INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 193
A Scalable Retrieval-Augmented Generation Pipeline for Domain-
Specific Knowledge Applications
Emmanuel A. Olanrewaju
DOI: https://doi.org/10.51584/IJRIAS.2025.1010000014
Received: 26 Sep 2025; Accepted: 02 Oct 2025; Published: 28 October 2025
ABSTRACT
This work presents a Retrieval-Augmented Generation (RAG) pipeline that integrates document preprocessing,
embedding-based retrieval, and large language model (LLM) generation into a unified framework. The pipeline
begins with the ingestion of PDF documents, followed by text cleaning, sentence segmentation, and chunking
to ensure compatibility with embedding model constraints. High-dimensional vector representations are
generated using transformer-based embedding models and stored for downstream use. Semantic similarity
search, implemented via dot product and cosine similarity, enables efficient retrieval of contextually relevant
text. For scalability, the framework is designed to accommodate vector indexing methods such as Faiss. On
the generation side, locally hosted LLM (Gemma-7B) is employed with optional quantization for reduced
resource consump- tion. Retrieved context is integrated with user queries to enhance the accuracy and
relevance of generated responses. This pipeline demonstrates a practical approach for building domain-
specific, retrieval-augmented applications that balance efficiency, scalability, and adaptability to local com-
pute environments.
INTRODUCTION
Digital transformation signifies the integration of dig- ital technology across various facets of a business, re-
shaping both its operations and the delivery of value to customers [1]. At the forefront of this transforma-
tion are Large Language Models (LLMs), advanced ma- chine learning models trained extensively on textual
data to understand and generate human-like text [1]. LLMs, such as the Generative Pre-training Transformer
(GPT) series [2, 3] and others, have demonstrated remarkable capabilities in natural language processing
(NLP) tasks [4]. Despite these advances, LLMs often struggle with domain-specific queries, generating
inaccurate or irrele- vant information—commonly referred to as “hallucina- tions”—especially when data is
sparse [5]. Consequently, applying LLMs in practical scenarios is challenging, as their outputs cannot always
be relied upon for accuracy.
Pre-trained models are capable of learning substantial amounts of in-depth knowledge directly from data [6],
functioning as parameterized implicit knowledge bases without requiring external memory [7, 8]. While this
capability is promising, such models have limitations: they cannot easily expand or revise their knowledge, can-
not transparently explain their predictions, and remain prone to hallucinations [9]. Hybrid approaches that com-
bine parametric memory with non-parametric, retrieval- based memory [10–12] address some of these limitations
by allowing knowledge to be updated dynamically and providing transparency in the retrieved information.
Retrieval-augmented generation (RAG) has emerged as a powerful approach in NLP, effectively combining
the strengths of retrieval and generative models [13]. RAG has been applied to reduce hallucinations, provide
knowledge grounding, and enable personalization [14– 17]. Evaluating RAG systems is crucial for ensuring
the effectiveness of integrating retrieval-based methods with generative models [18, 19]. Traditionally, such
evalua- tion relies on end-to-end assessment, comparing gener- ated outputs with one or more ground truth
references [20]. While essential, this approach presents limitations, particularly for assessing the performance of
the retrieval component in RAG systems.
In this work, we present a Retrieval-Augmented Gen- eration (RAG) pipeline that integrates document pre-
processing, embedding-based retrieval, and LLM gener- ation into a cohesive framework. The pipeline begins
with the ingestion of PDF documents, followed by text cleaning, sentence segmentation, and chunking to ensure
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 194
compatibility with embedding model constraints. High- dimensional vector representations are generated using
transformer-based embedding models and stored for ef- ficient downstream retrieval. Semantic similarity search,
implemented via dot product and cosine similarity, en- ables rapid identification of contextually relevant text,
while vector indexing methods such as Faiss provide scal- ability for large document collections.
On the generation side, locally hosted LLM (Gemma- 7B) is employed, with optional quantization to reduce
computational resource requirements. Retrieved con- text is combined with user queries to enhance the rele-
vance and accuracy of generated responses. This pipeline provides a practical approach for constructing domain-
specific, retrieval-augmented applications that are effi- cient, scalable, and adaptable to local compute environ-
ments, establishing a foundation for robust knowledge- driven systems.
