langchain_google_vertexai[patch]: Add support for SystemMessage for Gemini chat model (#15933)

- **Description:** In Google Vertex AI, Gemini Chat models currently
doesn't have a support for SystemMessage. This PR adds support for it
only if a user provides additional convert_system_message_to_human flag
during model initialization (in this case, SystemMessage would be
prepended to the first HumanMessage). **NOTE:** The implementation is
similar to #14824


- **Twitter handle:** rajesh_thallam

---------

Co-authored-by: Erick Friis <erick@langchain.dev>

docs/docs/integrations/chat/google_vertex_ai_palm.ipynb

@@ -11,7 +11,6 @@
    ]
   },
   {
-   "attachments": {},
    "cell_type": "markdown",
    "metadata": {},
    "source": [
@@ -95,7 +94,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "If we want to construct a simple chain that takes user specified parameters:"
+    "Gemini doesn't support SystemMessage at the moment, but it can be added to the first human message in the row. If you want such behavior, just set the `convert_system_message_to_human` to `True`:"
    ]
   },
   {
@@ -106,7 +105,7 @@
     {
      "data": {
       "text/plain": [
-       "AIMessage(content=' プログラミングが大好きです')"
+       "AIMessage(content=\"J'aime la programmation.\")"
       ]
      },
      "execution_count": 9,
@@ -114,6 +113,40 @@
      "output_type": "execute_result"
     }
    ],
+   "source": [
+    "system = \"You are a helpful assistant who translate English to French\"\n",
+    "human = \"Translate this sentence from English to French. I love programming.\"\n",
+    "prompt = ChatPromptTemplate.from_messages([(\"system\", system), (\"human\", human)])\n",
+    "\n",
+    "chat = ChatVertexAI(model_name=\"gemini-pro\", convert_system_message_to_human=True)\n",
+    "\n",
+    "chain = prompt | chat\n",
+    "chain.invoke({})"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "If we want to construct a simple chain that takes user specified parameters:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "AIMessage(content=' プログラミングが大好きです')"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
    "source": [
     "system = (\n",
     "    \"You are a helpful assistant that translates {input_language} to {output_language}.\"\n",
@@ -121,6 +154,8 @@
     "human = \"{text}\"\n",
     "prompt = ChatPromptTemplate.from_messages([(\"system\", system), (\"human\", human)])\n",
     "\n",
+    "chat = ChatVertexAI()\n",
+    "\n",
     "chain = prompt | chat\n",
     "\n",
     "chain.invoke(\n",
@@ -133,7 +168,6 @@
    ]
   },
   {
-   "attachments": {},
    "cell_type": "markdown",
    "metadata": {
     "execution": {
@@ -352,7 +386,7 @@
        "AIMessage(content=' Why do you love programming?')"
       ]
      },
-     "execution_count": 6,
+     "execution_count": 7,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -428,8 +462,14 @@
   }
  ],
  "metadata": {
+  "environment": {
+   "kernel": "python3",
+   "name": "common-cpu.m108",
+   "type": "gcloud",
+   "uri": "gcr.io/deeplearning-platform-release/base-cpu:m108"
+  },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
@@ -443,7 +483,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.11.4"
+   "version": "3.10.10"
   },
   "vscode": {
    "interpreter": {

libs/partners/google-vertexai/langchain_google_vertexai/chat_models.py

@@ -111,7 +111,9 @@ def _is_url(s: str) -> bool:
 
 
 def _parse_chat_history_gemini(
-    history: List[BaseMessage], project: Optional[str]
+    history: List[BaseMessage],
+    project: Optional[str] = None,
+    convert_system_message_to_human: Optional[bool] = False,
 ) -> List[Content]:
     def _convert_to_prompt(part: Union[str, Dict]) -> Part:
         if isinstance(part, str):
@@ -155,9 +157,25 @@ def _parse_chat_history_gemini(
         return [_convert_to_prompt(part) for part in raw_content]
 
     vertex_messages = []
+    raw_system_message = None
     for i, message in enumerate(history):
-        if i == 0 and isinstance(message, SystemMessage):
-            raise ValueError("SystemMessages are not yet supported!")
+        if (
+            i == 0
+            and isinstance(message, SystemMessage)
+            and not convert_system_message_to_human
+        ):
+            raise ValueError(
+                """SystemMessages are not yet supported!
+                
+To automatically convert the leading SystemMessage to a HumanMessage,
+set  `convert_system_message_to_human` to True. Example:
+
+llm = ChatVertexAI(model_name="gemini-pro", convert_system_message_to_human=True)
+"""
+            )
+        elif i == 0 and isinstance(message, SystemMessage):
+            raw_system_message = message
+            continue
         elif isinstance(message, AIMessage):
             raw_function_call = message.additional_kwargs.get("function_call")
             role = "model"
@@ -170,6 +188,8 @@ def _parse_chat_history_gemini(
                 )
                 gapic_part = GapicPart(function_call=function_call)
                 parts = [Part._from_gapic(gapic_part)]
+            else:
+                parts = _convert_to_parts(message)
         elif isinstance(message, HumanMessage):
             role = "user"
             parts = _convert_to_parts(message)
@@ -188,6 +208,15 @@ def _parse_chat_history_gemini(
                 f"Unexpected message with type {type(message)} at the position {i}."
             )
 
+        if raw_system_message:
+            if role == "model":
+                raise ValueError(
+                    "SystemMessage should be followed by a HumanMessage and "
+                    "not by AIMessage."
+                )
+            parts = _convert_to_parts(raw_system_message) + parts
+            raw_system_message = None
+
         vertex_message = Content(role=role, parts=parts)
         vertex_messages.append(vertex_message)
     return vertex_messages
@@ -258,6 +287,11 @@ class ChatVertexAI(_VertexAICommon, BaseChatModel):
     model_name: str = "chat-bison"
     "Underlying model name."
     examples: Optional[List[BaseMessage]] = None
+    convert_system_message_to_human: bool = False
+    """Whether to merge any leading SystemMessage into the following HumanMessage.
+    
+    Gemini does not support system messages; any unsupported messages will 
+    raise an error."""
 
