Types

VECTOR

from milvusql_sqlalchemy.types import VECTOR

Column("embedding", VECTOR(768))

Renders as VECTOR(768) in DDL — MilvusQL's FLOAT_VECTOR field type. Bind values pass through as plain Python lists; nothing is stringified (bind_processor copies the value into a new list and returns it, with no serialization), because milvusql's DBAPI takes real Python values in its parameters dict and never serializes a vector into SQL text.

Comparator methods

Deliberately parallels pgvector.sqlalchemy almost line for line:

Item.embedding.l2_distance(query_vec)        # <->
Item.embedding.cosine_distance(query_vec)    # <=>
Item.embedding.max_inner_product(query_vec)  # <#>
Item.embedding.l1_distance(query_vec)        # <+>

Each generates a custom_op via SQLAlchemy's own ColumnOperators.op() — the base SQLCompiler already knows how to render a custom op as plain infix text (left <-> right), so none of these need any compiler-side code of their own. Only the VECTOR(n) column type needed dialect work.

select(Item.id).order_by(Item.embedding.cosine_distance(query_vec)).limit(10)

SPARSEVEC

from milvusql_sqlalchemy.types import SPARSEVEC

Column("sparse_embedding", SPARSEVEC)

MilvusQL's SPARSE_FLOAT_VECTOR field type. Only .max_inner_product() is exposed — Milvus's sparse index only supports the IP metric.