Abstract
We demonstrate LEGEND, a hybrid vector-graph index where knowledge graph structure guides HNSW construction, creating ''semantic shortcuts'' that improve navigation for graph-constrained queries. Unlike TigerVector and NaviX, which treat graph predicates as post-construction filters, LEGEND incorporates KG edges during neighbor selection, achieving 14-18% higher recall. Our system introduces KG-HNSW (reserving HNSW neighbor slots for KG-connected nodes), community-aligned partitioning via Louvain detection, attention-based score fusion with interpretable weight decompositions, and a learned query router selecting among four execution strategies. On FB15K-237 and a 100K-node Amazon product subgraph, LEGEND achieves 0.92 Recall@10 versus 0.78 for TigerVector, with ablations confirming KG-HNSW contributes the largest gain (+12%). Our demonstration enables attendees to visualize semantic shortcuts in real-time, inspect per-result score explanations, and stress-test recall under low-selectivity filters. A demo video is available at https://drive.google.com/file/d/1w1mN-jG69Wr0jOVCAaf_p26EmFhu6yc/view?usp=sharing [HERE].

This publication has 11 references indexed in Scilit: