Introduction

In pursuit of optimizing real-time vector search within a large, company-segmented dataset , a shift from the traditional single HNSW graph to a more tailored approach wa necessary. The introduction of a HNSW Manager allows for searches across multiple, segmented HNSW graphs, enhancing efficiency by directly targeting relevant company data instead of searching the entire dataset. This approach prompts a comparison of memory consumption, build time, and response time between the singular and multiple graph approaches.

HNSW Manager

Untitled

Single vs. Multiple HNSW Graphs: A Comparative Analysis

Single HNSW Graph

Total vector	num HNSW	num_vectors per HNSW	Build Time	Memory
100k	1	100k	350s	230 MB
1M	1	1M	4435s	808 MB

dim = 512, M = 32, ef_construction = 500

Multiple HNSW Graphs

Total vector	num HNSW	num vectors per HNSW	Build Time	Memory
100k	2	50k	267s	252 MB
100k	10	10k	94s	293 MB
100k	100	1k	27s	747 MB
1M	10	100k	2989s	2419 MB
1M	100	10k	909s	2887 MB
1M	1000	1k	161s	3670 MB

dim = 512, M = 32, ef_construction = 500

Key Findings

Single Graphs: are simple and more memory-efficient. But build-time takes longer
Multiple Graphs: greatly reduces build time but takes more memory usage.

Pros and Cons of Single vs. Multiple Divided HNSW Graphs

Single HNSW Graphs