In this work, we present PARSAL, a retrieval pipeline to obtain relevant scientific articles in a standardized format, given some relevant keyword. The pipeline exploits the API of scientific publishers to retrieve relevant full-text articles in PDF, JSON, or XML format. Afterwards, a parser was implemented to standardize the retrieved articles in a unique format, thus they can be inserted in a Mongo DB database and accessed via a custom GUI. In addition, papers are arranged in a Knowledge Graph, built via LLamaIndex framework, to allow users to make queries to the collected articles and obtain a verbose answer. The code of the developed pipeline, GUI and Knowledge Graph creation and inference is available on GitHub1.
can be seen in Figure 1.
The field of scientific research is constantly expanding and is going through a golden moment in the
digital era. In fact, scientific output across all disciplines is growing exponentially year after year [
Early projections suggest that this publishing trend will continue to grow over the next few years, increasing the number of peer-reviewed articles available to the scientific community, which will
CEUR Workshop
ISSN1613-0073
end up with a growing number of articles to read and review. In fact, bibliographic analysis is an
important phase in scientific research and an essential preliminary step in any work. In emerging
ifelds, where new articles are published daily, it is desirable to keep track of continuous changes and
highlight trends to build a reliable systematic overview of a given research topic. In Natural Language
Processing (NLP) field, the advent of the concept of Retrieval-Augmented Generation (RAG) [
In this paper, we introduce PARSAL, an automatic open-source retrieval pipeline that can talk to all APIs of the most popular publishing houses and, through an intuitive graphical interface, download the most relevant scientific articles based on keyword-search strategy. In addition, PARSAL automatically formats documents, simplifying the complex pre-processing phase, making them suitable for building RAG-based inference systems. Thus, the main contributions of this work can be summarized as follows. 1. An extraction pipeline across API of diferent editors given a target keyword; 2. A parser which processes retrieved articles in XML, JSON or PDF and return them in a standardized
JSON format; 3. A Graphical User Interface (GUI) based over a Mongo DB Database, to interactively access the retrieved articles; 4. An end-to-end solution for supporting the build of an efective RAG architecture; This work is organized as follows. Section 2 reports an overview of the state-of-the-art approaches. Section 3 reports the API available to download full-text articles, while the retrieval pipeline is presented in Section 4 along with its GUI. To assess the potentialities of the proposed pipeline, a case study is presented in the field of drug discovery, in which a Knowledge Graph (KG) has been created and queried accordingly, as a retrieval source in a RAG setup (Section 5), and concluding remarks are drawn in Section 6.
The availability of advanced tools has recently simplified automatic access to the full-text of scientific
articles, responding to the need to manage an increasingly vast academic output. Over the years,
various open source bots and scripts have been developed to meet this demand, capable of integrating
diferent sources. PyPaperBot 1 is an example of combining searches in Google Scholar and CrossRef to
sequentially download PDFs from DOI lists. Similarly, many oficial platforms provide APIs to facilitate
automatic content retrieval: both options for metadata retrieval from a given DOI of full-text PDF with
services such as CrossRef API [
At the same time, the scientific community has developed tools to support the literature reviews based
on NLP techniques. A good example is LiteRev [
The collected texts are then preprocessed and represented using text mining techniques (e.g., TF-IDF
vectorization), and then analyzed with dimensional reduction and unsupervised clustering algorithms.
In this way, LiteRev organizes the corpus of documents into key topics, highlighting groups of works
with related keywords, and facilitates the identification of relevant research topics [
Another recent development involves the usage of LLMs to query and synthesize scientific knowledge,
combining automatic document retrieval with LLM generative capabilities. PaperQA [
Finally, several researchers are exploring the use of RAG to automate the review of the scientific
literature. For example, Han et al. provide an exhaustive review of the role of RAG in automating
literature reviews [
Our proposed pipeline, following an open-source philosophy, in contrast with tools such as ScopusAI3, permits a customized keyword-based article retrieval from multiple editors’ API. The flexibility of the framework allows us to arrange collected articles in both structured formats (XLM, JSON) and PDFs, in a standard format, thus allowing the building of both a custom retrieval and generative component, which better meets the user’s needs. To stress this point, we propose a case study in the field of drug discovery, where the retrieval component consists of a KG. Given a user query, the output obtained may consist of both a verbose answer and a list of interesting sources.