METHODOLOGY
This study implements a RAG pipeline by following a structured computational workflow developed in Python.
The methodology consists of several sequential stages: data acquisition, preprocessing, embedding generation,
retrieval system design, large language model integration, and evaluation using specialized metrics. Each stage
is implemented as described below.
Data Acquisition and Parsing
The PDF for this work was downloaded from Press- books OER Hawaii Human Nutrition. Raw data was
ingested from PDF documents using the fitz library (PyMuPDF) [21]. This parser was selected due to its ro-
bustness in extracting structured and unstructured text compared to alternatives such as PyPDF [22]. The ex-
tracted text was organized into structured tabular form using pandas [23] and datasets [24] to facilitate fur- ther
downstream processing. Metadata such as docu- ment name, section identifiers, and chunk indices were
preserved.
Text Preprocessing
To prepare the text for embedding and retrieval, sev- eral preprocessing steps were executed:
Tokenization and sentence boundary detection were performed using the spacy.lang.en tokenizer [25].
Text was segmented into coherent chunks to bal- ance retrieval efficiency and context length require- ments.
Cleaning procedures included lowercasing, removal of extraneous whitespace, and preservation of domain-
specific entities.
This ensured consistency across documents while retain- ing semantic information critical for embedding.
Embedding Generation
Semantic embeddings were generated using the all-mpnet-base-v2 model implemented in the
SentenceTransformer library [26]. A pre-trained transformer-based embedding model was employed to
convert text chunks into dense vector representations. The sentence_transformers.util module was used for
similarity computations, enabling effective nearest- neighbor retrieval. Embeddings were stored for rapid
querying during RAG execution.
The choice of all-mpnet-base-v2 was mo- tivated by its balance between semantic rich- ness and
computational efficiency.
Compared to alternative models such as MiniLM-L12-v2 or distilbert-base-nli-stsb-mean-tokens, it provides
superior performance on sentence-level semantic similar- ity tasks. Despite challenges posed by domain-specific
medical terminology, this model demonstrated robust generalization and strong clustering behavior, making it a
suitable baseline for our pipeline.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 195
Retrieval System Design
A vector-based retrieval system was constructed over the generated embeddings. Document similarity was
computed using cosine similarity. This enabled the re- trieval of top-k semantically relevant chunks given a user
query. The retrieval process was optimized with progress tracking using tqdm to monitor performance across
large datasets.
Large Language Model Integration
A causal language model was integrated using the
transformers library [27]. Specifically:
Tokenization was handled by Auto Tokenizer.
Model loading was performed with
Auto Model For Causal LM.
To improve memory efficiency, quantization was enabled using Bits And Bytes Config, with
conditional acceleration checks via is_flash_attn_2_available.
The retriever passed the top-ranked text chunks to the language model, enabling contextually grounded answer
generation.
Evaluation Metrics
The RAG pipeline was systematically evaluated using the ragas library and its metrics [28]. The following
quantitative metrics were computed:
Context Entity Recall: Measures the ability of retrieved passages to cover key entities relevant to the query.
Noise Robustness: Evaluates pipeline resilience to irrelevant or noisy inputs.
Additional evaluation functions from ragas. evaluate were applied to provide ag- gregate scores.
Timing benchmarks were recorded using time. perf_counter.
Workflow Summary
The entire methodology, from PDF ingestion through model response evaluation, was executed in a Jupyter
Notebook environment. Each step was modularized into reproducible code cells to ensure transparency and repro-
ducibility. The integration of preprocessing, retrieval, and generative modeling established a complete RAG
pipeline suitable for experimentation with real-world text corpora.
RESULTS AND DISCUSSION
Document Processing and Text Chunking
The RAG pipeline successfully processed the Hu- man Nutrition textbook (2020 Edition) containing 1,208
pages. The text extraction and preprocessing pipeline demonstrated robust performance in handling PDF con-
tent with varying structural elements. Statistical anal- ysis of the processed text revealed an average of 10.32
sentences per page with a standard deviation of 6.30, in- dicating consistent content density throughout the doc-
ument.
