Slow Query

[wilsonlimaneto]

Hello I just created an account (free tier) end using Deepset haystack, I wrote about 600 1024 dim vectors. To query those (also using haystack), it took me 157 seconds. My collection was setup like:

client.recreate_collection(
    collection_name="document_vector",
    vectors_config=models.VectorParams(size=1024, distance=models.Distance.COSINE),
    quantization_config=models.ProductQuantization(
        product=models.ProductQuantizationConfig(
            compression=models.CompressionRatio.X16,
            always_ram=True,
        ),
    ),
)

I do not know if this is due to the free tier or collection setup.

[generall]

it might be possible if you are dropping connection on each request of something like this.
Could you please share a code to reproduce?

[wilsonlimaneto]

Hello... well I firstly ran this code and created the collection

client.recreate_collection(
    collection_name="document_vector",
    vectors_config=models.VectorParams(size=1024, distance=models.Distance.COSINE),
    quantization_config=models.ProductQuantization(
        product=models.ProductQuantizationConfig(
            compression=models.CompressionRatio.X16,
            always_ram=True,
        ),
    ),
)

Then, I wrote using haystack... 600 Document objects with ids and embeddings with this:

qdrant_document_store = QdrantDocumentStore("https://<<>>", index="document_vector", api_key="<<>>", embedding_dim=1024, recreate_index=False, embedding_field='embedding', duplicate_documents='skip', similarity='cosine' )
qdrant_document_store.write_documents(lst_Documents_vetor)

And I queried like:

str_query = "very small string"
retriever_qd = EmbeddingRetriever(document_store=qdrant_document_store, use_gpu=False, progress_bar=False, embedding_model='path to transformer')
pipe1 = DocumentSearchPipeline(retriever_qd)
print('Querying Qdrant... ')
time1= time.time()
prediction1 = pipe1.run (query=str_query, params={"Retriever": {"top_k": 3000} })
time2 = time.time()
print('   ...  Query time:', time2-time1, 's')

[fritzprix]

@wilsonlimaneto by the way, do you have any reason for using top_k of 3000, when you have only 600 items though. I'm not sure what is your use case exactly but setting top_k as 3000 is kinda rare.