genai: fix pydantic structured_output with array

libs/genai/langchain_google_genai/_function_utils.py

@@ -269,20 +269,97 @@ def _convert_pydantic_to_genai_function(
         name=tool_name if tool_name else schema.get("title"),
         description=tool_description if tool_description else schema.get("description"),
         parameters={
-            "properties": {
-                k: {
-                    "type_": _get_type_from_schema(v),
-                    "description": v.get("description"),
-                }
-                for k, v in schema["properties"].items()
-            },
+            "properties": _get_properties_from_schema_any(
+                schema.get("properties")
+            ),  # TODO: use _dict_to_gapic_schema() if possible
+            # "items": _get_items_from_schema_any(
+            #     schema
+            # ),  # TODO: fix it https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/function-calling?hl#schema
             "required": schema.get("required", []),
             "type_": TYPE_ENUM[schema["type"]],
         },
     )
     return function_declaration
 
 
+def _get_properties_from_schema_any(schema: Any) -> Dict[str, Any]:
+    if isinstance(schema, Dict):
+        return _get_properties_from_schema(schema)
+    return {}
+
+
+def _get_properties_from_schema(schema: Dict) -> Dict[str, Any]:
+    properties = {}
+    for k, v in schema.items():
+        if not isinstance(k, str):
+            logger.warning(f"Key '{k}' is not supported in schema, type={type(k)}")
+            continue
+        if not isinstance(v, Dict):
+            logger.warning(f"Value '{v}' is not supported in schema, ignoring v={v}")
+            continue
+        properties_item: Dict[str, Union[str, int, Dict, List]] = {}
+        if v.get("type"):
+            properties_item["type_"] = _get_type_from_schema(v)
+
+        if v.get("enum"):
+            properties_item["enum"] = v["enum"]
+
+        description = v.get("description")
+        if description and isinstance(description, str):
+            properties_item["description"] = description
+
+        if v.get("type") == "array" and v.get("items"):
+            properties_item["items"] = _get_items_from_schema_any(v.get("items"))
+
+        if v.get("type") == "object" and v.get("properties"):
+            properties_item["properties"] = _get_properties_from_schema_any(
+                v.get("properties")
+            )
+        if k == "title" and "description" not in properties_item:
+            properties_item["description"] = k + " is " + str(v)
+
+        properties[k] = properties_item
+
+    return properties
+
+
+def _get_items_from_schema_any(schema: Any) -> Dict[str, Any]:
+    if isinstance(schema, Dict):
+        return _get_items_from_schema(schema)
+    if isinstance(schema, List):
+        return _get_items_from_schema(schema)
+    if isinstance(schema, str):
+        return _get_items_from_schema(schema)
+    return {}
+
+
+def _get_items_from_schema(schema: Union[Dict, List, str]) -> Dict[str, Any]:
+    items: Dict = {}
+    if isinstance(schema, List):
+        for i, v in enumerate(schema):
+            items[f"item{i}"] = _get_properties_from_schema_any(v)
+    elif isinstance(schema, Dict):
+        item: Dict = {}
+        for k, v in schema.items():
+            item["type_"] = _get_type_from_schema(v)
+            if not isinstance(v, Dict):
+                logger.warning(
+                    f"Value '{v}' is not supported in schema, ignoring v={v}"
+                )
+                continue
+            if v.get("type") == "object" and v.get("properties"):
+                item["properties"] = _get_properties_from_schema_any(
+                    v.get("properties")
+                )
+            if k == "title" and "description" not in item:
+                item["description"] = v
+        items = item
+    else:
+        # str
+        items["type_"] = TYPE_ENUM.get(str(schema), glm.Type.STRING)
+    return items
+
+
 def _get_type_from_schema(schema: Dict[str, Any]) -> int:
     if "anyOf" in schema:
         types = [_get_type_from_schema(sub_schema) for sub_schema in schema["anyOf"]]
@@ -293,10 +370,7 @@ def _get_type_from_schema(schema: Dict[str, Any]) -> int:
             pass
     elif "type" in schema:
         stype = str(schema["type"])
-        if stype in TYPE_ENUM:
-            return TYPE_ENUM[stype]
-        else:
-            pass
+        return TYPE_ENUM.get(stype, glm.Type.STRING)
     else:
         pass
     return TYPE_ENUM["string"]  # Default to string if no valid types found

