---
title: "Optimizing RAG Pipelines for Arabic — Alsubhi 2025"
description: "The paper addresses a gap in the literature: retrieval-augmented generation (RAG) has been tuned mostly for English and other high-resource languages…"
authors: ["Jumana Alsubhi", "Mohammad D. Alahmadi", "Ahmed Alhusayni", "Ibrahim Aldailami", "Israa Hamdine", "Ahmad Shabana", "Yazeed Iskandar", "Suhayb Khayyat"]
year: 2025
venue: "arXiv:2506.06339"
source: "https://arxiv.org/abs/2506.06339"
doi: "10.48550/arxiv.2506.06339"
arxiv: "2506.06339"
licence: "CC0 1.0 (arXiv)"
licence_url: "http://creativecommons.org/publicdomain/zero/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/alsubhi-2025-optimizing-rag-pipelines-arabic-systematic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/alsubhi-2025-optimizing-rag-pipelines-arabic-systematic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/) / Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components

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

**Authors:** Jumana Alsubhi, Mohammad D. Alahmadi, Ahmed Alhusayni, Ibrahim Aldailami, Israa Hamdine, Ahmad Shabana, Yazeed Iskandar, Suhayb Khayyat · **Year:** 2025 · **Venue:** arXiv:2506.06339 · **Licence:** [CC0 1.0 (arXiv)](http://creativecommons.org/publicdomain/zero/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2506.06339) · [PDF](https://arxiv.org/pdf/2506.06339) · [DOI 10.48550/arxiv.2506.06339](https://doi.org/10.48550/arxiv.2506.06339) · [arXiv 2506.06339](https://arxiv.org/abs/2506.06339) · [OpenAlex](https://openalex.org/W4417097942)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14) · 2 in our library, of the 140884 papers whose bibliographies we read

## Summary
The paper addresses a gap in the literature: retrieval-augmented generation (RAG) has been tuned mostly for English and other high-resource languages, and there is little guidance on which components work best for Arabic. The authors built a modular Arabic RAG pipeline using LangChain, FAISS and the RAGAS evaluation framework. They changed one component at a time and measured the effect across six public Arabic corpora: ARCD, ArSQUAD, SaudiWiki, QA4MRE, Quran Tafseer and Hindawi Books. The corpora range from encyclopedic facts to religious commentary and literary prose. The components tested were four chunking methods, six embedding models, one reranker and two generators. Each configuration was scored on four RAGAS metrics, two for retrieval and two for generation. Sentence-aware chunking gave the most consistent gains. BGE-M3 and Multilingual-E5-large were the strongest embedders. Adding bge-reranker-v2-m3 mainly improved faithfulness on harder data, and Aya-8B produced better answers than StableLM-1.6B. These results give practitioners an evidence-based starting configuration for Arabic RAG, and the study design itself can be reused for other languages.

## Key points
- The study covers four chunking strategies (fixed-size, recursive, semantic, sentence-aware), six embedding models, the bge-reranker-v2-m3 reranker, and two generators (Aya-8B, StableLM-1.6B). All were scored with the RAGAS metrics context precision, context recall, faithfulness and answer relevancy.
- In the embedding comparison, BGE-M3 averaged 70.99 and E5-Large averaged 70.31, ahead of GTE-Base (68.48) and a Matryoshka model (66.46).
- Adding the reranker on top of the BGE-M3 setup raised the reported average to 74.15, compared with 70.99 without it. The benefit was largest on complex or ambiguous retrieval.
- With chunking and embedding held fixed (sentence-aware chunking, bge-m3), Aya-8B averaged 72.06 against 67.08 for StableLM-1.6B. The gaps were widest on QA4MRE (61.8 vs 55.07) and Quran Tafseer (about 78.4 vs 71.24). Aya-8B's best single result was 88.81 on SaudiWiki.
- The authors acknowledge that all scores come from automated, LLM-based RAGAS judging with no human check, and that Arabic dialects are not covered.