     @classmethod
     def is_lc_serializable(self) -> bool:
@@ -327,7 +361,11 @@ class ChatVertexAI(_VertexAICommon, BaseChatModel):
             msg_params["candidate_count"] = params.pop("candidate_count")
 
         if self._is_gemini_model:
-            history_gemini = _parse_chat_history_gemini(messages, project=self.project)
+            history_gemini = _parse_chat_history_gemini(
+                messages,
+                project=self.project,
+                convert_system_message_to_human=self.convert_system_message_to_human,
+            )
             message = history_gemini.pop()
             chat = self.client.start_chat(history=history_gemini)
 
@@ -396,7 +434,11 @@ class ChatVertexAI(_VertexAICommon, BaseChatModel):
             msg_params["candidate_count"] = params.pop("candidate_count")
 
         if self._is_gemini_model:
-            history_gemini = _parse_chat_history_gemini(messages, project=self.project)
+            history_gemini = _parse_chat_history_gemini(
+                messages,
+                project=self.project,
+                convert_system_message_to_human=self.convert_system_message_to_human,
+            )
             message = history_gemini.pop()
             chat = self.client.start_chat(history=history_gemini)
             # set param to `functions` until core tool/function calling implemented
@@ -441,7 +483,11 @@ class ChatVertexAI(_VertexAICommon, BaseChatModel):
     ) -> Iterator[ChatGenerationChunk]:
         params = self._prepare_params(stop=stop, stream=True, **kwargs)
         if self._is_gemini_model:
-            history_gemini = _parse_chat_history_gemini(messages, project=self.project)
+            history_gemini = _parse_chat_history_gemini(
+                messages,
+                project=self.project,
+                convert_system_message_to_human=self.convert_system_message_to_human,
+            )
             message = history_gemini.pop()
             chat = self.client.start_chat(history=history_gemini)
             # set param to `functions` until core tool/function calling implemented

libs/partners/google-vertexai/tests/integration_tests/test_chat_models.py

@@ -182,3 +182,36 @@ def test_vertexai_single_call_fails_no_message() -> None:
         str(exc_info.value)
         == "You should provide at least one message to start the chat!"
     )
+
+
+@pytest.mark.parametrize("model_name", ["gemini-pro"])
+def test_chat_vertexai_gemini_system_message_error(model_name: str) -> None:
+    model = ChatVertexAI(model_name=model_name)
+    text_question1, text_answer1 = "How much is 2+2?", "4"
+    text_question2 = "How much is 3+3?"
+    system_message = SystemMessage(content="You're supposed to answer math questions.")
+    message1 = HumanMessage(content=text_question1)
+    message2 = AIMessage(content=text_answer1)
+    message3 = HumanMessage(content=text_question2)
+    with pytest.raises(ValueError):
+        model([system_message, message1, message2, message3])
+
+
+@pytest.mark.parametrize("model_name", model_names_to_test)
+def test_chat_vertexai_system_message(model_name: str) -> None:
+    if model_name:
+        model = ChatVertexAI(
+            model_name=model_name, convert_system_message_to_human=True
+        )
+    else:
+        model = ChatVertexAI()
+
+    text_question1, text_answer1 = "How much is 2+2?", "4"
+    text_question2 = "How much is 3+3?"
+    system_message = SystemMessage(content="You're supposed to answer math questions.")
+    message1 = HumanMessage(content=text_question1)
+    message2 = AIMessage(content=text_answer1)
+    message3 = HumanMessage(content=text_question2)
+    response = model([system_message, message1, message2, message3])
+    assert isinstance(response, AIMessage)
+    assert isinstance(response.content, str)

libs/partners/google-vertexai/tests/unit_tests/test_chat_models.py

@@ -13,6 +13,7 @@ from vertexai.language_models import ChatMessage, InputOutputTextPair  # type: i
 from langchain_google_vertexai.chat_models import (
     ChatVertexAI,
     _parse_chat_history,
+    _parse_chat_history_gemini,
     _parse_examples,
 )
 
@@ -112,6 +113,24 @@ def test_parse_chat_history_correct() -> None:
     ]
 
 
+def test_parse_history_gemini() -> None:
+    system_input = "You're supposed to answer math questions."
+    text_question1, text_answer1 = "How much is 2+2?", "4"
+    text_question2 = "How much is 3+3?"
+    system_message = SystemMessage(content=system_input)
+    message1 = HumanMessage(content=text_question1)
+    message2 = AIMessage(content=text_answer1)
+    message3 = HumanMessage(content=text_question2)
+    messages = [system_message, message1, message2, message3]
+    history = _parse_chat_history_gemini(messages, convert_system_message_to_human=True)
+    assert len(history) == 3
+    assert history[0].role == "user"
+    assert history[0].parts[0].text == system_input
+    assert history[0].parts[1].text == text_question1
+    assert history[1].role == "model"
+    assert history[1].parts[0].text == text_answer1
+
+
 def test_default_params_palm() -> None:
     user_prompt = "Hello"