The chunking strategy employing spaCy’s sentence to- kenizer with a fixed size of 10 sentences per chunk proved
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 196
effective. This approach resulted in 1,843 text chunks with an average token count of 183.61 (approximately
734 characters), well within the embedding model’s 384- token capacity limit. The chunk size distribution
showed:
Minimum chunk size: 12 characters (3 tokens)
Maximum chunk size: 1,831 characters (457 to- kens)
Interquartile range: 315-1,118 characters (78-279 tokens)
Filtering chunks with fewer than 30 tokens eliminated insignificant content such as page headers, footers, and
isolated phrases, improving the quality of embeddings while maintaining contextual coherence.
A. Embedding Generation and Vector Representation
B. The sentence-transformers library with the all-mpnet-base-v2 model generated high-quality 768-
dimensional embeddings for each text chunk. The embedding process successfully captured semantic
relationships between nutritional concepts, as evidenced by the coherent clustering of related topics in the vector
space, as illustrated in Figure 1. The model’s 384-token input capacity was optimally utilized, with most
chunks occupying less than 50% of available capacity, ensuring minimal information loss.
Preliminary qualitative analysis of embedding similar- ities revealed that:
Chunks discussing similar micronutrients (e.g., Vi- tamin A and Vitamin E) showed higher cosine sim- ilarity
Related physiological processes (e.g., digestion and absorption) formed natural clusters in the embed- ding space
Taxonomic relationships (e.g., macronutrients to their subcategories) were preserved in the vector
representations
FIG. 1: Similarity score of each chunk with respect to the query.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 197
Computational Efficiency
The pipeline demonstrated efficient processing capabil- ities:
PDF text extraction: Completed in approximately 2 minutes for 1,208 pages
Sentence tokenization: Processed using spaCy’s op- timized pipeline with minimal computational over- head
Embedding generation: Leveraged GPU accelera- tion when available, with average processing time of 100–
200 chunks per minute
In addition, we compared the performance of the Gemma-7B LLM in both quantized and non-quantized
modes. Results showed that quantization reduced mem- ory usage by approximately 40% and improved
through- put by 1.3×, while maintaining comparable generation quality. Non-quantized inference exhibited
slightly lower latency for short sequences, but quantized execution pro- vided a more favorable trade-off for large-
scale retrieval- augmented generation tasks. This substantiates the in- clusion of optional quantization as a
design choice in the pipeline.
Evaluation of the RAG Pipeline
To assess the performance of the RAG pipeline, we defined a set of evaluation questions along with corre-
sponding ground truth answers. The questions cover fundamental nutritional topics, including infant feeding,
micronutrients, digestion, protein intake, and deficiency symptoms. The ground truth answers were obtained
from authoritative sources to serve as benchmarks for evaluating the generated responses.
For each evaluation question, relevant context was re- trieved from the document corpus using semantic em-
beddings generated with the all-mpnet-base-v2 model implemented in the SentenceTransformer library. Re-
trieved context chunks were formatted into prompts, which were then passed to the locally hosted LLM to
generate answers. The procedure can be summarized as follows:
Retrieve context chunks most relevant to the query using cosine similarity and dot product scores.
Format the query and context chunks into a prompt suitable for the LLM.
Generate the answer using the LLM, ensuring that only the generated response is extracted (prompt text is
removed).
Record both the generated answer and the context chunks used for retrieval.
The evaluation of the RAG pipeline was conducted us- ing several metrics to measure the quality and reliability
of generated answers:
Context Precision: Proportion of retrieved con- text that was relevant to the query.
Context Recall: Proportion of all relevant con- text that was successfully retrieved.
Answer Relevancy: Degree to which the gener- ated answer correctly addressed the question.
Faithfulness: Accuracy of the generated content with respect to the retrieved context.
Context Entity Recall (optional): Recall of named entities present in the retrieved context.