A core point of the proposed pipeline is to obtain the full-text article in an automatic manner. As a consequence, a study was conducted on the APIs available from the most popular and relevant publishers for the scientific community. The idea is to involve APIs that allow the search of relevant articles given a target keyword. More in detail, the available APIs can be mapped as: • Metadata APIs, which provide only bibliographic information; • Full-text APIs, which allow full access to the textual context of the requested article. It can be restricted to open-access publications or include no open-access articles with previous authentication. 3https://www.elsevier.com/products/scopus/scopus-ai
A complete overview of selected publishers and characteristics of the available API is reported in the following subsections, while their main characteristics are summarized in Table 1. Despite MDPI ofers valuable API endpoint, it has not been included in this analysis as its automatic usage for download full-text articles is not permitted.
ACL Anthology hosts open-access articles related to computational linguistics and NLP. Articles can be accessed via their Python library 4 which allows interaction with the entire articles repository in both streaming access and fully local. Articles have to be searched programmatically since a custom search or ranking algorithm is not implemented, in which custom filters may be implemented over the articles’ metadata, such as keyword, title, or publishing year.
Once the articles’ ID are collected, full-text PDFs can be downloaded without any authentication, as all articles in the anthology are open-access. 3.2. API @ arXiv arXiv is a free distribution service and an open-access archive of pre-print articles not peer-reviewed, and it mainly serves as a platform for sharing improvements in researcher’s work and for articles waiting to be published in journals or presented to conferences.
Its native open repository [
In the context of this work, both APIs have been used, and retrieval was limited to ScienceDirect openaccess articles. During this process, some limitations have been encountered, such as the maximum number of results per query, the maximum quota available per user, and the rate limit per request.
Springer Nature is a German-British academic publishing company created by the fusion of Springer Science Business Media and Nature Publishing Group. Articles published are open-access and accessible under payment, and APIs allow full-text download for open-access articles, while institutional API access is required for downloading no open-access articles.
Springer Metadata API and Springer Meta API allow retrieval of metadata information about all articles
in the catalog, while Springer Open Access API allows both retrieval of metadata and downloadable
full-text content in JSON, PDF, and JATS XML format [
Due to the lack of an API endpoint to retrieve identifiers of relevant articles given a target keyword or
other criteria, CrossRef API7 [
In this Section, the proposed retrieval pipeline (Figure 2) is described, which is composed of three diferent modules, namely the API module, the OCR module, and the Parser module. The API module employs the publicly accessible APIs described in Section 3, for automatic retrieval of relevant articles given a keyword-based query. The articles obtained in XML, JSON, or PDF format are then processed to obtain a unique and standardized format. The OCR module is used for extracting the textual content from PDF leveraging an Optical Character Recognition (OCR) model, while the parser module creates a unique standard JSON-based format from either XML or JSON files as well as the obtained from the OCR model.
To properly process articles in PDF format, extraction of their textual content is needed. Despite some metadata information can also be obtained from the PDF (e.g. authors, title, abstract), they can be easier accessed via Metadata API queries. In addition, metadata retrieved from API provides a native well-formatted structure of that information. Therefore, the focus in the OCR module is limited towards the extraction of the actual textual content of the article, as the title of the sections, and the corresponding content.
As the core of the OCR module, the Open Language Model OCR (olmOCR) [13] was used. olmOCR is an open source toolkit developed by Allen Institute for AI (AI2), based on Qwen2-VL-7B-Instruct [14], a 7B parameters Visual Language Model (VLM). It has been fine-tuned over the olmOCR-mix-0225 data
set [15], which consists of 260.000 pages from various PDF files, including scientific papers, books, and legal documents [13].