## Interesting ideas
- Each component is varied in isolation while every other stage stays fixed, so a difference in score can be credited to that one component instead of an interaction between several. — For Mova (learn): Swapping one module at a time, such as the parser, the role labeller or the coreference step, against a frozen rest of the pipeline would show which stage actually drives end-to-end answer accuracy.
- Retrieval quality (whether the right context was fetched) is scored separately from generation quality (whether the answer stays true to that context), which lets a failure be traced to the stage that caused it. — For Mova (answer): The readers could be graded both on whether they selected the correct sentences or world facts and on whether the final answer follows from them, which separates lookup errors from reasoning errors.
- Splitting text at sentence boundaries, rather than by character count, keeps meaningful units intact and gave the most reliable gains. This suggests that respecting linguistic structure pays off even in pipelines that otherwise ignore it. — For Mova (tokens): This is a reminder that sentence segmentation quality directly constrains everything downstream, so the en splitter deserves its own evaluation on narrative and dialogue-heavy text.
- A cheap first-pass retriever is followed by a stronger second-pass rescorer of the top candidates, and this paid off most on the hardest inputs. — For Mova (tree): The fast perceptron parser could propose several candidate trees and gparse could rerank them, with the expensive pass applied only to sentences where the first parser is uncertain.
- The test corpora were deliberately chosen to differ in genre and structure, which showed that the best component choice can shift with text type. — For Mova (learn): Reporting evaluation results separately for each genre (fables, textbook passages, word problems) would reveal genre-specific weaknesses that a single pooled score hides.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{alsubhi2025optimizing,
  title = {{Optimizing RAG Pipelines for Arabic: A Systematic Analysis of Core Components}},
  author = {Alsubhi, Jumana and Alahmadi, Mohammad D. and Alhusayni, Ahmed and Aldailami, Ibrahim and Hamdine, Israa and Shabana, Ahmad and Iskandar, Yazeed and Khayyat, Suhayb},
  year = {2025},
  doi = {10.48550/arxiv.2506.06339},
  eprint = {2506.06339},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2506.06339}
}
```

- [← previous OpenThoughts: Data Recipes for Reasoning Models](https://md.research.elamur.ai/free/en/papers/guha-2025-openthoughts-data-recipes-reasoning-models/)
- [next → Overcoming Data Scarcity in Multi-Dialectal Arabic ASR via Whisper Fine-Tuning](https://md.research.elamur.ai/free/en/papers/ozyilmaz-2025-overcoming-data-scarcity-multi-dialectal/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/)

## References

21 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. 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:2312.10997 2, 1 (2023) [↗](https://arxiv.org/abs/2312.10997) [our card](https://md.research.elamur.ai/free/en/papers/gao-2023-retrieval-augmented-generation-large-language/)
2. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020) [find](https://scholar.google.com/scholar?q=Lewis%2C+P%2E%2C+Perez%2C+E%2E%2C+Piktus%2C+A%2E%2C+Petroni%2C+F%2E%2C+Karpukhin%2C+V%2E%2C+Goyal%2C+N%2E%2C+Ku%CC%88ttler%2C+H%2E%2C+Lewis%2C+M%2E%2C+Yih%2C+W%2E%2Dt%2E%2C+Rockta%CC%88schel%2C+T%2E%2C+et+al%2E%3A+Retrieval%2Daugmented+generation+for+knowledge%2Dintensive+nlp+tasks)
3. El-Beltagy, S.R., Abdallah, M.A.: Exploring retrieval augmented generation in arabic. Procedia Computer Science 244, 296–307 (2024) [find](https://scholar.google.com/scholar?q=El%2DBeltagy%2C+S%2ER%2E%2C+Abdallah%2C+M%2EA%2E%3A+Exploring+retrieval+augmented+generation+in+arabic%2E+Procedia+Computer+Science+244%2C+296%E2%80%93307+%282024%29) [our card](https://md.research.elamur.ai/free/en/papers/el-2024-beltagy-exploring-retrieval-augmented-generation-arabic/)
4. Wu, S., Tang, J., Yang, B., Wang, A., Jia, K., Yu, J., Yao, J., Su, J.: Not all languages are equal: Insights into multilingual retrieval-augmented generation. arXiv preprint arXiv:2410.21970 (2024) [↗](https://arxiv.org/abs/2410.21970)
5. 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 processing in the arab world. Communications of the ACM 64(4), 72–81 (2021) [find](https://scholar.google.com/scholar?q=Darwish%2C+K%2E%2C+Habash%2C+N%2E%2C+Abbas%2C+M%2E%2C+Al%2DKhalifa%2C+H%2E%2C+Al%2DNatsheh%2C+H%2ET%2E%2C+Bouamor%2C+H%2E%2C+Bouzoubaa%2C+K%2E%2C+Cavalli%2DSforza%2C+V%2E%2C+El%2DBeltagy%2C+S%2ER%2E%2C+El%2DHajj%2C+W%2E%2C+et+al%2E%3A+A+panoramic+survey+of+natural+language+proc)
