Qdrant connector docs: VectorSchema(dtype=np.float32) (the class) fails to reload on the second run #2535

Describe the bug

The Qdrant connector docs pass the np.float32 class (not a np.dtype instance) to VectorSchema:

from cocoindex.resources.schema import VectorSchema
import numpy as np

schema = await qdrant.CollectionSchema.create(
    vectors={
        "text": qdrant.QdrantVectorDef(schema=VectorSchema(dtype=np.float32, size=384), distance="cosine"),
        ...
    }
)

— see the Named vectors and Dense + sparse vectors sections of https://cocoindex.io/docs/connectors/qdrant/ (source: docs/src/content/docs/connectors/qdrant.mdx).

The first cocoindex update succeeds. The second run fails while CocoIndex reconciles its persisted collection-tracking record:

cocoindex._internal.serde.DeserializationError: Failed to deserialize msgspec payload
  (type_hint=cocoindex.connectorkits.statediff.MutualTrackingRecord[cocoindex.connectors.qdrant._target._CollectionTrackingRecordCore],
   source=prev_possible_records param of cocoindex.connectors.qdrant._target._CollectionHandler.reconcile())
_pickle.UnpicklingError: Forbidden global during unpickling: numpy.float32

Why

VectorSchema (a msgspec.Struct) declares dtype: np.dtype. np.float32 is a class, so msgspec reaches enc_hook, whose type is type → in _SERIALIZE_BY_PICKLE_TYPES → the value is pickled by reference to the global numpy.float32. On reload, _RestrictedUnpickler only allows ("numpy", "dtype"), numpy.ndarray, the registered dtype subclasses and _frombuffer — not the numpy.float32 scalar type class — so it raises.

Passing an instance (np.dtype(np.float32) or np.dtype("float32")) round-trips fine, because ("numpy", "dtype") is allow-listed.

To Reproduce

# run 1: OK, run 2: DeserializationError
import numpy as np, cocoindex as coco
from cocoindex.connectors import qdrant
from cocoindex.resources.schema import VectorSchema

schema = await qdrant.CollectionSchema.create(
    vectors={"v": qdrant.QdrantVectorDef(schema=VectorSchema(dtype=np.float32, size=384))}
)
await qdrant.mount_collection_target(DB, "coll", schema)

Minimal check of the round-trip, no indexing needed:

import io, pickle, numpy as np
from cocoindex._internal.serde import _RestrictedUnpickler
_RestrictedUnpickler(io.BytesIO(pickle.dumps(np.float32, 5))).load()        # UnpicklingError: numpy.float32
_RestrictedUnpickler(io.BytesIO(pickle.dumps(np.dtype(np.float32), 5))).load()  # OK -> float32

Expected behavior

Either the docs use np.dtype(np.float32), or (better) VectorSchema normalizes/coerces dtype to a np.dtype instance (e.g. np.dtype(dtype)) so the documented class form works and is persisted safely.

Versions

• cocoindex==1.0.25, qdrant-client==1.16.x, numpy 2.x
• macOS arm64, Python 3.12 (uv)

Workaround

Use VectorSchema(dtype=np.dtype(np.float32), size=...) (or VectorSchema(dtype=np.dtype("float32"), ...)).

Possibly related to #2011 (different root cause: numpy 1.x _frombuffer module path), but this one reproduces on numpy 2.x and is about the dtype argument form.