Andrey Vasnetsov

icon indicating a website link https://t.me/neural_network_engineering icon indicating an email [email protected]

icon company Qdrant icon location Berlin icon location CTO, creator of Qdrant. Interested in Vector Search, Metric Learning, Self-Supervised and One-Shot learning.

Results 529 comments of Andrey Vasnetsov

[experiment] Hnsw support for duplicated vectors

The approach didn't work, the update heuristic breaks the index in case of too many duplicates

Read-only API key

/bounty $250

memory leak

Hi @hyhanyan, it might be not a memory leak, but caching and memory fragmentation. This is an expected behavior, as allocator does not necessary return memory to the OS if...

memory leak

I am going to close this issue, as this is not a memory leak. Please feel free to open another issue or re-open this one if you will find other...

Render effective configuration

it is `details_level`

Support for older GLIBC version in binary releases

Resolved by introducing MUSL builds

could you tell me the maximum number of collections to create? Is one empty collection consumes ~10MB memory?

This tutorial might be relevant to your use-case https://qdrant.tech/documentation/tutorials/multiple-partitions/

a collection, contains ~170000 points with 1536 embedding-dimension, and some payloads, one search consumes ~700ms, is that normal?

Hi @chendulearn, could you pls share collection info and example of requests you are doing

a collection, contains ~170000 points with 1536 embedding-dimension, and some payloads, one search consumes ~700ms, is that normal?

@chendulearn please also share an example of request

a collection, contains ~170000 points with 1536 embedding-dimension, and some payloads, one search consumes ~700ms, is that normal?

You are using filters without the payload index. That is, most likely, the reason for the slow search