PAPER / ARXIV:2609.13452
Song Young Oh, Amal Gueroudji, Seth Ockerman, Rob Latham, Orcun Yildiz, Ian Foster, Kyle Chard, Robert Ross
RESUMO
Vector databases hash-partition data across shards, destroying semantic locality and forcing scatter-gather queries limited by the slowest shard. COMPASS uses a knowledge graph to determine data placement and query-time shard selection, detecting communities and routing queries to a small set of shards. Across four biomedical knowledge graphs, COMPASS searches only 13-18% of the corpus while preserving recall and recovering up to 2.6x more multi-hop evidence than an embedding-based baseline, sustaining 7.9x higher throughput on 15 HPC nodes.
NO MESMO MAPA