Hyperbolic Search on Zilliz Cloud

This post measures the performance of hyperbolic embeddings on Zilliz Cloud, comparing raw and centered vectors. It demonstrates that centering vectors improves recall@10 by 88% and that a single search can achieve 0.98 recall. It also quantifies capacity, showing that a 1 CU cluster can hold at least 2.8 million 513-dimensional vectors. Latency benchmarks compare different search levels and strategies, including reranking within Zilliz Cloud, showing that centered Strategy A at level=10 achieves 0.977 recall at 3.75 ms with high throughput, and that reranking within the database is faster than application-side reranking for larger candidate sets.