libs/genai/tests/integration_tests/test_chat_models.py

@@ -335,43 +335,56 @@ def _check_tool_calls(response: BaseMessage, expected_name: str) -> None:
     assert isinstance(response, AIMessage)
     assert isinstance(response.content, str)
     assert response.content == ""
+
+    # function_call
     function_call = response.additional_kwargs.get("function_call")
     assert function_call
     assert function_call["name"] == expected_name
     arguments_str = function_call.get("arguments")
     assert arguments_str
     arguments = json.loads(arguments_str)
-    assert arguments == {
-        "name": "Erick",
-        "age": 27.0,
-    }
+    _check_tool_call_args(arguments)
+
+    # tool_calls
     tool_calls = response.tool_calls
     assert len(tool_calls) == 1
     tool_call = tool_calls[0]
     assert tool_call["name"] == expected_name
-    assert tool_call["args"] == {"age": 27.0, "name": "Erick"}
+    _check_tool_call_args(tool_call["args"])
+
+
+def _check_tool_call_args(tool_call_args: dict) -> None:
+    assert tool_call_args == {
+        "age": 27.0,
+        "name": "Erick",
+        "likes": ["apple", "banana"],
+    }
 
 
 @pytest.mark.extended
 def test_chat_vertexai_gemini_function_calling() -> None:
     class MyModel(BaseModel):
         name: str
         age: int
+        likes: list[str]
 
     safety: Dict[HarmCategory, HarmBlockThreshold] = {
         HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_ONLY_HIGH  # type: ignore[dict-item]
     }
     # Test .bind_tools with BaseModel
-    message = HumanMessage(content="My name is Erick and I am 27 years old")
+    message = HumanMessage(
+        content="My name is Erick and I am 27 years old. I like apple and banana."
+    )
     model = ChatGoogleGenerativeAI(model=_MODEL, safety_settings=safety).bind_tools(
         [MyModel]
     )
     response = model.invoke([message])
+    print("response=", response)
     _check_tool_calls(response, "MyModel")
 
     # Test .bind_tools with function
-    def my_model(name: str, age: int) -> None:
-        """Invoke this with names and ages."""
+    def my_model(name: str, age: int, likes: list[str]) -> None:
+        """Invoke this with names and age and likes."""
         pass
 
     model = ChatGoogleGenerativeAI(model=_MODEL, safety_settings=safety).bind_tools(
@@ -382,8 +395,8 @@ def my_model(name: str, age: int) -> None:
 
     # Test .bind_tools with tool
     @tool
-    def my_tool(name: str, age: int) -> None:
-        """Invoke this with names and ages."""
+    def my_tool(name: str, age: int, likes: list[str]) -> None:
+        """Invoke this with names and age and likes."""
         pass
 
     model = ChatGoogleGenerativeAI(model=_MODEL, safety_settings=safety).bind_tools(
@@ -405,7 +418,9 @@ def my_tool(name: str, age: int) -> None:
     assert len(gathered.tool_call_chunks) == 1
     tool_call_chunk = gathered.tool_call_chunks[0]
     assert tool_call_chunk["name"] == "my_tool"
-    assert tool_call_chunk["args"] == '{"age": 27.0, "name": "Erick"}'
+    arguments_str = tool_call_chunk["args"]
+    arguments = json.loads(str(arguments_str))
+    _check_tool_call_args(arguments)
 
 
 # Test with model that supports tool choice (gemini 1.5) and one that doesn't

libs/genai/tests/unit_tests/test_function_utils.py

@@ -136,6 +136,7 @@ def test_tool_to_dict_pydantic() -> None:
     class MyModel(BaseModel):
         name: str
         age: int
+        likes: list[str]
 
     tool = convert_to_genai_function_declarations([MyModel])
     tool_dict = tool_to_dict(tool)