LogosKG: Scalable and Interpretable Knowledge Graph Retrieval

Published

July 1, 2025

LogosKG: Hardware-Optimized Knowledge Graph Retrieval

LogosKG is a model-agnostic engine for efficient, scalable, and interpretable multi-hop retrieval over large knowledge graphs. It represents graph traversal as hardware-efficient operations over decomposed subject, object, and relation matrices, reducing the memory and pointer-chasing bottlenecks of conventional graph systems.

For billion-edge biomedical graphs, LogosKG combines degree-aware partitioning, cross-graph routing, and on-demand caching. Its integration with large language models supports evidence-grounded clinical reasoning and enables analysis of how graph topology shapes LLM diagnostic decisions.

Publication

Published at the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026).