Other OCR models have been considered, such as Tesseract [16], EasyOCR [17], PaddleOCR [18], docTR [19]) but their performance was poor compared to olmOCR both in terms of accuracy and computational resources needed (GPU and inference time). In addition, the aforementioned models needed an additional external model for layout analysis, such as LayoutParser [20], to correctly process the complex layout of scientific articles. olmOCR, on the other hand, has an embedded component for layout analysis of the given document, thus allowing the automatic recognition of images, multi-column content, tables, and equation.
To extract the textual content of a given PDF file, olmOCR converts each page into a high-resolution image over which a layout analysis is performed and then the textual content is extracted and arranged in textual (txt) and markdown (md) format to preserve the structure of the provided document. The generated output can then be used to extract the target textual content.
The parser module receives as input the metadata information of the given article, the XML or JSON full-text articles obtained via API requests, and the output of the OCR module. Its objective is to provide, for each diferent full-text file format, a unique JSON file following a well-defined schema, including both metadata and textual content in the given article as in Figure 3. The so obtained parsed documents, sharing a standard and common format, serve as input to build an end-to-end RAG architecture, as shown in Section 5. 4.3. GUI In addition to the proposed retrieval pipeline, a user interface was developed using the CustomTkinter [21] Python library based on Tkinter. This library allows for the creation of custom and intuitive graphical interfaces thanks to its large number of widgets. The interface allows users to run queries in a simple way. The left section of the interface allows users to enter keywords using the Search Keyword bar, select the starting year for the article search, and finally select the publishing houses included in PARSAL. Once all these parameters have been set, the search can be started by clicking on the Start Research button, and the results will be displayed in the central box of the app. The query results can be downloaded or filtered using the check boxes next to each article. For each article, the title, DOI, associated keywords, and abstract will be returned, allowing the user to evaluate them manually. A snapshot of the GUI is shown in Figure 4.
{ } `doi': `10.1016/j.rico.2024.100489', `title': `Satellite imagery, big data, IoT and deep learning techniques
for wheat yield prediction in Morocco', `authors': [`Abdelouafi Boukhris', `Antari Jilali',
`Abderrahmane Sadiq'], `keywords': [`Satellite imagery', `IoT', `ArcGis',
`Deep learning', `RMSE', `NoSQL Database'], `sections': { `Introduction': `The efficient oversight of agricultural operations,
ensuring [...]', `Crop wheat yield prediction in morocco: a proposed system': `Yield
prediction plays [...]', }, `abstract': `In the domain of efficient management of resources and
ensuring nutritional consistency, accuracy [...]', `editor': `Elsevier'
A case study in the field of computer-aided drug discovery was conducted to test the eficiency of the proposed pipeline. This topic was selected because it can be validated by authors who are experts in the field in terms of the validity of the articles retrieved, adherence to the reference topic, and the subsequent construction and querying of the KG.
In particular, five meaningful keywords were carefully chosen as meaningful for the target field of interest in the selected study case and have been considered both in their expanded version and in their condensed form, as they are used in scientific publications. Therefore, Drug Discovery, Bi Accuracy Prediction, Graph Convolutional Network (GCN), Ligand Based Virtual Screening (LBVS) and Structure Based Virtual Screening, have been used to retrieve full-text articles from ACL Anthology, ArXiv, Elsevier, Springer, and Wiley, as reported in Table 2. Despite the wide availability of articles that are freely accessible or obtainable thanks to agreements between our University and publishers, only a relatively small quantity was efectively available for full-text download in an automatic manner.
Once the articles have been downloaded, we ensured that any duplicates were deleted by DOI checking, as the same article can be retrieved twice, e.g. among its keywords appear both “GCN” and “Drug Discovery”. In this phase, each API was queried with a single keyword, so the aforementioned duplications can be noticed. Despite some publisher allow for complex queries involving logical OR or AND, we decide to use a single query for each keyword and its abbreviated form, to simulate the same retrieval conditions. After this retrieval phase, articles in XML and JSON format are injected into a parser, while the retrieved PDF are preliminary analyzed by the OCR module and then parsed to the same unique format (Section 4.2).
