Merge pull request #9437 from BerriAI/litellm_dev_03_21_2025_p1

Consistent anthropic response_format streaming/non-streaming behaviour

litellm/llms/anthropic/chat/transformation.py

@@ -387,7 +387,7 @@ class AnthropicConfig(BaseConfig):
             _input_schema["additionalProperties"] = True
             _input_schema["properties"] = {}
         else:
-            _input_schema["properties"] = {"values": json_schema}
+            _input_schema.update(cast(AnthropicInputSchema, json_schema))
 
         _tool = AnthropicMessagesTool(
             name=RESPONSE_FORMAT_TOOL_NAME, input_schema=_input_schema

tests/litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py

@@ -0,0 +1,35 @@
+import json
+import os
+import sys
+
+import pytest
+from fastapi.testclient import TestClient
+
+sys.path.insert(
+    0, os.path.abspath("../../../../..")
+)  # Adds the parent directory to the system path
+from unittest.mock import MagicMock, patch
+
+from litellm.llms.anthropic.chat.transformation import AnthropicConfig
+
+
+def test_response_format_transformation_unit_test():
+    config = AnthropicConfig()
+
+    response_format_json_schema = {
+        "description": 'Progress report for the thinking process\n\nThis model represents a snapshot of the agent\'s current progress during\nthe thinking process, providing a brief description of the current activity.\n\nAttributes:\n    agent_doing: Brief description of what the agent is currently doing.\n                Should be kept under 10 words. Example: "Learning about home automation"',
+        "properties": {"agent_doing": {"title": "Agent Doing", "type": "string"}},
+        "required": ["agent_doing"],
+        "title": "ThinkingStep",
+        "type": "object",
+        "additionalProperties": False,
+    }
+
+    result = config._create_json_tool_call_for_response_format(
+        json_schema=response_format_json_schema
+    )
+
+    assert result["input_schema"]["properties"] == {
+        "agent_doing": {"title": "Agent Doing", "type": "string"}
+    }
+    print(result)

tests/llm_translation/base_llm_unit_tests.py

@@ -20,6 +20,7 @@ from litellm.utils import (
     get_optional_params,
     ProviderConfigManager,
 )
+from litellm.main import stream_chunk_builder
 from typing import Union
 
 # test_example.py
@@ -338,7 +339,7 @@ class BaseLLMChatTest(ABC):
 
     @pytest.mark.flaky(retries=6, delay=1)
     def test_json_response_pydantic_obj(self):
-        litellm.set_verbose = True
+        litellm._turn_on_debug()
         from pydantic import BaseModel
         from litellm.utils import supports_response_schema
 
@@ -995,3 +996,73 @@ class BaseOSeriesModelsTest(ABC):  # test across azure/openai
             ), "temperature should not be in the request body"
         except Exception as e:
             pytest.fail(f"Error occurred: {e}")
+
+
+class BaseAnthropicChatTest(ABC):
+    """
+    Ensures consistent result across anthropic model usage
+    """
+
+    @abstractmethod
+    def get_base_completion_call_args(self) -> dict:
+        """Must return the base completion call args"""
+        pass
+
+    @property
+    def completion_function(self):
+        return litellm.completion
+
+    def test_anthropic_response_format_streaming_vs_non_streaming(self):
+        litellm.set_verbose = True
+        args = {
+            "messages": [
+                {
+                    "content": "Your goal is to summarize the previous agent's thinking process into short descriptions to let user better understand the research progress. If no information is available, just say generic phrase like 'Doing some research...' with the given output format. Make sure to adhere to the output format no matter what, even if you don't have any information or you are not allowed to respond to the given input information (then just say generic phrase like 'Doing some research...').",
+                    "role": "system",
+                },
+                {
+                    "role": "user",
+                    "content": "Here is the input data (previous agent's output): \n\n Let's try to refine our search further, focusing more on the technical aspects of home automation and home energy system management:",
+                },
+            ],
+            "response_format": {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": "final_output",
+                    "strict": True,
+                    "schema": {
+                        "description": 'Progress report for the thinking process\n\nThis model represents a snapshot of the agent\'s current progress during\nthe thinking process, providing a brief description of the current activity.\n\nAttributes:\n    agent_doing: Brief description of what the agent is currently doing.\n                Should be kept under 10 words. Example: "Learning about home automation"',
+                        "properties": {
+                            "agent_doing": {"title": "Agent Doing", "type": "string"}
+                        },
+                        "required": ["agent_doing"],
+                        "title": "ThinkingStep",
+                        "type": "object",
+                        "additionalProperties": False,
+                    },
+                },
+            },
+        }
+
+        base_completion_call_args = self.get_base_completion_call_args()
+
+        response = self.completion_function(
+            **base_completion_call_args, **args, stream=True
+        )
+
+        chunks = []
+        for chunk in response:
+            print(f"chunk: {chunk}")
+            chunks.append(chunk)
+
+        print(f"chunks: {chunks}")
+        built_response = stream_chunk_builder(chunks=chunks)
+
+        non_stream_response = self.completion_function(
+            **base_completion_call_args, **args, stream=False
+        )
+
+        assert (
+            json.loads(built_response.choices[0].message.content).keys()
+            == json.loads(non_stream_response.choices[0].message.content).keys()
+        ), f"Got={json.loads(built_response.choices[0].message.content)}, Expected={json.loads(non_stream_response.choices[0].message.content)}"

tests/llm_translation/test_anthropic_completion.py

@@ -36,7 +36,7 @@ from litellm.types.llms.openai import ChatCompletionToolCallFunctionChunk
 from litellm.llms.anthropic.common_utils import process_anthropic_headers
 from litellm.llms.anthropic.chat.handler import AnthropicChatCompletion
 from httpx import Headers
-from base_llm_unit_tests import BaseLLMChatTest
+from base_llm_unit_tests import BaseLLMChatTest, BaseAnthropicChatTest
 
 
 def streaming_format_tests(chunk: dict, idx: int):
@@ -455,14 +455,15 @@ def test_create_json_tool_call_for_response_format():
     _input_schema = tool.get("input_schema")
     assert _input_schema is not None
     assert _input_schema.get("type") == "object"
-    assert _input_schema.get("properties") == {"values": custom_schema}
+    assert _input_schema.get("name") == custom_schema["name"]
+    assert _input_schema.get("age") == custom_schema["age"]
     assert "additionalProperties" not in _input_schema
 
 
 from litellm import completion
 
 
-class TestAnthropicCompletion(BaseLLMChatTest):
+class TestAnthropicCompletion(BaseLLMChatTest, BaseAnthropicChatTest):
     def get_base_completion_call_args(self) -> dict:
         return {"model": "anthropic/claude-3-5-sonnet-20240620"}