Skip to content

The agent engineering platform.

fix(core): avoid `dict` shadowing in language models (#38480)

Mason Daughertyd3eb2296e72c
diff --git a/libs/core/langchain_core/language_models/base.py b/libs/core/langchain_core/language_models/base.py
index a39979d14c..44eb7aeee5 100644
--- a/libs/core/langchain_core/language_models/base.py
+++ b/libs/core/langchain_core/language_models/base.py
@@ -2,6 +2,7 @@
from __future__ import annotations
+import builtins # noqa: TC003 # runtime-evaluated; subclass `dict()` shadows the builtin
import warnings
from abc import ABC, abstractmethod
from collections.abc import Callable, Mapping, Sequence
@@ -191,7 +192,7 @@ class BaseLanguageModel(
tags: list[str] | None = Field(default=None, exclude=True)
"""Tags to add to the run trace."""
- metadata: dict[str, Any] | None = Field(default=None, exclude=True)
+ metadata: builtins.dict[str, Any] | None = Field(default=None, exclude=True)
"""Metadata to add to the run trace."""
custom_get_token_ids: Callable[[str], list[int]] | None = Field(
diff --git a/libs/core/langchain_core/language_models/chat_models.py b/libs/core/langchain_core/language_models/chat_models.py
index 01737308eb..3ed4b9455c 100644
--- a/libs/core/langchain_core/language_models/chat_models.py
+++ b/libs/core/langchain_core/language_models/chat_models.py
@@ -3,7 +3,7 @@
from __future__ import annotations
import asyncio
-import builtins # noqa: TC003
+import builtins # noqa: TC003 # runtime-evaluated; subclass `dict()` shadows the builtin
import contextlib
import inspect
import json
diff --git a/libs/core/tests/unit_tests/language_models/test_imports.py b/libs/core/tests/unit_tests/language_models/test_imports.py
index d99f37f99d..cfd2cc18fb 100644
--- a/libs/core/tests/unit_tests/language_models/test_imports.py
+++ b/libs/core/tests/unit_tests/language_models/test_imports.py
@@ -1,4 +1,9 @@
-from langchain_core.language_models import __all__
+from typing import Any
+
+from langchain_core.callbacks import Callbacks
+from langchain_core.language_models import BaseLanguageModel, __all__
+from langchain_core.outputs import LLMResult
+from langchain_core.prompt_values import PromptValue
EXPECTED_ALL = [
"BaseLanguageModel",
@@ -26,3 +31,41 @@ EXPECTED_ALL = [
def test_all_imports() -> None:
assert set(__all__) == set(EXPECTED_ALL)
+
+
+def test_pydantic_rebuild_handles_subclass_dict_method_shadowing_builtin() -> None:
+ """Regression for Pydantic field collection with subclasses that define `dict()`.
+
+ Pydantic 2.14.0a1 evaluates inherited field annotations during subclass
+ rebuilds. If `BaseLanguageModel.metadata` uses a plain `dict[...]`
+ annotation, the subclass `dict()` method can shadow the builtin and make
+ annotation evaluation fail with `'function' object is not subscriptable`.
+ """
+
+ class DictMethodLanguageModel(BaseLanguageModel[str]):
+ name: str = "test"
+
+ def generate_prompt(
+ self,
+ prompts: list[PromptValue],
+ stop: list[str] | None = None,
+ callbacks: Callbacks = None,
+ **kwargs: Any,
+ ) -> LLMResult:
+ raise NotImplementedError
+
+ async def agenerate_prompt(
+ self,
+ prompts: list[PromptValue],
+ stop: list[str] | None = None,
+ callbacks: Callbacks = None,
+ **kwargs: Any,
+ ) -> LLMResult:
+ raise NotImplementedError
+
+ def dict(self, **_kwargs: Any) -> dict[str, Any]:
+ return {}
+
+ DictMethodLanguageModel.model_rebuild(force=True)
+
+ assert "metadata" in DictMethodLanguageModel.model_fields