Knowledge Sovereignty in Disability Information Retrieval: Architecting Privacy-Preserving and Sustainable Infrastructure

KNOWLEDGE SOVEREIGNTY IN DISABILITY INFORMATION RETRIEVAL: ARCHITECTING PRIVACY-PRESERVING AND SUSTAINABLE INFRASTRUCTURE Noor A. Fathima∗1, Noor K. Kubra2, A. Wagner1 1CERN, Geneva, Switzerland 2University of Mysore, Mysore, India Abstract Knowledge ecosystems shaped by disabled people, caregivers, and other marginalized groups remain systematically underrepresented in mainstream search and AI systems [1 – 3]. These ecosystems contain embodied practices, ethnographies, and everyday adaptations that rarely surface online, leaving critical insights absent from datasets that inform decision-making in healthcare, employment, and social inclusion [4,5]. Addressing this gap requires infrastructures that can surface and structure such knowledge in ways that are transparent, trustworthy, and open to community participation [6,7]. The OpenWebSearch.eu (OWS) initiative is developing the Open Web Index (OWI) to provide reusable, open web data for research and innovation [8, 9]. Building on this foundation, we demonstrate how a vertical search engine can be prototyped by using the OWILIX tool [10] to extract targeted slices of the OWI and consuming them within CIFFcompatible search frameworks such as MOSAIC [9, 11] for interactive exploration. This workflow serves as a technical backbone for Nooon , a privacy-preserving search engine envisioned to surface disability-related knowledge through multimodal contributions [12,13]. BACKGROUND Underrepresented Knowledge Ecosystems Mainstream search engines and AI systems prioritize large, standardized datasets [2]. As a result, knowledge ecosystems 1 shaped by marginalized groups, such as disabled people, their caregivers, and immediate families who are often the primary witnesses of daily life, remain largely invisible [1, 4]. These ecosystems are heterogeneous and context-dependent, encompassing both formal sources (e.g., policy documents, research reports) and informal contributions (e.g., lived experiences, workplace adaptations, family care practices) [5]. Within Nooon’s design, these ecosystems are recognized as microdata: fine-grained, individual-level or communitylevel traces that reveal embodied practices, ethnographic insights, and everyday adaptations [3]. Making such microdata visible, trustworthy, and reusable is essential for ∗[email protected] 1 In information science, a knowledge ecosystem refers to a dynamic and interconnected system of actors, practices, and artifacts that collectively produce, store, and circulate knowledge. addressing structural exclusion, where disability narratives are too often reduced to tokenized symbols (e.g., wheelchair icons) instead of nuanced, situated information [13]. Need for Open, Transparent Search Commercial search engines provide little visibility into their data collection and ranking processes [14, 17]. This opacity limits the ability of researchers, advocacy groups, and smaller communities to shape how their knowledge is indexed and surfaced [15,16]. Transparency and openness are therefore critical if search is to serve underrepresented groups fairly. Beyond technical access, open infrastructures also support epistemic sovereignty: the capacity of communities to contribute to and represent their knowledge on their own terms [18]. The OpenWebSearch Initiative OpenWebSearch.eu (OWS) is a European research initiative building the Open Web Index (OWI), an open, reusable corpus of web data [8]. OWI provides a foundation for building alternative and domain-specific search engines without dependence on proprietary indices [9]. By making web data accessible to researchers, developers, and civil society, OWS aims to create a sustainable and pluralistic search ecosystem in Europe and beyond [6]. The OWILIX Tool To make the OWI usable at scale, the project developed OWILIX, a command-line tool that allows targeted extraction of subsets (or “slices”) of the index [10]. OWILIX enables developers to tailor datasets to specific domains by filtering based on URLs, keywords, or metadata [9]. This functionality is particularly valuable for building vertical search engines, where precision and domain relevance are more important than exhaustive coverage. The MOSAIC Framework MOSAIC is an open-source search and visualization framework that can integrate multiple indices, including slices created with OWILIX [11]. It provides a flexible frontend for exploring search results, experimenting with ranking strategies, and evaluating the user experience [9]. For vertical search prototypes, MOSAIC offers a ready-to-use environment to demonstrate the impact of targeted indexing. https://doi.org/10.5281/zenodo.17253014 The Nooon Project Building on these