 (2022)

arXiv 2022

Show all 21 references

[9]

arXiv preprint arXiv:2402.03216 (2024)

Chen, X., Zheng, S., Jiang, S., Huang, P., Chen, J., Zhang, Z., Gong, Y., Liu, Z., Sun, M.: Bge m3-embedding: Multi-lingual, multi- functionality, multi-granularity text embed- dings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 (2024)

2024 arXiv
[10]

arXiv preprint arXiv:2102.10684 (2021)

Abdelali, A., Hassan, S., Mubarak, H., Dar- wish, K., Samih, Y.: Pre-training bert on arabic tweets: Practical considerations. arXiv preprint arXiv:2102.10684 (2021)

2021 arXiv
[11]

arXiv preprint arXiv:2310.11511 (2023)

Ma, Z., Ren, X., Zhao, X., Zhao, T., Liu, Z., Sun, M.: Fine-grained reranking for retrieval-augmented generation. arXiv preprint arXiv:2310.11511 (2023)

2023 arXiv
[12]

arXiv preprint arXiv:2003.00104 (2020)

Antoun, W., Baly, F., Hajj, H.: Arabert: Transformer-based model for arabic lan- guage understanding. arXiv preprint arXiv:2003.00104 (2020)

2020 arXiv
[13]

arXiv preprint arXiv:2402.07827 (2024)

Ousidhoum, N., Abdulhai, H., Abdulhai, O., Aguila, A., Aliyu, I., Alkhamissi, B., Altakrori, M., Bari, M.S., Basyal, D., Chakraborty, B., et al.: Aya model: An open science approach to align language models with human values across cultures. arXiv preprint arXiv:2402.07827 (2024)

2024 arXiv
[14]

Transactions of the Association for Compu- tational Linguistics (2024)

Thakur, N., Reimers, N., R¨ uckl´ e, A., Srivas- tava, A., Gurevych, I.: Nomiracl: Multilingual information retrieval across 18 languages. Transactions of the Association for Compu- tational Linguistics (2024)

2024
[15]

In: Proceedings of the Fourth Arabic Natural Language Processing Workshop, pp

Mozannar, H., Maamary, E., El Hajal, K., Hajj, H.: Neural arabic question answering. In: Proceedings of the Fourth Arabic Natural Language Processing Workshop, pp. 108–118 (2019)

2019
[16]

arXiv preprint arXiv:1606.05250 (2019)

Arkady, A., Bernal, B., Rajpurkar, P., Zhang, J., Lopyrev, K., Liang, P.: Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250 (2019)

2019 arXiv
[17]

arXiv preprint arXiv:2211.15147 (2022)

Alqahtani, S., Alotaibi, F., Abdelali, Springer Nature 2021 LATEX template 12Article Title A., Sajjad, H.: Saudiwiki: A large-scale arabic wikipedia dataset for knowledge extraction and retrieval. arXiv preprint arXiv:2211.15147 (2022)

2022 arXiv
[18]

In: International Conference of the Cross-Language Evaluation Forum for European Languages, pp

Pe˜ nas, A., Hovy, E., Forner, P., Rodrigo, ´A., Sutcliffe, R., Sporleder, C., Forascu, C., Benajiba, Y., Osenova, P.: Qa4mre 2011- 2013: Overview of question answering for machine reading evaluation. In: International Conference of the Cross-Language Evaluation Forum for Euro...

2013
[19]

In: Proceedings of the 12th Language Resources and Evaluation Conference, pp

Altammami, S., Atwell, E., Alsalka, A.: Auto- matic extraction of quranic knowledge. In: Proceedings of the 12th Language Resources and Evaluation Conference, pp. 4318–4324 (2020)

2020
[20]

arXiv preprint arXiv:1204.4327 (2012)

Zerrouki, T., Balla, A.: Hindawi: An open source corpus framework for arabic. arXiv preprint arXiv:1204.4327 (2012)

2012 arXiv
[21]

arXiv preprint arXiv:2309.15217 (2023)

Es, S., Subramanian, N., Konda, S., Agar- wal, S., Ganesan, V., Gupta, I., Gupta, A., Narayanan, V., Khapra, M.M.: Ragas: Automated evaluation of retrieval augmented generation. arXiv preprint arXiv:2309.15217 (2023)

2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

arXiv preprint arXiv:2312.109972, 1 (2023)

Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, H., Wang, H.: Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.109972, 1 (2023)

arXiv 2023

[2] [2]

Advances in neural information processing systems33, 9459–9474 (2020)

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.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33, 9459–9474 (2020)

work page 2020

[3] [3]

Procedia Computer Science244, 296–307 (2024)

El-Beltagy, S.R., Abdallah, M.A.: Explor- ing retrieval augmented generation in arabic. Procedia Computer Science244, 296–307 (2024)

work page 2024

[4] [4]

arXiv preprint arXiv:2410.21970 (2024)

Wu, S., Tang, J., Yang, B., Wang, A., Jia, K., Yu, J., Yao, J., Su, J.: Not all languages are equal: Insights into multilin- gual retrieval-augmented generation. arXiv preprint arXiv:2410.21970 (2024)

arXiv 2024

[5] [5]

Communications of the ACM64(4), 72–81 (2021)

Darwish, K., Habash, N., Abbas, M., Al- Khalifa, H., Al-Natsheh, H.T., Bouamor, H., Bouzoubaa, K., Cavalli-Sforza, V., El- Beltagy, S.R., El-Hajj, W.,et al.: A panoramic survey of natural language pro- cessing in the arab world. Communications of the ACM64(4), 72–81 (2021)

work page 2021

[6] [6]

Morgan & Claypool Publishers, ??? (2010)

Habash, N.Y.: Introduction to Arabic Natu- ral Language Processing. Morgan & Claypool Publishers, ??? (2010)

work page 2010

[7] [7]

IEEE Transactions on Learning Technologies16(1), 112–125 (2023)

Chen, J., Wang, Z., Zhao, T., Yin, H.: Retrieval-based question answering for e- learning. IEEE Transactions on Learning Technologies16(1), 112–125 (2023)

work page 2023

[8] [8]

arXiv preprint arXiv:2112.07899 (2022)

Ni, J., Bian, Y., Qu, W., Cheng, H.: Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899 (2022)

arXiv 2022