Overall Comparison #

PatANN consistently outperforms leading vector database libraries across multiple datasets and metrics. Our comprehensive benchmarks below demonstrate PatANN’s superior performance in query time, recall, and throughput.

PatANN Benchmark

Algorithm Performance Comparison #

Algorithm	Recall @ 50000 QPS	Recall @ 100000 QPS	QPS Geometric Mean	QPS @ 95%	Epsilon Median	AUC Normalized
PatANN	0.99991	0.99991	182,526	357,321	0.99718	223,090
ScaNN	0.93428	-	40,818	81,245	0.97712	57,529
HNSWlib	0.68671	-	20,504	53,226	0.96344	32,429
FAISS-IVFPQFS	0.85664	0.56150	15,845	302,140	0.88150	71,710
Milvus-IVFSQ8	-	-	495	2,219	0.99487	859
QDrant	-	-	14	74	0.99397	26

Note: A dash (-) indicates no data available at that performance level.

PatANN delivers the best combination of speed and accuracy across all metrics. Our comprehensive benchmarks show PatANN has 68% higher AUC than HNSWlib and 4.6x better QPS at 95% recall. These metrics demonstrate PatANN’s superior efficiency in both the speed-accuracy tradeoff (AUC) and practical high-throughput scenarios. The pattern-aware partitioning approach particularly excels on large, diverse datasets, providing dramatic improvements in real-world performance.

These results are for in-memory index implementations. PatANN On-Disk has an even better memory footprint, while HNSWlib does not support On-Disk functionality. All benchmarks were conducted in our standardized benchmark environment as detailed in the Benchmark Environment section.