Noise Robustness (optional): Sensitivity of the system to irrelevant context or noisy input.
Each metric was computed for all evaluation questions, and average scores were used to categorize performance
as Excellent (≥ 0.8), Good (0.6–0.79), Fair (0.4–0.59), or
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 198
Poor (<0.4).
The evaluation revealed the following key results:
Best Performing Metric: Faithfulness (1.000, Excellent)
Worst Performing Metric: Context Entity Re- call (0.280, Poor)
Overall Performance: Three metrics were Ex- cellent, one Good, zero Fair, and one Poor.
To address the inadequate Context Entity Recall, we plan to explore both reduced fixed chunk sizes (e.g., 5
sentences per chunk) and dynamic chunking strategies based on semantic boundaries. Preliminary experiments
with smaller chunk sizes already indicate an improvement in entity coverage, suggesting that finer granularity can
better capture domain-specific terminology. Addition- ally, we will conduct error analysis on evaluation items
where entity recall failed, to identify whether missed enti- ties arise from chunking limitations, embedding sparsity,
or preprocessing artifacts.
Analysis of these results provides actionable insights:
Retrieval was generally effective, as indicated by high context precision and recall scores.
Answer generation was reliable, achieving high rel- evancy and faithfulness.
The system struggled with retrieving all relevant entities, suggesting potential improvements in em- bedding
quality or chunking strategy.
Detailed evaluation results, including per- question answers and contexts, were saved to
rag_evaluation_results.csv for further inspection and reproducibility.
Limitations and Challenges
Several challenges were identified during implementa- tion:
Text Quality Variations: Some pages contained formatting artifacts and special characters that required
additional cleaning. These issues occa- sionally resulted in fragmented or noisy text seg- ments, which may
have reduced embedding fi- delity and contributed to the low Context Entity Recall scores. Improved
preprocessing strategies such as advanced OCR normalization and entity- preserving cleaning will be evaluated
in future iter- ations.
Embedding Model Constraints: The 384-token limit occasionally truncated longer, complex nutri- tional
explanations.
Medical Terminology: Domain-specific terms required careful handling to ensure accurate seman- tic
representation.
These limitations are addressed in the pipeline en- hancements, including experimenting with reduced and
dynamic chunking strategies to improve entity coverage, refining preprocessing to handle formatting artifacts, and
evaluating embedding and LLM configurations (quan- tized vs. non-quantized) to optimize both performance
and fidelity, as discussed in the Conclusion.
CONCLUSION
The implemented RAG pipeline successfully demon- strates an end-to-end framework for processing and em-
bedding nutritional textbook content. Key achievements include:
Robust Text Processing: The pipeline effec- tively handles real-world PDF documents with complex
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 199
formatting and structure
Semantic Preservation: The chunking strategy maintains contextual coherence while optimizing for
embedding model constraints
Scalable Architecture: The modular design sup- ports easy extension to larger document collections and
different domains
Knowledge Accessibility: The vectorized repre- sentation enables efficient semantic search and re- trieval of
nutritional information
The pipeline provides a solid foundation for developing intelligent nutritional assistance systems, with potential
applications in:
Educational tools for nutrition students and pro- fessionals
Clinical decision support systems
Public health information retrieval platforms
Personalized dietary recommendation engines
In response to the evaluation results, particularly the inadequate Context Entity Recall score, we have out-
lined several enhancements. These include experiment- ing with reduced chunk sizes and dynamic chunking ap-
proaches to improve entity coverage, as well as perform- ing detailed error analysis of missed entities. Further-
more, we provided a clear rationale for selecting the all-mpnet-base-v2 embedding model, acknowledging its
strengths and limitations in handling domain-specific medical terminology. To strengthen computational effi-
ciency, a comparison of quantized versus non-quantized Gemma-7B inference was conducted, demonstrating
the practical benefits of quantization for large-scale tasks. Fi- nally, the influence of formatting artifacts on text
quality was examined, highlighting their impact on preprocess- ing and embedding fidelity. Together, these
refinements provide a roadmap for future iterations of the pipeline and enhance the robustness and
interpretability of our methodology.
