The recall rate of glass is lower than hnswlib under the same parameters.

[cwj0bzxg]

In the hnswlib library, there is a distinction between internal IDs and external IDs. However, I noticed that in glass's hnsw index, what is returned during the search is the internal ID of the neighbor. When using single thread, the internal ID and external ID are completely consistent, so returning the internal ID will not affect the recall rate. However, in the case of multi-threading, this can become a problem.

After a series of experimental verifications, I confirmed the existence of this problem.
Server configuration: 96 cores, 512G memory
Dataset: deep1M, deep10M
Method：glass(hnsw)，hnswlib
Parameters: R=32(M=16), efc=200, efs=500

Result(recall@100)：
deep1M: glass(90.0532%), hnswlib(99.3282%)
deep10M: glass(92.6417%), hnswlib(97.8977%)

[cwj0bzxg]

It is worth noting that when using multi-core, recall is not stable because the internal ID and external ID are inconsistent due to concurrency issues.

[Wainberg]

@hhy3 my own tests confirm that Glass's multi-threaded HNSW is unusable due to this bug, please fix it!