Japanese Reading Performance of VisionAPI

[sum]

Hi.
I am the manager of a project to introduce a new version of VisionAPI "builtin/latest".
I am not an engineer, so I would appreciate it if you could give me a brief explanation.

I was using the old version "builtin/legacy" to read Japanese receipts with VisionAPI.
When I switched the API to "builtin/latest" in the same environment, the Japanese reading performance seemed to drop.

Specifically, Japanese "half-width katakana" characters were misread more often.
Some examples of "half-width katakana" are shown below.

ｱ, ｲ, ｳ, ｴ, ｵ
ｶ, ｷ, ｸ, ｹ, ｺ
etc...

I would like to know possible reasons for the increase in misreading of Japanese between the new version and the old version.

Thank you in advance.