[Feature Request]: Async support for Qdrant Vector store

[pradeep-suresh2002]

Feature Description

Need Async support for Async QdrantClients

Reason

To create collections

• For a PDF with 345 pages, when enable_hybrid = True is taking nearly 21 minutes when batch_size is 64 (default) and the same takes around 16 minutes when batch size is 20.
• But the same when enable_hybrid = False, irrespective of the batch size, it took around 41s.
• So I planned to try with AsyncQdrandClient, when I try to create collections asynchronously, I am facing the issue of "None object has collection exists". It looks like it's not creating the QdrantClient object. So, need of support for Async QdrantClients.
• It would also be great, to know why the reason for creating collection is huge when enable_hybrid is set True and also interested to know if there is any ways to reduce this time.

Value of Feature

• If we get the Async support, we might create this collection faster when enable_hybrid is True.
• This might greatly reduce the time required to reduce the time for creating collections

[logan-markewich]

@pradeep-suresh2002 async is supported, but you need to supply the async client

vector_store = QdrantVectorStore(
  ..., 
  client=qdrant_client.QdrantClient(...), 
  aclient=qdrant_client.AsyncQdrantClient(...)
)

However, sparse vector generation is not async, its compute bound to your machine, so it will be blocking either way

[pradeep-suresh2002]

For creating AsyncClients, why we are giving both client and aclient? We can provide only aclient right? vector_store = QdrantVectorStore( ..., client=qdrant_client.QdrantClient(...), aclient=qdrant_client.AsyncQdrantClient(...) )

• Is the sparse vector generation can be made async?
• Is there any way to decrease the computation time for sparse vector generation?
• Is there any recommended machine configuration to achieve faster computation for this ?

[logan-markewich]

Because the code uses both a sync and async client, and qdrant doesn't provide a way to create one client from the other

• It cannot (at least currently), because its running a sparse model locally
• Better hardware? Different model or approach for generating sparse embeddings?
• Probably running the sparse model on a GPU