Merge pull request #4522 from andrewmjc/matchingToolMessageSpec

[Bug Fix] Use OpenAI Tool Response Spec When Converting To Gemini/VertexAI Tool Response

litellm/llms/prompt_templates/factory.py

@@ -1026,16 +1026,17 @@ def convert_to_gemini_tool_call_invoke(
 
 def convert_to_gemini_tool_call_result(
     message: dict,
+    last_message_with_tool_calls: dict|None,
 ) -> litellm.types.llms.vertex_ai.PartType:
     """
     OpenAI message with a tool result looks like:
     {
         "tool_call_id": "tool_1",
-        "role": "tool",
-        "name": "get_current_weather",
+        "role": "tool", 
         "content": "function result goes here",
     },
 
+    # NOTE: Function messages have been deprecated
     OpenAI message with a function call result looks like:
     {
         "role": "function",
@@ -1044,7 +1045,19 @@ def convert_to_gemini_tool_call_result(
     }
     """
     content = message.get("content", "")
-    name = message.get("name", "")
+    name = ""
+
+    # Recover name from last message with tool calls
+    if last_message_with_tool_calls:
+        tools = last_message_with_tool_calls.get("tool_calls", [])
+        msg_tool_call_id = message.get("tool_call_id", None)
+        for tool in tools:
+            prev_tool_call_id = tool.get("id", None)
+            if msg_tool_call_id and prev_tool_call_id and msg_tool_call_id == prev_tool_call_id:
+                name = tool.get("function", {}).get("name", "")
+
+    if not name:
+        raise Exception("Missing corresponding tool call for tool response message")
 
     # We can't determine from openai message format whether it's a successful or
     # error call result so default to the successful result template

litellm/llms/vertex_ai.py

@@ -328,6 +328,8 @@ def _gemini_convert_messages_with_history(messages: list) -> List[ContentType]:
     user_message_types = {"user", "system"}
     contents: List[ContentType] = []
 
+    last_message_with_tool_calls = None
+
     msg_i = 0
     try:
         while msg_i < len(messages):
@@ -383,6 +385,7 @@ def _gemini_convert_messages_with_history(messages: list) -> List[ContentType]:
                             messages[msg_i]["tool_calls"]
                         )
                     )
+                    last_message_with_tool_calls = messages[msg_i]
                 else:
                     assistant_text = (
                         messages[msg_i].get("content") or ""
@@ -397,7 +400,7 @@ def _gemini_convert_messages_with_history(messages: list) -> List[ContentType]:
 
             ## APPEND TOOL CALL MESSAGES ##
             if msg_i < len(messages) and messages[msg_i]["role"] == "tool":
-                _part = convert_to_gemini_tool_call_result(messages[msg_i])
+                _part = convert_to_gemini_tool_call_result(messages[msg_i], last_message_with_tool_calls)
                 contents.append(ContentType(parts=[_part]))  # type: ignore
                 msg_i += 1
             if msg_i == init_msg_i:  # prevent infinite loops

litellm/tests/test_amazing_vertex_completion.py

@@ -1159,8 +1159,7 @@ async def test_gemini_pro_function_calling(provider, sync_mode):
             # The result of the tool call is added to the history
             {
                 "role": "tool",
-                "tool_call_id": "call_123",
-                "name": "get_weather",
+                "tool_call_id": "call_123", 
                 "content": "27 degrees celsius and clear in San Francisco, CA",
             },
             # Now the assistant can reply with the result of the tool call.
@@ -1382,6 +1381,52 @@ async def test_vertexai_aembedding():
     except Exception as e:
         pytest.fail(f"Error occurred: {e}")
 
+@pytest.mark.asyncio
+def test_tool_name_conversion():
+    messages = [
+        {
+            "role": "system",
+            "content": "Your name is Litellm Bot, you are a helpful assistant",
+        },
+        # User asks for their name and weather in San Francisco
+        {
+            "role": "user",
+            "content": "Hello, what is your name and can you tell me the weather?",
+        },
+        # Assistant replies with a tool call
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "id": "call_123",
+                    "type": "function",
+                    "index": 0,
+                    "function": {
+                        "name": "get_weather",
+                        "arguments": '{"location":"San Francisco, CA"}',
+                    },
+                }
+            ],
+        },
+        # The result of the tool call is added to the history
+        {
+            "role": "tool",
+            "tool_call_id": "call_123",
+            "content": "27 degrees celsius and clear in San Francisco, CA",
+        },
+        # Now the assistant can reply with the result of the tool call.
+    ]
+
+    translated_messages = _gemini_convert_messages_with_history(messages=messages)
+
+    print(f"\n\ntranslated_messages: {translated_messages}\ntranslated_messages")
+
+    # assert that the last tool response has the corresponding tool name
+    assert (
+        translated_messages[-1]["parts"][0]["function_response"]["name"] == "get_weather"
+    )
+
 
 # Extra gemini Vision tests for completion + stream, async, async + stream
 # if we run into issues with gemini, we will also add these to our ci/cd pipeline
@@ -1531,7 +1576,6 @@ def test_prompt_factory():
         {
             "role": "tool",
             "tool_call_id": "call_123",
-            "name": "get_weather",
             "content": "27 degrees celsius and clear in San Francisco, CA",
         },
         # Now the assistant can reply with the result of the tool call.
@@ -1541,7 +1585,6 @@ def test_prompt_factory():
 
     print(f"\n\ntranslated_messages: {translated_messages}\ntranslated_messages")
 
-
 def test_prompt_factory_nested():
     messages = [
         {"role": "user", "content": [{"type": "text", "text": "hi"}]},
@@ -1563,4 +1606,4 @@ def test_prompt_factory_nested():
         assert "text" in message["parts"][0], "Missing 'text' from 'parts'"
         assert isinstance(
             message["parts"][0]["text"], str
-        ), "'text' value not a string."
+        ), "'text' value not a string."