The articles obtained in JSON format are then loaded into the proposed GUI app (Section 4.3) to be browsed interactively. In addition, we explored an interaction modality based on Generative AI where the user’s query leverages a KG built from the collected articles. In this section, both the KG building and inference processes are reported. The experiments, for both the OCR and parser module, KG building and inference have been executed on a cluster with 1 NVIDIA A100 64 GB GPUs from the Leonardo supercomputer8 via an ISCRA-C application.
The objective is to create a semantic Knowledge Graph (KG) in which meaningful relationships between diferent papers can be highlighted. To do it, each parsed article is arranged in multiple documents, each containing a section of the original article in textual form, with the corresponding metadata, including the DOI, authors, name of the article, reference section and keywords, as reported in Figure 5.
Document( metadata={ `doi': `10.1016/j.rico.2024.100489', `authors': [`Abdelouafi Boukhris', `Antari Jilali', `Abderrahmane
Sadiq'], `keywords': [`Satellite imagery', `Iot', `Arcgis', `Deep learning',
`RMSE', `NoSQL Database'], `title': `Satellite imagery, big data, IoT and deep learning
techniques for wheat yield prediction in Morocco', `section': `abstract', `editor': `Elsevier' }, textual_document={ `text': `In the domain of efficient management of resources and
ensuring nutritional consistency, accuracy [...]' )
}
From the collected and parsed articles, a KG was automatically built using the LlamaIndex Framework9. Starting from the collected articles, the KG is created querying an external LLM to extract the semantic triplets found in the provided textual content and arrange them in a graph-like structure, thus enabling inter-document connections.
To this aim, diferent decoder-only generative model [ 22] has been used to create and query the built KG. In particular, we choose among instruction-tuned LLMs of 4B parameters, namely Phi 4 mini [23] and Gemma 3 4B [24], and 8B parameters such as Llama 3.1 8B [25], Qwen 3 8B [26]. It is worth noticing that the choice of the generative model is crucial, since diferent models generate diferent KGs, despite having been injected with the same documents. We are also aware that more powerful models can have been considered, such as GPT-4o [27] or Gemini [28], but their usage is under paid APIs and our priority was open-source models and reproducibility with relatively small computational resources.
To create the KG, documents have been divided into chunks of 4096 tokens for each of which a maximum of 5 triplets are extracted, and all selected LLMs were used with the generative parameters suggested by their authors. A prompt-based generative approach was used to extract relevant semantic triplets from the given documents. This approach is eficient since it is fully automatized and the obtained KGs are completely data driven and do not include any external knowledge injection, as for the typology of relations to be extracted. On the other hand, less significant triplets with respect to more meaningful ones can be extracted and included in the KG.
In Figure 6 a simplified graphical version of the built KG is reported that is obtained by selecting the “Drug-target interaction prediction” node and including its neighbors up to 2-hop distance10.
Once the KG was built, we selected five meaningful questions, reported in Table 3. The answer to these questions can be obtained using multiple articles, thus stressing the retrieval and generative capabilities of the KG coupled with the considered LLMs.
More in detail, we adopted the LlamaIndex query engine, which generates the desired answer from the KG in multiple steps, involving the diverse LLMs considered as a generative model. First, the input query is analyzed to extract the meaningful words in the query itself (keywords11). The obtained results are then used to retrieve the more meaningful triplets and documents contained in the graph, and,
What is drug discovery? What AI tools can be used in this field? How can a GCN be used in drug discovery applications? Which molecular descriptors are widely used in virtual screening approaches? Which neural networks are most widely used in ligand-based virtual screening approaches? Which explainability methods are used in drug discovery approaches? in a generative manner, the retrieved content is used as a context for the LLM to properly answer the user’s question. If the retrieved context exceeds the maximum number of tokens for the model window, multiple queries are given to the LLM with the newer context and the yet generated answer, until the last retrieved context is reached and the last answer generated. KGs created with the selected LLMs have been queried (Table 3) and the retrieved and generated content has been evaluated and the computational statics as for the inference time and the GPU required are reported in Table 4.
