Fix Pydantic 2.12+ compatibility for custom FieldInfo with Annotated types

[abrookins]

Pydantic 2.12+ FieldInfo Subclass Regression

The Problem

Pydantic Issue: #12359 — FieldInfo.from_annotations no longer returns subclasses of FieldInfo.

What Changed in Pydantic 2.12

Pydantic 2.12 introduced a complete refactor of the FieldInfo class (PRs #11388 and #11898) to fix ~10 separate bugs.
This refactor changed the behavior of FieldInfo.from_annotation():

• Pydantic 2.11.x and earlier:
  When a custom FieldInfo subclass appeared in Annotated metadata, from_annotation() preserved the subclass type.

• Pydantic 2.12+:
  from_annotation() now returns a plain PydanticFieldInfo instance instead of preserving the subclass.

How This Affected redis-om-python

Redis OM Python uses a custom FieldInfo subclass with additional Redis-specific attributes:

class FieldInfo(PydanticFieldInfo):
    def __init__(self, default: Any = ..., **kwargs: Any) -> None:
        # Custom attributes for Redis indexing
        self.primary_key = kwargs.pop("primary_key", False)
        self.sortable = kwargs.pop("sortable", None)
        self.index = kwargs.pop("index", None)
        self.full_text_search = kwargs.pop("full_text_search", None)
        # ...

When users define fields with Annotated types like Coordinates:

class Location(JsonModel, index=True):
    coordinates: Coordinates = Field(index=True)

Where Coordinates is:

Coordinates = Annotated[
    CoordinateType,
    PlainSerializer(...),
    BeforeValidator(parse_redis),
]

Under Pydantic 2.12+, the internal steps are:

1. Process the Annotated type and read metadata.
2. Call FieldInfo.from_annotation().
3. Receive a plain PydanticFieldInfo (not the custom subclass).
4. Lose custom attributes:
   • index
   • sortable
   • primary_key
   • full_text_search
   • vector_options
   • case_sensitive

The Fix

The fix preserves the original custom FieldInfo objects by capturing them before Pydantic's metaclass runs and restoring them afterward:

# Before Pydantic processes the model:
original_field_infos: Dict[str, FieldInfo] = {}
for attr_name, attr_value in attrs.items():
    if isinstance(attr_value, FieldInfo):
        original_field_infos[attr_name] = attr_value

# After Pydantic processes the model:
if type(field) is PydanticFieldInfo:
    if field_name in original_field_infos:
        field = original_field_infos[field_name]
        field.annotation = pydantic_field.annotation
        field.metadata = pydantic_field.metadata

Will Pydantic Restore the Original Behavior?

No. Per Pydantic issue #12374:

1. The old behavior was considered buggy — even older versions failed when multiple FieldInfo instances existed in metadata.
2. Subclassing FieldInfo is discouraged — composition is recommended instead:

   “If possible, I would encourage libraries to avoid subclassing FieldInfo, and use composition if possible.”

3. No guarantee of full compatibility:

   “Many external libraries are setting too many expectations on the FieldInfo logic… I don't know if we'll be able to fully preserve compatibility for every use case.”

Long-term Implications

The workaround in this PR is appropriate for now.
A more future-proof approach would be refactoring redis-om-python to use composition (custom metadata in Annotated) rather than inheritance, though this would be a larger change.

[justin-cechmanek]

LGTM!