An ultra-scalable graph-based nearest neighbor indexing algorithm that builds state-of-the-art indexes up to 11.6× faster than Vamana (DiskANN) and 12.9× faster than HNSW. PiPNN uses HashPrune, a novel online pruning algorithm that enables efficient billion-scale index construction on a single machine.
Loading more......
PiPNN (Pick-in-Partitions Nearest Neighbors) is a research breakthrough in graph-based ANN index construction published in February 2026. The algorithm fundamentally addresses the "search bottleneck" that existing graph-based methods suffer from during index construction.
HashPrune is a novel online pruning algorithm that dynamically maintains sparse collections of edges during graph construction. This innovation enables:
The algorithm employs a three-step process:
Speed Improvements:
Quality: Produces indexes that achieve higher query throughput than competing methods
Scalability: Enables building high-quality ANN indexes on billion-scale datasets in under 20 minutes using a single multicore machine
Unlike traditional approaches that require significant intermediate memory, HashPrune guarantees bounded memory during index construction. This makes PiPNN practical for building very large indexes on machines with limited RAM.
PiPNN represents a significant advancement in making graph-based ANN indexing practical at billion-scale. The combination of speed, quality, and memory efficiency makes it particularly valuable for production systems.
Published as arXiv preprint arXiv:2602.21247 (2026) by Rubel, Tobias, et al.