From 74e605a31465f74ef602c0af91ca4dfbc4f8a9f0 Mon Sep 17 00:00:00 2001
From: Sydney Runkle <sydneymarierunkle@gmail.com>
Date: Fri, 24 Apr 2026 09:30:16 -0400
Subject: [PATCH] perf(core): cache _format_tool_to_openai_function per tool
 instance
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Stash the OpenAI function description dict on the BaseTool instance under
`tool.__dict__["_openai_function_dict"]`. BaseTool.__setattr__ already pops
`tool_call_schema` and `args` when `args_schema`, `description`, or `name`
change; extend the invalidation set to include the new key so the cache
matches the schema caching lifecycle.

Previously, every call to `convert_to_openai_tool(tool)` re-ran
`schema.model_json_schema()` on the cached tool_call_schema pydantic model,
rebuilding the full JSON-schema tree on every model invocation. Summarization
middleware's `count_tokens_approximately` (called twice per model call) plus
the prompt-caching middleware's `bind_tools` meant three fresh schema
generations per model call × 15-ish tools × 500 model calls in a 100-turn
agent run — tens of seconds of pydantic work that's identical every time.

With this cache the first call pays the schema-gen cost once per tool; all
subsequent calls are a dict lookup.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 libs/core/langchain_core/tools/base.py        |  1 +
 .../langchain_core/utils/function_calling.py  | 59 +++++++++++--------
 2 files changed, 37 insertions(+), 23 deletions(-)

diff --git a/libs/core/langchain_core/tools/base.py b/libs/core/langchain_core/tools/base.py
index 013c6a43c52..70900f971a0 100644
--- a/libs/core/langchain_core/tools/base.py
+++ b/libs/core/langchain_core/tools/base.py
@@ -560,6 +560,7 @@ class ChildTool(BaseTool):
         if name in {"args_schema", "description", "name"}:
             self.__dict__.pop("tool_call_schema", None)
             self.__dict__.pop("args", None)
+            self.__dict__.pop("_openai_function_dict", None)
 
     @property
     def is_single_input(self) -> bool:
diff --git a/libs/core/langchain_core/utils/function_calling.py b/libs/core/langchain_core/utils/function_calling.py
index c2b59fc254c..13caf23e378 100644
--- a/libs/core/langchain_core/utils/function_calling.py
+++ b/libs/core/langchain_core/utils/function_calling.py
@@ -337,40 +337,53 @@ def _format_tool_to_openai_function(tool: BaseTool) -> FunctionDescription:
     Returns:
         The function description.
     """
+    # The result is cached on the tool instance under `_openai_function_dict`.
+    # `BaseTool.__setattr__` pops this key when `args_schema` / `description` /
+    # `name` are mutated (alongside the existing `tool_call_schema` and `args`
+    # caches), so the invalidation path is already wired up.
+    cached = tool.__dict__.get("_openai_function_dict")
+    if cached is not None:
+        return cached
+
     is_simple_oai_tool = (
         isinstance(tool, langchain_core.tools.simple.Tool) and not tool.args_schema
     )
     schema = tool.tool_call_schema
     if schema and not is_simple_oai_tool:
         if isinstance(schema, dict):
-            return _convert_json_schema_to_openai_function(
+            result = _convert_json_schema_to_openai_function(
                 schema, name=tool.name, description=tool.description
             )
-        if issubclass(schema, (BaseModel, BaseModelV1)):
-            return _convert_pydantic_to_openai_function(
+        elif issubclass(schema, (BaseModel, BaseModelV1)):
+            result = _convert_pydantic_to_openai_function(
                 schema, name=tool.name, description=tool.description
             )
-        error_msg = (
-            f"Unsupported tool call schema: {schema}. "
-            "Tool call schema must be a JSON schema dict or a Pydantic model."
-        )
-        raise ValueError(error_msg)
-    return {
-        "name": tool.name,
-        "description": tool.description,
-        "parameters": {
-            # This is a hack to get around the fact that some tools
-            # do not expose an args_schema, and expect an argument
-            # which is a string.
-            # And Open AI does not support an array type for the
-            # parameters.
-            "properties": {
-                "__arg1": {"title": "__arg1", "type": "string"},
+        else:
+            error_msg = (
+                f"Unsupported tool call schema: {schema}. "
+                "Tool call schema must be a JSON schema dict or a Pydantic model."
+            )
+            raise ValueError(error_msg)
+    else:
+        result = {
+            "name": tool.name,
+            "description": tool.description,
+            "parameters": {
+                # This is a hack to get around the fact that some tools
+                # do not expose an args_schema, and expect an argument
+                # which is a string.
+                # And Open AI does not support an array type for the
+                # parameters.
+                "properties": {
+                    "__arg1": {"title": "__arg1", "type": "string"},
+                },
+                "required": ["__arg1"],
+                "type": "object",
             },
-            "required": ["__arg1"],
-            "type": "object",
-        },
-    }
+        }
+
+    tool.__dict__["_openai_function_dict"] = result
+    return result
 
 
 def convert_to_openai_function(