infrastructures, Nooon is conceptualized as a privacy-preserving search engine designed to surface disability-related knowledge. It envisions combining slices of OWI with additional indices created through multimodal knowledge contributions (e.g., text, gestures, sensory adaptations) provided directly by disabled people and caregivers [12,13]. Nooon’s design extends the open, transparent ethos of OWS by embedding privacy, trust, and sustainability into its core [6], ensuring that sensitive disability narratives can be surfaced without fear of exploitation or misrepresentation. Scope of This Work In this paper, we focus on constructing a first proof-ofconcept slice of the OWI that we refer to as the Nooon index. The slice is obtained by selecting web documents whose HTML content contains the keyword ‘disability‘, providing a domain-specific corpus for further exploration. While this approach is deliberately simple, it establishes a reproducible baseline for evaluating how knowledge related to disability, often hidden or tokenized in mainstream systems, can be isolated as a distinct corpus. By extracting and surfacing a slice of web data where the term “disability” appears in the underlying HTML, we begin to make visible a knowledge ecosystem that is otherwise fragmented or invisible. The subsequent sections describe how this slice, referred to as the Nooon index, is prepared, structured, and made accessible as a first step toward building search infrastructures that respect and sustain marginalized knowledge. SYSTEM DESIGN & ARCHITECTURE Our design follows a modular workflow: first, identifying and extracting relevant subsets of the OWI using OWILIX; second, preparing the resulting data in formats compatible with CIFF-based search libraries; and finally, deploying the data through MOSAIC for interactive exploration. Each stage of the workflow emphasizes transparency, reproducibility, and sustainability, ensuring that the resulting slice not only serves as a technical proof of concept but also as a first step toward surfacing disability-related knowledge ecosystems in open search infrastructures. The following subsections describe this workflow in brief, from dataset access to frontend deployment. Dataset Access, Querying and Slice Selection with OWILIX To construct domain-specific corpora for our prototype, we accessed the Open Web Index (OWI) through the OWILIX command-line interface. OWILIX enables remote inspection, synchronization, and querying of OWI datasets in a manner similar to Git, providing versioned, auditable workflows. Our environment was deployed on a dedicated bare-metal server (open-science-search.ch), configured with Python 3.11 using pyenv. We created an isolated virtual environment (owi) to install the required dependencies, including py4lexis and owilix, directly from the project’s PyPI mirror. This setup ensured reproducibility and avoided dependency conflicts. The full configuration is provided in Appendix A (Puppet manifests). We began by listing available datasets to identify suitable partitions. Using the command: owilix remote ls it4i:latest files=**/language=eng/* we retrieved the most recent English-language snapshot from the IT4I data center. This confirmed the availability of a daily dataset containing over 5.6 million documents (24.3 GiB across 1,037 files). After verifying scope, we synchronized the relevant subset locally with: owilix remote pull it4i:latest "files=*/language=eng/*" This operation pulled 184 Parquet files containing both metadata and plain text. OWILIX’s incremental synchronization ensured that only new or changed files were downloaded, reducing both storage requirements and network load. By selecting only English-language partitions rather than mirroring full daily datasets ( 600 GB/day), we further minimized environmental and computational costs. This selective approach aligns with the project’s sustainability goals: smaller slices not only reduce energy intensity but also lower barriers for replication by other researchers and advocacy groups. Once the dataset was locally available, we generated a focused slice for disability-related content using a SQL-like filter: owilix query slice --local all:latest \ "where=main_content like ’%disability%’" \ collection_name="disability" \ creator="NoorAF" This created a named slice (disability) annotated with provenance metadata, including a creator tag and timestamp. Preserving these slice specifiers (it4i:latest files=... , where=...) makes the workflow reproducible and auditable, ensuring that future researchers can regenerate the exact same dataset. Together, the listing, pulling, and querying steps formed a reproducible workflow for constructing thematic corpora tailored to underrepresented knowledge ecosystems. In this study, the resulting slice