Description

nathanjmcdougall

opened on Dec 4, 2024

Issue body actions

Initial Checks

• I have read and followed the docs and still think this is a bug

Description

I would expect floats and ints to be consistently either able to be coerced to NdArrays or not.

Floats cannot be coerced.

Ints can be coerced - but not in the expected way. It seems to use np.empty to initialize an array with N element where N is the int being coerced.

Generally, I would expect that an int or a float would be coerced to a 1-dimensional array with a single element.

Example Code

from docarray import BaseDoc
from docarray.typing import NdArray

class MyDoc(BaseDoc):
    arr: NdArray

MyDoc(arr=10).arr # Undefined behaviour; uninitialized array
>> NdArray([7.17677097e-312, 1.77863633e-322, 0.00000000e+000,
         0.00000000e+000, 5.28555753e+180, 5.98847221e+174,
         5.92957407e-038, 7.42262642e-091, 2.80246406e-032,
                     nan])
MyDoc(arr=10.0).arr # ValidationError

I would expect instead:

MyDoc(arr=10).arr
>> NdArray([10])
MyDoc(arr=10.0).arr
>> NdArray([10.0])

Python, DocArray & OS Version

Python 3.11.9
docarray 0.40.0
Windows 10

Affected Components

• Vector Database / Index
• Representing
• Sending
• storing
• multi modal data type

Activity

1. nathanjmcdougall

   changed the title [-]NdArray inconsistently coerces floats and int[/-] [+]NdArray inconsistently coerces floats and inst[/+] on Dec 4, 2024
2. nathanjmcdougall

   changed the title [-]NdArray inconsistently coerces floats and inst[/-] [+]NdArray inconsistently coerces floats and ints[/+] on Dec 4, 2024
3. JoanFM commented on Dec 6, 2024

   

   JoanFM

   on Dec 6, 2024

   Member

   More actions

   Hey @nathanjmcdougall ,

   Thanks for reporting the issue. We will look into it. Any chance you would be able to contribute the feature?
4. nathanjmcdougall commented on Dec 6, 2024

   

   nathanjmcdougall

   on Dec 6, 2024

   Author

   More actions

   I've not contributed to DocArray before - I'll take a look at the codebase this coming week sometime to see whether this is something within my capacity.
5. 23Tarandeep57 commented on Nov 24, 2025

   

   23Tarandeep57

   on Nov 24, 2025

   More actions

   The code was using np.ndarray(value) which treats integers as shape specifications, not values. changed the code to Check for scalar types first and use np.array([value]) to explicitly wrap them in 1D arrays.

Issue actions