feat(internal/types): support eagerly validating pydantic iterators

Author: stainless-app[bot]

src/openai/_models.py

@@ -27,7 +27,9 @@
     Protocol,
     Required,
     Sequence,
+    Annotated,
     ParamSpec,
+    TypeAlias,
     TypedDict,
     TypeGuard,
     final,
@@ -81,7 +83,15 @@
 from ._constants import RAW_RESPONSE_HEADER
 
 if TYPE_CHECKING:
+    from pydantic import GetCoreSchemaHandler, ValidatorFunctionWrapHandler
+    from pydantic_core import CoreSchema, core_schema
     from pydantic_core.core_schema import ModelField, ModelSchema, LiteralSchema, ModelFieldsSchema
+else:
+    try:
+        from pydantic_core import CoreSchema, core_schema
+    except ImportError:
+        CoreSchema = None
+        core_schema = None
 
 __all__ = ["BaseModel", "GenericModel"]
 
@@ -422,6 +432,76 @@ def model_dump_json(
             )
 
 
+class _EagerIterable(list[_T], Generic[_T]):
+    """
+    Accepts any Iterable[T] input (including generators), consumes it
+    eagerly, and validates all items upfront.
+
+    Validation preserves the original container type where possible
+    (e.g. a set[T] stays a set[T]).  Serialization (model_dump / JSON)
+    always emits a list — round-tripping through model_dump() will not
+    restore the original container type.
+    """
+
+    @classmethod
+    def __get_pydantic_core_schema__(
+        cls,
+        source_type: Any,
+        handler: GetCoreSchemaHandler,
+    ) -> CoreSchema:
+        (item_type,) = get_args(source_type) or (Any,)
+        item_schema: CoreSchema = handler.generate_schema(item_type)
+        list_of_items_schema: CoreSchema = core_schema.list_schema(item_schema)
+
+        return core_schema.no_info_wrap_validator_function(
+            cls._validate,
+            list_of_items_schema,
+            serialization=core_schema.plain_serializer_function_ser_schema(
+                cls._serialize,
+                info_arg=False,
+            ),
+        )
+
+    @staticmethod
+    def _validate(v: Iterable[_T], handler: "ValidatorFunctionWrapHandler") -> Any:
+        original_type: type[Any] = type(v)
+
+        # Normalize to list so list_schema can validate each item
+        if isinstance(v, list):
+            items: list[_T] = v
+        else:
+            try:
+                items = list(v)
+            except TypeError as e:
+                raise TypeError("Value is not iterable") from e
+
+        # Validate items against the inner schema
+        validated: list[_T] = handler(items)
+
+        # Reconstruct original container type
+        if original_type is list:
+            return validated
+        # str(list) produces the list's repr, not a string built from items,
+        # so skip reconstruction for str and its subclasses.
+        if issubclass(original_type, str):
+            return validated
+        try:
+            return original_type(validated)
+        except (TypeError, ValueError):
+            # If the type cannot be reconstructed, just return the validated list
+            return validated
+
+    @staticmethod
+    def _serialize(v: Iterable[_T]) -> list[_T]:
+        """Always serialize as a list so Pydantic's JSON encoder is happy."""
+        if isinstance(v, list):
+            return v
+        return list(v)
+
+
+EagerIterable: TypeAlias = Annotated[Iterable[_T], _EagerIterable]
+
+
 def _construct_field(value: object, field: FieldInfo, key: str) -> object:
     if value is None:
         return field_get_default(field)

tests/test_models.py

@@ -1,15 +1,16 @@
 import json
-from typing import TYPE_CHECKING, Any, Dict, List, Union, Optional, cast
+from typing import TYPE_CHECKING, Any, Dict, List, Union, Iterable, Optional, cast
 from datetime import datetime, timezone
-from typing_extensions import Literal, Annotated, TypeAliasType
+from collections import deque
+from typing_extensions import Literal, Annotated, TypedDict, TypeAliasType
 
 import pytest
 import pydantic
 from pydantic import Field
 
 from openai._utils import PropertyInfo
 from openai._compat import PYDANTIC_V1, parse_obj, model_dump, model_json
-from openai._models import DISCRIMINATOR_CACHE, BaseModel, construct_type
+from openai._models import DISCRIMINATOR_CACHE, BaseModel, EagerIterable, construct_type
 
 
 class BasicModel(BaseModel):
@@ -961,3 +962,56 @@ def __getattr__(self, attr: str) -> Item: ...
     assert model.a.prop == 1
     assert isinstance(model.a, Item)
     assert model.other == "foo"
+
+
+# NOTE: Workaround for Pydantic Iterable behavior.
+# Iterable fields are replaced with a ValidatorIterator and may be consumed
+# during serialization, which can cause subsequent dumps to return empty data.
+# See: https://github.com/pydantic/pydantic/issues/9541
+@pytest.mark.parametrize(
+    "data, expected_validated",
+    [
+        ([1, 2, 3], [1, 2, 3]),
+        ((1, 2, 3), (1, 2, 3)),
+        (set([1, 2, 3]), set([1, 2, 3])),
+        (iter([1, 2, 3]), [1, 2, 3]),
+        ([], []),
+        ((x for x in [1, 2, 3]), [1, 2, 3]),
+        (map(lambda x: x, [1, 2, 3]), [1, 2, 3]),
+        (frozenset([1, 2, 3]), frozenset([1, 2, 3])),
+        (deque([1, 2, 3]), deque([1, 2, 3])),
+    ],
+    ids=["list", "tuple", "set", "iterator", "empty", "generator", "map", "frozenset", "deque"],
+)
+@pytest.mark.skipif(PYDANTIC_V1, reason="this is only supported in pydantic v2")
+def test_iterable_construction(data: Iterable[int], expected_validated: Iterable[int]) -> None:
+    class TypeWithIterable(TypedDict):
+        items: EagerIterable[int]
+
+    class Model(BaseModel):
+        data: TypeWithIterable
+
+    m = Model.model_validate({"data": {"items": data}})
+    assert m.data["items"] == expected_validated
+
+    # Verify repeated dumps don't lose data (the original bug)
+    assert m.model_dump()["data"]["items"] == list(expected_validated)
+    assert m.model_dump()["data"]["items"] == list(expected_validated)
+
+
+@pytest.mark.skipif(PYDANTIC_V1, reason="this is only supported in pydantic v2")
+def test_iterable_construction_str_falls_back_to_list() -> None:
+    # str is iterable (over chars), but str(list_of_chars) produces the list's repr
+    # rather than reconstructing a string from items. We special-case str to fall
+    # back to list instead of attempting reconstruction.
+    class TypeWithIterable(TypedDict):
+        items: EagerIterable[str]
+
+    class Model(BaseModel):
+        data: TypeWithIterable
+
+    m = Model.model_validate({"data": {"items": "hello"}})
+
+    # falls back to list of chars rather than calling str(["h", "e", "l", "l", "o"])
+    assert m.data["items"] == ["h", "e", "l", "l", "o"]
+    assert m.model_dump()["data"]["items"] == ["h", "e", "l", "l", "o"]