Hnsw config update

[davidschrooten]

Hello,

I am running qdrant v1.14 in kubernetes. With 46 million dense and sparse vectors. And have a few questions.

When I update the collection config to change hnsw config values such as ef_construct, m values and on_disk will the index be rebuild. Or do I need to reinsert the data for the changes to be reflected.

Also I see a hnsw config can also be configured for named vectors separately. Currently I just seems to have it configured for the whole collection. Is this wrong? And whats the difference when configuring it specifically within named vector config.

Thanks,

David

[timvisee]

When I update the collection config to change hnsw config values such as ef_construct, m values and on_disk will the index be rebuild. Or do I need to reinsert the data for the changes to be reflected.

Yes, the indices will be rebuilt. This can take some time, and will be done once the collection reports a green status.

Also I see a hnsw config can also be configured for named vectors separately. Currently I just seems to have it configured for the whole collection. Is this wrong? And whats the difference when configuring it specifically within named vector config.

Configuring it at collection level is perfectly fine, and is the most common way to configure it. If you ever want a different configuration for a specific named vector (maybe tune space vs accuracy, or don't generate a global HNSW at all), you can configure it there.

To be specific, by default it'll use the collection level HNSW configuration for all vectors. But if you specify a different configuration for a specific named vector, it'll use that instead.