Recall increase while having same HNSW parameters

[nuvotex-tk]

Hello,
when testing different DBs using https://github.com/zilliztech/VectorDBBench I found that there was an increase in recall when the dataset size increases.
Now I don't know how this increase is possible while having the same HNSW configuration throughout all tests. I looked through the documentation and couldn't find anything that could cause this. Are there maybe some dynamic adjustments to the index when the dataset is getting larger?

Benchmark results where achieved using the following configuration:

• Milvus standalone deployed with 4vCPU and 70GiB RAM

HNSW parameters used:

• M: 16
• efConstruction: 200
• ef: 100

[yhmo]

768D 1M = 3GB
768D 10M = 30GB
1536D 50K = 300MB
1536D 500K = 3GB
1536D 5M = 30GB

In milvus, segments can be merged into 1GB per segment. Each segment has an independent index with the same index parameters.
300MB = 1 segment
3GB = 3 segments
30GB = 30 segments

A search request will scan each segment and retrieve the topk nearest items from each segment. Finally, N * topk items are merged into one topk result set.
Let's say, ideally, each segment's recall rate is 95%. Assume topk = 100.

If there is only 1 segment, 100 items are retrieved immediately, the recall rate of a search request is 95%.
If there are 30 segments, 30 * 100 items are retrieved in the first step. In the second step, milvus sorts the 3000 items by distance, picks the top 100 items as the final result. The recall rate of the first step is 95%. Is the recall rate of the second step higher than the first step? It is a mathematical proof problem. I don't know how to prove it, but I believe the answer is yes.