served as the foundation for Nooon, which further aims to have indices with multimodal livedexperience contributions as microdata. Consuming Datasets with CIFF-compatible search frameworks Once slices of the Open Web Index (OWI) were retrieved locally, the next step was to make them consumable within https://doi.org/10.5281/zenodo.17253014 CIFF-compatible search frameworks, enabling interactive exploration. For this purpose, we first experimented with MOSAIC, the reference open-source framework developed in the OpenWebSearch.eu project. MOSAIC integrates Lucene-based search indices with Parquet-formatted metadata, providing a unified environment for retrieval and visualization. This made it well suited for our initial experiments with disability-focused slices. To prepare the OWILIX-generated slice, we exported the data into a CIFF index and corresponding Parquet files. MOSAIC requires these to be arranged in a specific directory hierarchy, with Lucene indices and metadata separated into dedicated folders. In our case, the OWILIX export produced a compressed Lucene index (index.ciff.gz) alongside Parquet metadata. These were organized under a serve/ directory, mirroring the structure expected by MOSAIC. Preparing Data for MOSAIC After constructing the disability-focused slice, we exported it from OWILIX into a format consumable by external search frameworks. OWILIX provides a local export function that outputs both a CIFF file (compressed Lucene index) and associated Parquet metadata. This export was performed as follows: cd ~/tmp/ mkdir data owilix local export all/id=id outdir=$(PWD) /data The resulting directory contained an index.ciff.gz file together with metadata files prefixed metadata_. These artifacts formed the raw input for MOSAIC. To align with MOSAIC’s expected structure, we converted the CIFF file into a Lucene index using the official converter image: mkdir -p data/serve/lucene podman run \ --rm \ -v "$PWD/data":/data:Z \ opencode.it4i.eu:5050/ openwebsearcheu-public/mosaic/lucene-ciff \ /data/index.ciff.gz \ /data/serve/lucene/disability-index The metadata was then organized into a corresponding folder: mkdir -p data/serve/metadata/ disability-index mv data/*metadata_* data/serve/metadata/ disability-index This resulted in a directory hierarchy under /tmp/data/serve/ containing both a Lucene index (disabilityindex) and the associated Parquet metadata. At the file level, the Lucene directory contained the expected segment files (e.g., _0.fdt, _0.fdx, _0_Lucene90_0.doc, segments_1), confirming that the CIFF-to-Lucene conversion had succeeded. At this stage, the slice was fully prepared for consumption by MOSAIC. lucene/ disability-index/ # Lucene index files (_0.fdt, _0.fdx, segments_1, etc.) metadata/ disability-index/ # Parquet metadata files Here our domain-specific slice, labeled disability-index is served. This ensured compatibility with MOSAIC’s configuration while preserving semantic clarity about the dataset’s thematic scope. Deployment of MOSAIC With the Lucene and Parquet structures in place, we deployed MOSAIC as a containerized service. Rather than building from source, we relied on pre-built images from the OpenWebSearch.eu GitLab registry, executed via Podman for compatibility with our infrastructure. The backend service was launched as follows: podman run -d --name mosaic --network host \ -v "$PWD/data":/data:Z \ localhost/mosaic:latest \ --lucene-dir-path $PWD/data/serve/lucene/ \ --parquet-dir-path $PWD/data/serve/metadata/ This deployment was hosted on the bare-metal server, integrated into CERN’s internal infrastructure. For security, the host is not directly exposed to the internet; instead, inbound traffic is routed via a load balancer running Nginx as a reverse proxy. This design ensured reliable user access while isolating the backend from direct external exposure. The backend search service ( mosaic ) consumed the Lucene and Parquet directories prepared in the earlier step and exposed them via a JSON API. Verification at the /mosaic/index-info endpoint confirmed that the nooon index (our disability-focused slice) was successfully loaded, containing over 2.59M English-language documents. Additional indices such as simplewiki and unis-graz were also present, demonstrating that OWILIX-exported slices can be integrated as first-class indices within MOSAIC. Deploying the MOSAIC Frontend Following backend validation, we deployed the frontend container ( mosaic-fe ), which serves a lightweight web interface built on nginx. The frontend connects directly to the backend API, enabling interactive queries and faceted exploration of multiple indices. Running backend and frontend as separate services provided flexibility: indices could be updated or swapped at the backend without interrupting the user interface, while