6. Habash, N.Y.: Introduction to Arabic Natural Language Processing. Morgan & Claypool Publishers, ??? (2010) [find](https://scholar.google.com/scholar?q=Habash%2C+N%2EY%2E%3A+Introduction+to+Arabic+Natural+Language+Processing%2E+Morgan+%26+Claypool+Publishers%2C+%3F%3F%3F+%282010%29)
7. Chen, J., Wang, Z., Zhao, T., Yin, H.: Retrieval-based question answering for elearning. IEEE Transactions on Learning Technologies 16(1), 112–125 (2023) [find](https://scholar.google.com/scholar?q=Chen%2C+J%2E%2C+Wang%2C+Z%2E%2C+Zhao%2C+T%2E%2C+Yin%2C+H%2E%3A+Retrieval%2Dbased+question+answering+for+elearning%2E+IEEE+Transactions+on+Learning+Technologies+16%281%29%2C+112%E2%80%93125+%282023%29)
8. Ni, J., Bian, Y., Qu, W., Cheng, H.: Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899 (2022) [↗](https://arxiv.org/abs/2112.07899) [our card](https://md.research.elamur.ai/free/en/papers/ni-2021-large-dual-encoders-generalizable-retrievers/)
9. 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 embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 (2024) [↗](https://arxiv.org/abs/2402.03216) [our card](https://md.research.elamur.ai/free/en/papers/chen-2024-m3-embedding-multi-linguality/)
10. Abdelali, A., Hassan, S., Mubarak, H., Darwish, K., Samih, Y.: Pre-training bert on arabic tweets: Practical considerations. arXiv preprint arXiv:2102.10684 (2021) [↗](https://arxiv.org/abs/2102.10684) [our card](https://md.research.elamur.ai/free/en/papers/abdelali-2021-pre-training-bert-arabic-tweets/)
11. 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) [↗](https://arxiv.org/abs/2310.11511) [our card](https://md.research.elamur.ai/free/en/papers/asai-2023-self-rag-learning-retrieve-generate/)
12. Antoun, W., Baly, F., Hajj, H.: Arabert: Transformer-based model for arabic language understanding. arXiv preprint arXiv:2003.00104 (2020) [↗](https://arxiv.org/abs/2003.00104) [our card](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/)
13. 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) [↗](https://arxiv.org/abs/2402.07827) [our card](https://md.research.elamur.ai/free/en/papers/ustun-2024-aya-model-instruction-finetuned-open/)
14. Thakur, N., Reimers, N., Rücklé, A., Srivastava, A., Gurevych, I.: Nomiracl: Multilingual information retrieval across 18 languages. Transactions of the Association for Computational Linguistics (2024) [find](https://scholar.google.com/scholar?q=Thakur%2C+N%2E%2C+Reimers%2C+N%2E%2C+Ru%CC%88ckle%CC%81%2C+A%2E%2C+Srivastava%2C+A%2E%2C+Gurevych%2C+I%2E%3A+Nomiracl%3A+Multilingual+information+retrieval+across+18+languages%2E+Transactions+of+the+Association+for+Computational+Linguistics+%2820)
15. 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) [find](https://scholar.google.com/scholar?q=Mozannar%2C+H%2E%2C+Maamary%2C+E%2E%2C+El+Hajal%2C+K%2E%2C+Hajj%2C+H%2E%3A+Neural+arabic+question+answering%2E+In%3A+Proceedings+of+the+Fourth+Arabic+Natural+Language+Processing+Workshop%2C+pp%2E+108%E2%80%93118+%282019%29)
16. 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) [↗](https://arxiv.org/abs/1606.05250) [our card](https://md.research.elamur.ai/free/en/papers/rajpurkar-2016-squad-100-000-questions-machine/)
17. Alqahtani, S., Alotaibi, F., Abdelali, A., Sajjad, H.: Saudiwiki: A large-scale arabic wikipedia dataset for knowledge extraction and retrieval. arXiv preprint arXiv:2211.15147 (2022) [↗](https://arxiv.org/abs/2211.15147)
18. Peñas, A., Hovy, E., Forner, P., Rodrigo, Á., 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 European Languages, pp. 303–320 (2013). Springer [find](https://scholar.google.com/scholar?q=Pen%CC%83as%2C+A%2E%2C+Hovy%2C+E%2E%2C+Forner%2C+P%2E%2C+Rodrigo%2C+A%CC%81%2E%2C+Sutcliffe%2C+R%2E%2C+Sporleder%2C+C%2E%2C+Forascu%2C+C%2E%2C+Benajiba%2C+Y%2E%2C+Osenova%2C+P%2E%3A+Qa4mre+2011%2D2013%3A+Overview+of+question+answering+for+machine+reading+evaluation%2E+I)
19. Altammami, S., Atwell, E., Alsalka, A.: Automatic extraction of quranic knowledge. In: Proceedings of the 12th Language Resources and Evaluation Conference, pp. 4318–4324 (2020) [find](https://scholar.google.com/scholar?q=Altammami%2C+S%2E%2C+Atwell%2C+E%2E%2C+Alsalka%2C+A%2E%3A+Automatic+extraction+of+quranic+knowledge%2E+In%3A+Proceedings+of+the+12th+Language+Resources+and+Evaluation+Conference%2C+pp%2E+4318%E2%80%934324+%282020%29)
20. Zerrouki, T., Balla, A.: Hindawi: An open source corpus framework for arabic. arXiv preprint arXiv:1204.4327 (2012) [↗](https://arxiv.org/abs/1204.4327)
21. Es, S., Subramanian, N., Konda, S., Agarwal, 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) [↗](https://arxiv.org/abs/2309.15217) [our card](https://md.research.elamur.ai/free/en/papers/es-2023-ragas-automated-evaluation-retrieval-augmented/)