For each question we evaluated the percentage of meaningful retrieved content: 4B models cannot perform well in the retrieval phase and fail since no context has been retrieved. In contrast, 8B models achieve comparable performances in the retrieval phase, with 70% of relevant context retrieved. As models have not been fine-tuned for both KG creation and inference, we argue that the diferent native instruction fine-tuning process over which models have been trained on, played a crucial role, along with models’ parameter size. More in detail, the proposed application employs not only the purely verbose generative capabilities of selected models, but also their abilities in traditional NLP tasks such as Relation Extraction in KG creation and the retrieval phase itself.
As for performance, while real-time applications with 16GB GPU on average can be implemented smaller models, 8B models needed more than 32 GB of GPU and more than one minute on average per each answer. As for the result obtained, the second group of larger models represents the minimum satisfactory level in terms of both performance and computational requirements.
In Table 5 the generated response to Q3 and the DOI of the retrieved articles are reported12. Regarding the validity of the generated content, all responses were evaluated by domain experts, and they found them definitely correct. Answers have been considered correct also in the case in which a model replies that no relevant context has been provided to proper answer to the given question, coherently to the human-judgment in Table 4.
Overall, both Llama 3.1 and Qwen 3 in instruct and 8B parameters version, result valuable alternative in the proposed KG and RAG-based application.
Improvement on the presented approach can be performed in terms of the retrieval phase, involving Graph DB and SPARQL queries. In addition, this can be beneficial in terms of eficiency in the case of a 12The full generated answers can be downloaded in from the GitHub repository
Which molecular descriptors are widely used in virtual screening approaches?
Qwen 3 8B 2D descriptors such as MACCS keys, PubChem fingerprints, and molecular formulae are widely used in virtual screening approaches.
The answer is based on the knowledge that molecular descriptors are used in virtual screening, and the knowledge that 2D descriptors include MACCS keys, PubChem fingerprints, and molecular formulae.
Now, answer the following query using the knowledge from the provided context information and not prior knowledge.
Sources
arXiv:2404.04559, arXiv:2411.12748 larger retrieval corpus. In this case, more time would be required for the retrieval pipeline for both the download and parsing phases with the OCR and Parser modules as in Section 4) and the KG generation phase.
Bibliographic research of the scientific literature is a crucial step in any new research activity. The downloading and organizing of articles is a time-consuming process that requires automated systems to keep up with the number of articles published every day in several research domains. In this work, we present PARSAL, an automatic retrieval pipeline aimed at the scientific community in the process of retrieving and structuring the scientific literature. PARSAL relies on a set of dedicated APIs to download either native PDF or JSON / HTML / XML articles from the repositories of the main open-access scientific publishers. PDF documents are converted in text through olmOCR, and all articles have been transformed in a suitable JSON format, thus enabling the creation of a MongoDB instance that can be browsed using an easy-to-use graphical interface. One of the main objectives of the paper is to demonstrate that such an information organization can be a valuable support for advanced retrieval systems based on generative AI.
To this aim, we validated PARSAL in the field of drug discovery. 1,840 articles were downloaded and formatted from various publishers based on 7 keywords in order to build a Knowledge Graph to be used for the RAG approach through diferent LLMs. Some domain experts formulated five questions that were posed to the KG; they assessed the responses of the system and found them satisfactory both as regards the retrieved references and the actual content, when Llama 3.1 and Qwen 3 8B have been used in their native instruct version.
