qdrant_client.get_fastembed_vector_params() with upload_collection

[uguraydrd]

Hello folks!

Firstly, I would like to run fastembed integration. I suppose there is a bug or problem or I am wrong. I am going to explain this problem.

Normally, you know we follow these steps for example if we want to give the vector size:

qdrant_client = QdrantClient("http://localhost:6333")
qdrant_client.set_model("BAAI/bge-small-en-v1.5")
qdrant_client.recreate_collection(
    collection_name="test",
    vectors_config=VectorParams(size=384, distance=Distance.COSINE),
)

We know there is no problem here.

But if we want the vector size to be retrieved dynamically based on the model we follow these steps as in documentation like:

qdrant_client = QdrantClient("http://localhost:6333")
qdrant_client.set_model("BAAI/bge-small-en-v1.5")
qdrant_client.recreate_collection(
    collection_name="test",
    vectors_config=qdrant_client.get_fastembed_vector_params()
)

unfortunately my vectors are not loaded into the collection. I do not get any error but my vectors are not loaded either.

We know vectors_config has two typehint like Union[types.VectorParams, Mapping[str, types.VectorParams]]. We also have a mapping. After calling qdrant_client.get_fastembed_vector_params() , it returns a dictionary for us to feed to vectors_config like this:

{
    "fast-bge-small-en-v1.5": models.VectorParams(
        size = embeddings_size,
        distance = distance,
        ...
    )
}

Finally,

I check the your dev codes. Based on my many tests and observations, I noticed if I do this there is no problem but this is not a solution, just an indication that it works if I do this:

qdrant_client = QdrantClient("http://localhost:6333")
qdrant_client.set_model("BAAI/bge-small-en-v1.5")
qdrant_client.recreate_collection(
    collection_name="test",
    vectors_config=qdrant_client.get_fastembed_vector_params()[self.client.get_vector_field_name()]
)

because if I do this 'vectors_config' is no longer a dictionary but a pydantic class and equals a value like below. It can then be verified directly by models.VectorParams.

VectorParams(size=384, distance=<Distance.COSINE: 'Cosine'>, ...)

What is your comment on this topic? Please let me know what is the right approach or solution?

Cheers!

--
Uguray

[joein]

Hi @uguraydrd

Could you show please how you upload points to your collection?

[uguraydrd]

Hi @joein,

Of course. I do this:

metadata = data["col"].to_dict(orient="records")
model = qdrant_client.embedding_models[qdrant_client.embedding_model_name]

qdrant_client.upload_collection(
    collection_name="test",
    vectors=[
        arr.tolist()
        for arr in list(
            model.passage_embed(
                texts=data,
                batch_size=128,
            )
        )
    ],
    ids=range(len(data)),
    payload=metadata,
    batch_size=128,
)

[joein]

Fastembed creates a named vector with this name fast-bge-small-en-v1.5 in your case, however you are uploading unnamed vectors, that's why you they are not added to your collection.

You don't see any error on python client side, because upload_collection has param wait set to False by default, which means that client does not wait for the response from the server.

Here is a link to an example with upsert

You can check accepted param types of vectors in upload_collection signature.
You can also find examples of usage it with various input types in tests

[uguraydrd]

Thank you very much for your information. Finally, It works for me, but normally, I could not find or see this anything on your tutorials unfortunately . When I analyzed qdrant_client.add() deeply, I noticed the situation you mentioned.

I want to put the working code here, maybe it will be useful to other people.

metadata = data["col"].to_dict(orient="records")
model = qdrant_client.embedding_models[qdrant_client.embedding_model_name]
vector_field_name = qdrant_client.get_vector_field_name()

qdrant_client.upload_collection(
    collection_name="test",
    vectors=[
        {self.vector_field_name: arr.tolist()}
        for arr in self.model.passage_embed(
            texts=data,
            batch_size=128,
        )
    ],
    ids=range(len(data)),
    payload=metadata,
    batch_size=128,
)