Document Type : Research Paper

Author

• Amirali Ghajari

Islamic Azad University Central Tehran Branch

Abstract

State-of-the-art static Approximate Nearest Neighbor (ANN) search methods, like HNSW, are inefficient for dynamic environments due to costly index rebuilds. This paper addresses this gap by proposing the Hybrid Graph-Tree (HGT), a novel data structure for high-performance ANN search on streaming data. HGT synergistically combines a global navigational tree for rapid search space pruning with localized navigable graphs at its leaves for accurate local search. A key feature is an efficient leaf-splitting mechanism that maintains index balance and performance during continuous insertions without global reconstruction. Our extensive experiments demonstrate that HGT achieves query performance competitive with static HNSW while offering orders-of-magnitude faster insertions. The structure’s ability to maintain stable query latency and high recall under dynamic workloads establishes it as a robust solution for next-generation vector databases and real-time AI systems, bridging the critical gap between static index performance and dynamic data requirements.

Keywords

• Vector Databases
• Streaming Data
• Computational Geometry
• Curse of Dimensionality
• Vector Embeddings
• Incremental Indexing
• Metric Indexing
• Low-Latency Updates