Status report — 2026-05-25
Original Slovak report:
Sprava_stav_prac_ELVIRA_2026-05-25.docx· Ing. Jakub Dubec, FIIT STU · 25 May 2026
This page is an English condensation of the most recent status report on the research line "AI processing of multimodal data for the ELVIRA digital library". The Slovak original is the authoritative source.
Scope
The work extends ELVIRA — a multi-tenant OPDS publication catalog developed as open source at FIIT STU — with automated content processing. The end-to-end pipeline now runs from PDF ingest through extraction, semantic indexing and multilingual search to a conversational AI assistant and Readium LCP–protected lending.
Architecture
Two layers:
- Catalog core. Monolithic application holding metadata, accounts, permissions and loans.
- Worker fleet. Specialised microservices for text and structure extraction, equation recognition, semantic search and analytics. Scalable independently; new AI capabilities can be added without touching the catalog.
Users reach the system through the web portal and the embedded PDF viewer (with annotation support), or through standard e-reader protocols (OPDS 1.2; OPDS 2.0 in preparation). The full stack is publicly available under github.com/EvilFlowersCatalog.
Delivered capabilities
Automated content extraction
Text (including full Slovak diacritics), tables, figures and mathematical formulas are extracted automatically. The system detects whether a PDF is born-digital or scanned and routes it to direct extraction or to OCR accordingly — covering both new publications and the legacy digitised collection.
Structural decomposition
Content is processed at the level of pages, paragraphs and sentences, not as monolithic text. The same principle applies to images, formulas and tables. This unlocks sentence- and figure-level deep links and citation precision in search results.
Semantic and multilingual search
Extracted content is embedded with modern multilingual language models into a vector database. A Slovak query returns relevant passages from English, Czech and German literature alongside Slovak hits. Classical full-text search runs in parallel; the combined result blends keyword recall with semantic precision.
AI assistant over library content
A working AI assistant is deployed in the development environment. It uses the semantic search backend and returns answers grounded in specific source passages with inline citations. Production rollout to the academic community follows internal testing.
Readium LCP–protected lending
ELVIRA implements the Readium LCP open standard end to end — license issuance, return, renewal. Official EdRLab certification is in progress, which will unlock work with commercial publisher titles.
Dataverse integration
The catalog is wired to a Dataverse instance. Datasets published into public Dataverse storage automatically appear in the catalog, improving discoverability of STU research outputs.
Deployment status
| Environment | Status |
|---|---|
elvira.fiit.stuba.sk (FIIT STU) | Production |
elvira.stuba.sk (STU-wide) | Production |
elvira.digital (project home) | Public project site |
| AI microservices | Internal dev → staged rollout to production |
Publications and theses
Student theses (FIIT STU)
- Kubis — Cross-Lingual Citations Verification Using Multilingual Transformer Models. Contributes to ELVIRA's semantic layer.
- Zrutta — YoGaDoc: Graph-Based Caption-of Relationship Prediction for Semantic Indexing of Scientific Documents. Auto-assigns captions to figures in scientific documents.
Peer-reviewed publications
Indexed in IEEE Xplore:
Next steps
- Complete official Readium LCP certification with EdRLab.
- Promote AI microservices from development to production.
- Open the AI assistant to the wider academic community.
- Publish a consolidated paper on the integrated pipeline.
Contact
Ing. Jakub Dubec — Faculty of Informatics and Information Technologies, Slovak University of Technology in Bratislava.