This work is supported by the cup project B73C22000810001, project code ECS_00000022 “SAMOTHRACE” (Sicilian MicronanoTech Research And Innovation Center), and the cup project J73C24000070007, ”CAESAR” (Cognitive evolution in Ai: Explainable and Self-Aware Robots through multimodal data processing). The works presented were partially developed on the Leonardo supercomputer with the support of CINECA-Italian Super Computing Resource Allocation class C project IscrC_DOCVLM2 (HP10C97VNN).
During the preparation of this work, the authors used DeepL and Quillbot for translation, grammar and spelling check. After using these tools, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content. [11] S. E. Robertson, H. Zaragoza, M. Taylor, Simple bm25 extension to multiple weighted fields, in: Proceedings of the Thirteenth ACM International Conference on Information and Knowledge Management, CIKM ’04, ACM, New York, NY, USA, 2004, pp. 42–49. URL: https://doi.org/10.1145/ 1031171.1031181. doi:10.1145/1031171.1031181. [12] Wiley Online Library, Wiley text and data mining api documentation, 2024. URL: https:// onlinelibrary.wiley.com/library-info/resources/text-and-datamining, accessed: 2025-01-26. [13] J. Poznanski, J. Borchardt, J. Dunkelberger, R. Huf, D. Lin, A. Rangapur, C. Wilhelm, K. Lo, L. Soldaini, olmocr: Unlocking trillions of tokens in pdfs with vision language models, arXiv preprint arXiv:2502.18443 (2025). URL: https://arxiv.org/abs/2502.18443. arXiv:2502.18443, allen Institute for AI. [14] P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y. Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, J. Lin, Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution, arXiv preprint arXiv:2409.12191 (2024). arXiv:2409.12191. [15] J. Poznanski, A. Rangapur, J. Borchardt, others, olmOCR-mix-0225 Dataset, https://huggingface.
co/datasets/allenai/olmocr-mix-0225, 2024. [16] R. Smith, An overview of the Tesseract OCR engine, in: Proceedings of the Ninth International Conference on Document Analysis and Recognition (ICDAR 2007), volume 2, IEEE, Curitiba, Brazil, 2007, pp. 629–633. doi:10.1109/ICDAR.2007.4376991. [17] JaidedAI, EasyOCR: Ready-to-use OCR with 80+ supported languages, 2020. URL: https://github.
com/JaidedAI/EasyOCR, apache License 2.0. [18] P. Authors, Paddleocr, awesome multilingual ocr toolkits based on paddlepaddle., https://github.
com/PaddlePaddle/PaddleOCR, 2020. [19] Mindee, docTR: Document text recognition, 2021. URL: https://github.com/mindee/doctr, apache
License 2.0. [20] Z. Shen, R. Zhang, M. Dell, B. C. G. Lee, J. Carlson, W. Li, Layoutparser: A unified toolkit for deep learning based document image analysis, arXiv preprint arXiv:2103.15348 (2021). [21] T. Schimansky, Customtkinter: A modern and customizable python ui-library based on tkinter, 2022. URL: https://github.com/TomSchimansky/CustomTkinter, version 5.2.2. [22] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin,
Attention is All you Need, in: Advances in Neural Information Processing Systems, 2017. [23] A. Abouelenin, A. Ashfaq, A. Atkinson, H. Awadalla, N. Bach, et al., Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs, arXiv preprint arXiv:2503.01743 (2025). [24] GemmaTeam, A. Kamath, J. Ferret, S. Pathak, N. Vieillard, R. Merhej, et al., Gemma 3 Technical
Report, arXiv preprint arXiv:2503.19786 (2025). [25] LlamaTeam, The Llama 3 Herd of Models, arXiv preprint arXiv:2407.21783 (2024). [26] QwenTeam, Qwen3 Technical Report, arXiv preprint arXiv:2505.09388 (2025). [27] OpenAI, Gpt-4o system card, arXiv preprint arXiv:2410.21276 (2024). [28] GeminiTeam, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al., Gemini: A Family of Highly Capable Multimodal Models, 2024.