alternative frontends could be deployed https://doi.org/10.5281/zenodo.17253014 if needed. This modularity highlights MOSAIC’s suitability for multi-tenant deployments, where diverse domain-specific corpora can be hosted and accessed through a common interface. Figure 1 shows the frontend deployment with the disability-index slice active. The interface exposes filters for language, query limits, and geographic boundaries, enabling targeted exploration of the corpus. We then tested the interface with a query for the term “disability”. As shown in Figure 2, results were retrieved from the disability-index , including documents such as legal advice directories, government manuals, and parliamentary committee reports. Each result is enriched with metadata such as language, word count, index date, and links to the original source, allowing tailored exploration of the corpora. This demonstrates full end-to-end functionality: user input via the frontend, retrieval from the Lucene/Parquet backend, and structured result presentation to the user. CONCLUSION In this paper, we demonstrated how slices of the Open Web Index (OWI) can be extracted, prepared, and consumed within the MOSAIC framework. Using the OWILIX tool, we generated a disability-focused slice, exported it into Lucene and Parquet formats, and validated it through both backend API inspection and frontend deployment. Our deployment on bare-metal infrastructure behind a secure reverse proxy confirmed that OWILIX-exported slices can be hosted, queried, and visualized interactively using MOSAIC. Beyond the technical workflow, this case study illustrates a reproducible and sustainable model for constructing vertical search engines. By embedding provenance metadata, minimizing unnecessary data transfers, and reusing opensource frameworks, we show how targeted corpora can be surfaced in ways that are transparent and resource-efficient. Furthermore, the ability to host multiple corpora concurrently demonstrates MOSAIC’s scalability for multi-domain or multi-tenant search applications. Our prototype index, Nooon , exemplifies how these infrastructures can serve underrepresented knowledge ecosystems. Disability-related corpora, often absent or misrepresented in mainstream search and AI systems, can be surfaced as first-class indices within an open, auditable framework. This approach aligns with broader goals of privacy, trust, and inclusion, ensuring that minority knowledge is not erased but made visible in ethically responsible ways. FUTURE WORK Several extensions are planned. First, we will explore topic-level filtering using curlielabels derived from Curlie.org, enabling semantic slice construction by domain hierarchy (e.g., Society → Disability,Health → Conditions) rather than keyword alone. This would yield more robust corpora aligned with specific communities of knowledge. Second, while MOSAIC provides a flexible reference framework, we anticipate that many organizations will want to augment their internal search systems with OWI data. Since internal search is often based on ElasticSearch or OpenSearch, we will extend our pipeline to support direct export from OWILIX to JSON and ingestion into OpenSearch. This requires either developing a converter or adapting OWILIX’s export functions to natively push slices into OpenSearch indices. Demonstrating this workflow would broaden adoption, as it allows OWI slices to be embedded directly into enterprise search infrastructures. Third, improvements to the MOSAIC frontend are envisaged. Current functionality supports keyword queries, language filters, and metadata inspection; future iterations will explore category-level browsing, faceted search by Curlie labels, and cross-corpus comparison (e.g., Disability in Employment vs. Disability in Education). Enhancing the frontend will make vertical search applications more intuitive for non-technical users, which is critical for adoption by advocacy groups and organizations. Finally, we plan to extend beyond English-only text to multilingual and multimodal corpora, integrating images, video, and lived-experience narratives contributed directly by users. Combined with client-side preprocessing for privacy preservation, this will bring us closer to the vision of Nooon as a privacy-first, inclusive search engine for minority knowledge ecosystems. ACKNOWLEDGEMENTS This work was supported through guidance and supervision from the Open Search Foundation, in particular its working group on Ethics [19]. Additional feedback and community input were provided by the Disability Network, an informal group within the Diversity and Inclusion programme at CERN, where the first author has served as a contact point [20]. REFERENCES [1] World Health Organization. World Report on Disability. WHO Press, 2011. https://www.who.int/ publications/i/item/9789241564182. [2] Jacob Metcalf, Emily F. Keller, and danah boyd. Perspectives on Big Data, Ethics, and Society. Council for Big Data, Ethics, and Society, 2016. https://bdes.datasociety.net/council-output/ perspectives-on-big-data-ethics-and-society/. [3] Chris Williams. Disabled Data. Routledge, 2021. ISBN: 978-0-367-12345-6 (Placeholder, no confirmed ISBN found). [4] International Labour Organization. Labour market outcomes of persons with disabilities. ILO, 2023. https://ilostat.ilo.org/ new-ilo-database-highlights-labour-market-challenges-of-persons-with-disabilities/ . [5] Gary L. Albrecht, Katherine D. Seelman, and Michael Bury (eds.). Handbook of Disability Studies. Sage, 2001. ISBN: 9780761916529. DOI: 10.4135/9781412976251. https://uk.sagepub.com/en-gb/eur/book/ handbook-disability-studies#description. https://doi.org/10.5281/zenodo.17253014 Figure 1: MOSAIC frontend deployed via the mosaic-fe container on bare-metal infrastructure. The screenshot shows the disability-index slice, containing 2.59M English-language documents. [6] OpenWebSearch.eu Consortium. D4.2: Report of privacy, transparency, and trust models for search applications. OpenWebSearch.eu, 2023. DOI: 10.5281/zenodo.placeholder (Placeholder, no confirmed DOI found). https://openwebsearch.eu/deliverables/ (General project deliverables page, specific report not confirmed). [7] Paula Helm and Selin Gerlek. Empirical AI Ethics: Reconfiguring Ethics Towards a Situated Plural and Transformative Approach. Preprint, submitted to Cambridge Forum on AI: Culture and Society, 2025. [8] OpenWebSearch.eu Consortium. Annex 1 – Description of the Action. European Commission Horizon Europe Grant Agreement, 2022. https://cordis.europa.eu/ project/id/101070014 (General project page, specific annex not publicly available). [9] OpenWebSearch.eu Consortium. The OpenWebSearch Book. 2024. https://openwebsearcheu-public. pages.it4i.eu/ows-the-book. [10] OpenWebSearch.eu Consortium. OWILIX Command Line Interface. 2024. https://opencode.it4i.eu/ openwebsearcheu-public/owi-cli/. [11] OpenWebSearch.eu Consortium. MOSAIC Search Framework. 2024. https://openwebsearcheu-public. pages.it4i.eu/ows-the-book/content/howto/c_ mosaic.html. [12] Noor A. Fathima, A. Wagner, and The OSF Working Group Ethics. Knowledge Sovereignty: Empirical Ethics in PrivacyPreserving, Sustainable Disability Information Retrieval Infrastructure. Preprint, submitted to Cambridge Forum on AI: Culture and Society, 2025. OpenWebSearch.eu / CERN, Geneva, Switzerland, 2025. https://cds.cern.ch/ record/2932730/files/CERN-OPEN-2025-004.pdf. [13] B. Blaser and R. E. Ladner. Why is Data on Disability so Hard to Collect and Understand? In: 2020 Research on Equity and Sustained Participation in Engineering, Computing, and Technology (RESPECT), Portland, OR, USA, pp. 1-8, 2020. DOI: 10.1109/RESPECT49803.2020.9272466. https:// ieeexplore.ieee.org/document/9272466. [14] Frank Pasquale. The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard https://doi.org/10.5281/zenodo.17253014 Figure 2: MOSAIC frontend in use. A query for the term “disability” against the disability-index returns diverse results, including attorney directories, insurance manuals, and parliamentary reports. Metadata (language, word count, index date) and original source links are also displayed, demonstrating end-to-end functionality. https://doi.org/10.5281/zenodo.17253014 University Press, 2015. ISBN: 9780674970847. https: //www.hup.harvard.edu/books/9780674970847. [15] Safiya Umoja Noble. Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press, 2018. ISBN: 9781479837243. https://nyupress.org/ 9781479837243/algorithms-of-oppression/. [16] Cathy O’Neil. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown Publishing Group, 2016. ISBN: 9780553418835. https: //www.penguinrandomhouse.com/books/531209/ weapons-of-math-destruction-by-cathy-oneil/. [17] Luke Stark and Kate Crawford. The Work of Art in the Age of Artificial Intelligence: What Artists Can Teach Us About the Ethics of Data Practice. Surveillance & Society, 2019. DOI: 10.22215/surjsoc.2019.10821. https://ojs.library.queensu.ca/index.php/ surveillance-and-society/article/view/10821. [18] C. Estelle Smith, Avleen Kaur, Katie Z. Gach, Loren Terveen, Mary Jo Kreitzer, and Susan O’Conner-Von. What is Spiritual Support and How Might It Impact the Design of Online Communities? Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), 2021. DOI: 10.1145/3449117. https: //doi.org/10.1145/3449117. [19] Open Search Foundation. Ethics in Search Working Group. Available at: https://ethicsinsearch.org/en/. [20] CERN Diversity and Inclusion. Disability Network. Available at: https://diversity-and-inclusion.web. cern.ch/networks/disability-network. https://doi.org/10.5281/zenodo.17253014