Future work should focus on enhancing the pipeline with more sophisticated chunking strategies, domain-
specific embedding models, and integration with larger language models for improved question-answering capa-
bilities.
DATA AVAILABILITY
The primary data source for this project is the Human Nutrition: 2020 Edition textbook, available under
Creative Commons Attribution 4.0 International License from:
https://pressbooks.oer.hawaii.edu/ humannutrition2/
Processed Data Components:
Raw PDF document: human-nutrition-text.pdf
Extracted text chunks with metadata: JSON for- mat (1,843 entries)
Generated embeddings: NumPy array format (768- dimensional vectors)
Statistical summaries: CSV format with chunk characteristics
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 200
Data Characteristics
Total pages processed: 1,208
Final text chunks: 1,843
Vocabulary size: Approximately 15,000 unique terms
Domain coverage: Comprehensive nutritional sci- ence topics
All processed data and embeddings are available in the project repository for research and educational purposes.
Code Availability
The complete implementation of the RAG pipeline is available as an open-source project under MIT License.
Repository: https://github.com/ Olanrewajuemmanuelabiodun/RAG_Nutrition/tree/ main
The codebase is designed for easy modification and ex- tension, with comprehensive documentation and example
usage patterns.
1. K. I. Roumeliotis, N. D. Tselikas, and D. K. Nasiopou- los, “Llms in e-commerce: a comparative
analysis of gpt and llama models in product review evaluation,” Natural
2. Language Processing Journal, vol. 100056, pp. 1–6, 2024.
3. T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P.
Shyam, G. Sas- try, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger,
4. T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu,
5. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin,
6. S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish,
7. Radford, I. Sutskever, and D. Amodei, “Language models are few-shot learners,” in Advances in Neural
In- formation Processing Systems (NeurIPS 2020), vol. 33, 2020, pp. 1877–1901. OpenAI, “Gpt-4
technical report,” 2023.
8. Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, and Wang, “Retrieval-augmented generation for
large lan- guage models: A survey,” 2023.
9. N. Kandpal, H. Deng, A. Roberts, E. Wallace, and C. Raffel, “Large language models struggle to
learn long- tail knowledge,” in Proceedings of the International Con- ference on Machine Learning
(ICML). PMLR, 2023, pp. 15 696–15 707.
10. F. Petroni, T. Rocktäschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, and A. Miller,
“Language models as knowledge bases?” in Proceedings of the 2019 Conference on Empirical Methods
in Natural Language Processing and the 9th International Joint Conference on Natural Language
Processing (EMNLP-IJCNLP). Hong Kong, China: Association for Computational Linguistics, 2019,
pp. 2463–2473. [Online]. Available: https://www.aclweb.org/anthology/D19-1250
11. C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J.
Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” 2019. [Online].
