test test_create_json_tool_call_for_response_format

litellm/llms/anthropic/chat/transformation.py

@@ -7,6 +7,7 @@ from litellm.types.llms.anthropic import (
     AllAnthropicToolsValues,
     AnthropicComputerTool,
     AnthropicHostedTools,
+    AnthropicInputSchema,
     AnthropicMessageRequestBase,
     AnthropicMessagesRequest,
     AnthropicMessagesTool,
@@ -159,15 +160,17 @@ class AnthropicConfig:
         returned_tool: Optional[AllAnthropicToolsValues] = None
 
         if tool["type"] == "function" or tool["type"] == "custom":
+            _input_schema: dict = tool["function"].get(
+                "parameters",
+                {
+                    "type": "object",
+                    "properties": {},
+                },
+            )
+            input_schema: AnthropicInputSchema = AnthropicInputSchema(**_input_schema)
             _tool = AnthropicMessagesTool(
                 name=tool["function"]["name"],
-                input_schema=tool["function"].get(
-                    "parameters",
-                    {
-                        "type": "object",
-                        "properties": {},
-                    },
-                ),
+                input_schema=input_schema,
             )
 
             _description = tool["function"].get("description")
@@ -304,17 +307,10 @@ class AnthropicConfig:
                 - You should set tool_choice (see Forcing tool use) to instruct the model to explicitly use that tool
                 - Remember that the model will pass the input to the tool, so the name of the tool and description should be from the model’s perspective.
                 """
-                _tool_choice = None
                 _tool_choice = {"name": "json_tool_call", "type": "tool"}
-
-                _tool = AnthropicMessagesTool(
-                    name="json_tool_call",
-                    input_schema={
-                        "type": "object",
-                        "properties": {"values": json_schema},  # type: ignore
-                    },
+                _tool = self._create_json_tool_call_for_response_format(
+                    json_schema=json_schema,
                 )
-
                 optional_params["tools"] = [_tool]
                 optional_params["tool_choice"] = _tool_choice
                 optional_params["json_mode"] = True
@@ -341,6 +337,34 @@ class AnthropicConfig:
 
         return optional_params
 
+    def _create_json_tool_call_for_response_format(
+        self,
+        json_schema: Optional[dict] = None,
+    ) -> AnthropicMessagesTool:
+        """
+        Handles creating a tool call for getting responses in JSON format.
+
+        Args:
+            json_schema (Optional[dict]): The JSON schema the response should be in
+
+        Returns:
+            AnthropicMessagesTool: The tool call to send to Anthropic API to get responses in JSON format
+        """
+        _input_schema: AnthropicInputSchema = AnthropicInputSchema(
+            type="object",
+        )
+
+        if json_schema is None:
+            # Anthropic raises a 400 BadRequest error if properties is passed as None
+            # see usage with additionalProperties (Example 5) https://github.com/anthropics/anthropic-cookbook/blob/main/tool_use/extracting_structured_json.ipynb
+            _input_schema["additionalProperties"] = True
+            _input_schema["properties"] = {}
+        else:
+            _input_schema["properties"] = json_schema
+
+        _tool = AnthropicMessagesTool(name="json_tool_call", input_schema=_input_schema)
+        return _tool
+
     def is_cache_control_set(self, messages: List[AllMessageValues]) -> bool:
         """
         Return if {"cache_control": ..} in message content block

litellm/types/llms/anthropic.py

@@ -12,10 +12,16 @@ class AnthropicMessagesToolChoice(TypedDict, total=False):
     disable_parallel_tool_use: bool  # default is false
 
 
+class AnthropicInputSchema(TypedDict, total=False):
+    type: Optional[str]
+    properties: Optional[dict]
+    additionalProperties: Optional[bool]
+
+
 class AnthropicMessagesTool(TypedDict, total=False):
     name: Required[str]
     description: str
-    input_schema: Required[dict]
+    input_schema: Optional[AnthropicInputSchema]
     type: Literal["custom"]
     cache_control: Optional[Union[dict, ChatCompletionCachedContent]]
 

tests/llm_translation/base_llm_unit_tests.py

@@ -53,6 +53,32 @@ class BaseLLMChatTest(ABC):
         response = litellm.completion(**base_completion_call_args, messages=messages)
         assert response is not None
 
+    def test_json_response_format(self):
+        """
+        Test that the JSON response format is supported by the LLM API
+        """
+        base_completion_call_args = self.get_base_completion_call_args()
+        litellm.set_verbose = True
+
+        messages = [
+            {
+                "role": "system",
+                "content": "Your output should be a JSON object with no additional properties.  ",
+            },
+            {
+                "role": "user",
+                "content": "Respond with this in json. city=San Francisco, state=CA, weather=sunny, temp=60",
+            },
+        ]
+
+        response = litellm.completion(
+            **base_completion_call_args,
+            messages=messages,
+            response_format={"type": "json_object"},
+        )
+
+        print(response)
+
     @pytest.fixture
     def pdf_messages(self):
         import base64

tests/llm_translation/test_anthropic_completion.py

@@ -627,6 +627,38 @@ def test_anthropic_tool_helper(cache_control_location):
     assert tool["cache_control"] == {"type": "ephemeral"}
 
 
+def test_create_json_tool_call_for_response_format():
+    """
+    tests using response_format=json with anthropic
+
+    A tool call to anthropic is made when response_format=json is used.
+
+    """
+    # Initialize AnthropicConfig
+    config = AnthropicConfig()
+
+    # Test case 1: No schema provided
+    # See Anthropics Example 5 on how to handle cases when no schema is provided https://github.com/anthropics/anthropic-cookbook/blob/main/tool_use/extracting_structured_json.ipynb
+    tool = config._create_json_tool_call_for_response_format()
+    assert tool["name"] == "json_tool_call"
+    _input_schema = tool.get("input_schema")
+    assert _input_schema is not None
+    assert _input_schema.get("type") == "object"
+    assert _input_schema.get("additionalProperties") is True
+    assert _input_schema.get("properties") == {}
+
+    # Test case 2: With custom schema
+    # reference: https://github.com/anthropics/anthropic-cookbook/blob/main/tool_use/extracting_structured_json.ipynb
+    custom_schema = {"name": {"type": "string"}, "age": {"type": "integer"}}
+    tool = config._create_json_tool_call_for_response_format(json_schema=custom_schema)
+    assert tool["name"] == "json_tool_call"
+    _input_schema = tool.get("input_schema")
+    assert _input_schema is not None
+    assert _input_schema.get("type") == "object"
+    assert _input_schema.get("properties") == custom_schema
+    assert "additionalProperties" not in _input_schema
+
+
 from litellm import completion