Fix Bit constructor for NumPy arrays on NumPy >= 1.24 (np.bool removal)

[HUIIIM]

— vertciti AI team (assisting Jiahui Miao)

Summary

Bit(np.array([True, False, True])) raises AttributeError: module 'numpy' has no attribute 'bool' on NumPy >= 1.24, because pgvector/bit.py compares value.dtype against np.bool, which was removed in NumPy 1.24 (deprecated since 1.20). One-line fix: compare against np.bool_ instead.

Reproduction

import numpy as np
from pgvector import Bit
Bit(np.array([True, False, True]))
# AttributeError: module 'numpy' has no attribute 'bool'.

Root cause

In the np.ndarray branch of Bit.__init__, if value.dtype != np.bool: — np.bool no longer exists on NumPy >= 1.24, so attribute access itself raises before any comparison happens. The other constructors (list/str/bytes) and to_numpy() are unaffected.

Fix

-            if value.dtype != np.bool:
+            if value.dtype != np.bool_:

Rationale: np.bool_ is the NumPy scalar type NumPy's own migration guidance points to. np.dtype(bool) == np.dtype(np.bool_), so the comparison semantics are unchanged — this only restores the pre-1.24 behavior, no behavior change.

Note: np.bool also appears in the type annotations in the same file, but the module uses from __future__ import annotations, so those are lazily evaluated and never raise at runtime — left untouched to keep the PR minimal. Happy to fix the annotations too if preferred.

Verification performed

• Bit(np.array([True, False, True])) works after the fix
• np.uint8 0/1 arrays still accepted (the np.unpackbits compatibility path)
• np.uint8 arrays with values > 1 and float64 arrays still raise ValueError
• 2-D arrays still raise ValueError
• Roundtrip Bit(np.array([True, False, True])).to_numpy().tolist() == [True, False, True]
• list/str/bytes constructors unaffected

Regression suggestion

Consider adding a test guarding the dtype comparison across the supported NumPy range, e.g. asserting Bit(np.array([True, False, True])).to_numpy().tolist() == [True, False, True].

[ankane]

Thanks @HUIIIM. fwiw, numpy.bool was restored in NumPy 2.