feat(langchain): reference model profiles for provider strategy (#33974)

Author: ccurme

libs/langchain_v1/langchain/agents/factory.py

@@ -63,6 +63,18 @@
 
 STRUCTURED_OUTPUT_ERROR_TEMPLATE = "Error: {error}\n Please fix your mistakes."
 
+FALLBACK_MODELS_WITH_STRUCTURED_OUTPUT = [
+    # if langchain-model-profiles is not installed, these models are assumed to support
+    # structured output
+    "grok",
+    "gpt-5",
+    "gpt-4.1",
+    "gpt-4o",
+    "gpt-oss",
+    "o3-pro",
+    "o3-mini",
+]
+
 
 def _normalize_to_model_response(result: ModelResponse | AIMessage) -> ModelResponse:
     """Normalize middleware return value to ModelResponse."""
@@ -349,11 +361,13 @@ def _get_can_jump_to(middleware: AgentMiddleware[Any, Any], hook_name: str) -> l
     return []
 
 
-def _supports_provider_strategy(model: str | BaseChatModel) -> bool:
+def _supports_provider_strategy(model: str | BaseChatModel, tools: list | None = None) -> bool:
     """Check if a model supports provider-specific structured output.
 
     Args:
         model: Model name string or `BaseChatModel` instance.
+        tools: Optional list of tools provided to the agent. Needed because some models
+            don't support structured output together with tool calling.
 
     Returns:
         `True` if the model supports provider-specific structured output, `False` otherwise.
@@ -362,11 +376,26 @@ def _supports_provider_strategy(model: str | BaseChatModel) -> bool:
     if isinstance(model, str):
         model_name = model
     elif isinstance(model, BaseChatModel):
-        model_name = getattr(model, "model_name", None)
+        model_name = (
+            getattr(model, "model_name", None)
+            or getattr(model, "model", None)
+            or getattr(model, "model_id", "")
+        )
+        try:
+            model_profile = model.profile
+        except ImportError:
+            pass
+        else:
+            if (
+                model_profile.get("structured_output")
+                # We make an exception for Gemini models, which currently do not support
+                # simultaneous tool use with structured output
+                and not (tools and isinstance(model_name, str) and "gemini" in model_name.lower())
+            ):
+                return True
 
     return (
-        "grok" in model_name.lower()
-        or any(part in model_name for part in ["gpt-5", "gpt-4.1", "gpt-oss", "o3-pro", "o3-mini"])
+        any(part in model_name.lower() for part in FALLBACK_MODELS_WITH_STRUCTURED_OUTPUT)
         if model_name
         else False
     )
@@ -988,7 +1017,7 @@ def _get_bound_model(request: ModelRequest) -> tuple[Runnable, ResponseFormat |
         effective_response_format: ResponseFormat | None
         if isinstance(request.response_format, AutoStrategy):
             # User provided raw schema via AutoStrategy - auto-detect best strategy based on model
-            if _supports_provider_strategy(request.model):
+            if _supports_provider_strategy(request.model, tools=request.tools):
                 # Model supports provider strategy - use it
                 effective_response_format = ProviderStrategy(schema=request.response_format.schema)
             else:

libs/langchain_v1/pyproject.toml

@@ -57,6 +57,7 @@ test = [
     "pytest-mock",
     "syrupy>=4.0.2,<5.0.0",
     "toml>=0.10.2,<1.0.0",
+    "langchain-model-profiles",
     "langchain-tests",
     "langchain-openai",
 ]
@@ -75,6 +76,7 @@ test_integration = [
     "cassio>=0.1.0,<1.0.0",
     "langchainhub>=0.1.16,<1.0.0",
     "langchain-core",
+    "langchain-model-profiles",
     "langchain-text-splitters",
 ]
 
@@ -83,6 +85,7 @@ prerelease = "allow"
 
 [tool.uv.sources]
 langchain-core = { path = "../core", editable = true }
+langchain-model-profiles = { path = "../model-profiles", editable = true }
 langchain-tests = { path = "../standard-tests", editable = true }
 langchain-text-splitters = { path = "../text-splitters", editable = true }
 langchain-openai = { path = "../partners/openai", editable = true }

libs/langchain_v1/tests/unit_tests/agents/test_response_format.py

@@ -790,7 +790,7 @@ def wrap_model_call(
         # Track which model is checked for provider strategy support
         calls = []
 
-        def mock_supports_provider_strategy(model) -> bool:
+        def mock_supports_provider_strategy(model, tools) -> bool:
             """Track which model is checked and return True for ProviderStrategy."""
             calls.append(model)
             return True