[Question]: How to specify datatype Float16 of embeddings in Milvus vector store?

[shizhengLi]

[Question]: How to specify datatype Float16 of embeddings in Milvus vector store?

[dosubot]

To specify the datatype Float16 for embeddings in the Milvus vector store, you need to modify the collection creation parameters to include the desired datatype. However, the current implementation of the MilvusVectorStore class in LlamaIndex does not directly support setting the datatype for embeddings to Float16. The pymilvus library version used in LlamaIndex only supports float32 for vector fields ^[1].

Here's how you can modify the MilvusVectorStore class to specify the datatype for embeddings, assuming future support for Float16:

from pymilvus import DataType

# Modify the collection creation to include the datatype
self.milvusclient.create_collection(
    collection_name=self.collection_name,
    dimension=self.dim,
    primary_field_name=MILVUS_ID_FIELD,
    vector_field_name=self.embedding_field,
    id_type="string",
    metric_type=self.similarity_metric,
    max_length=65_535,
    consistency_level=self.consistency_level,
    fields=[
        {"name": self.embedding_field, "type": DataType.FLOAT_VECTOR, "params": {"dim": self.dim, "dtype": "float16"}},
        {"name": MILVUS_ID_FIELD, "type": DataType.INT64, "is_primary": True},
        # Add other fields as necessary
    ]
)

Please note that you need to ensure that both the pymilvus library and the Milvus server version you are using support the float16 datatype. As of now, the pymilvus library version used in LlamaIndex does not support float16 for vector fields ^[2][1].

To continue talking to Dosu, mention @dosu.