Available: https://arxiv.org/abs/1910.10683
12. A. Roberts, C. Raffel, and N. Shazeer, “How much knowledge can you pack into the parameters of a
language model?” 2020. [Online]. Available: https://arxiv.org/abs/2002.08910
13. G. Marcus, “The next decade in ai: Four steps towards robust artificial intelligence,” 2020. [Online].
Available: https://arxiv.org/abs/2002.06177
14. K. Guu, K. Lee, Z. Tung, P. Pasupat, and M.- W. Chang, “Realm: Retrieval-augmented
language model pre-training,” 2020. [Online]. Available: https://arxiv.org/abs/2002.08909
15. V. Karpukhin, B. Oguz, S. Min, L. Wu, S. Edunov, D. Chen, and W.-t. Yih, “Dense passage
retrieval for open-domain question answering,” 2020. [Online]. Available: https:// arxiv.org/ abs/ 2004
.04906
16. F. Petroni, P. Lewis, A. Piktus, T. Rocktäschel, Y. Wu, A. H. Miller, and S. Riedel, “How context
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 201
affects language models’ factual predictions,” in Automated Knowledge Base Construction (AKBC),
2020. [Online]. Available: https://openreview.net/forum?id=025X0zPfn
17. H. Zamani, F. Diaz, M. Dehghani, D. Metzler, and M. Bendersky, “Retrieval-enhanced machine
learning,” in Proceedings of the 45th International ACM SIGIR Conference on Research and
Development in Information Retrieval (SIGIR ’22). Madrid, Spain: Association for Computing
Machinery, 2022, pp. 2875–2886. [Online]. Available: https://doi.org/10.1145/3477495.3531722
18. G. Agrawal, T. Kumarage, Z. Alghami, and H. Liu, “Can knowledge graphs reduce hallucinations
in llms?: A survey,” 2023. [Online]. Available: https://arxiv.org/abs/2311.07914
19. G. Izacard and E. Grave, “Leveraging passage retrieval with generative models for open domain question
answering,” in Proceedings of the 16th Conference of the European Chapter of the Association for
Computational Linguistics: Main Volume, P. Merlo, J. Tiedemann, and
20. R. Tsarfaty, Eds. Online: Association for Computational Linguistics, 2021, pp. 874–880. [Online].
Available: https://doi.org/10.18653/v1/2021.eacl-main.74
21. A. Salemi, S. Kallumadi, and H. Zamani, “Optimization methods for personalizing large language models
through retrieval augmentation,” in Proceedings of the 47th An- nual International ACM SIGIR
Conference on Research and Development in Information Retrieval (SIGIR ’24), Washington, DC, USA,
2024, (to appear).
22. A. Salemi, S. Mysore, M. Bendersky, and H. Zamani, “Lamp: When large language
mod- els meet personalization,” 2023. [Online]. Available: https://arxiv.org/abs/2304.11406
23. J. James and S. Es, “Ragas: Evaluation framework for your retrieval-augmented genera- tion
(rag) pipelines,” 2023. [Online]. Available: https://github.com/explodinggradients/ragas
24. J. Saad-Falcon, O. Khattab, C. Potts, and M. Za- haria, “Ares: An automated evaluation framework
for retrieval-augmented generation systems,” 2023. [Online]. Available: https://arxiv.org/abs/2311.09476
25. F. Petroni, A. Piktus, A. Fan, P. Lewis, M. Yazdani, N. De Cao, J. Thorne, Y. Jernite, V.
Karpukhin, J. Maillard, V. Plachouras, T. Rocktäschel, and S. Riedel, “Kilt: a benchmark for
knowledge in- tensive language tasks,” in Proceedings of the 2021 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (NAACL-
HLT 2021), K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tur, I. Belt- agy, S. Bethard,
R. Cotterell, T. Chakraborty, and Y. Zhou, Eds. Online: Association for Computational Linguistics,
2021, pp. 2523–2544. [Online]. Available: https://doi.org/10.18653/v1/2021.naacl-main.200
26. A. S. . P. authors, “Pymupdf tutorial — docu- mentation,” https://pymupdf.readthedocs.io/en/latest/
tutorial.html, 2025, accessed: 2025-09-25.
27. M. Fenniak and pypdf Contributors, “pypdf documen- tation,” https://pypdf.readthedocs.io/en/stable/,
2025,
28. accessed: 2025-09-25.
29. The pandas development team, “pandas documentation,” https://pandas.pydata.org, 2025, accessed:
2025-09-25.
30. H. Face, “Datasets documentation,” https://huggingface. co/docs/datasets/en/index, 2025, accessed: 2025-
09-25.
31. E. AI, “Models & languages — spacy usage documen- tation,” https://spacy.io/usage/models, 2025,
accessed: 2025-09-25.
32. N. Reimers and I. Gurevych, “all-mpnet-base- v2,” https://huggingface.co/sentence-transformers/
all-mpnet-base-v2, 2021, accessed: 2025-09-25.
33. H. Face, “Transformers documentation,” https:// huggingface.co/docs/transformers/en/index, 2025,ac-
cessed: 2025-09-25.
34. E. Gradients, “Ragas documentation,” https://docs. ragas.io/en/stable/, 2025, accessed: 2025-09-25.