Designing Knowledge Graph Enhanced Retrieval Systems for Institutional Data

Abstrak

This demo article explores the design of a retrieval system that combines semantic search with an institutional knowledge graph. The graph represents fictional policies, organizational units, services, documents, and their temporal relationships, while vector retrieval identifies passages relevant to a natural-language query. A routing layer decides when graph traversal, text retrieval, or a hybrid strategy should be applied. Evaluation uses synthetic documents and invented questions that test entity resolution, multi-hop relationships, source freshness, and citation completeness. The simulated results suggest that graph constraints can improve precision for relationship-heavy questions and help expose conflicting versions of institutional records. The approach also introduces maintenance costs, including entity governance, schema evolution, and synchronization with source systems. The proposed design therefore emphasizes provenance, access control, and explicit uncertainty. Every document, entity, metric, and result described in this paper is fictional. It is original demo content prepared solely to test the OJS user interface, search index, article metadata, and PDF galley workflow.