merge

README.md

@@ -1,5 +1,5 @@
 # 🦜️🔗 LangChain Google
 
-This repository contains two packages with Google integrations with Langhchain:
+This repository contains two packages with Google integrations with LangChain:
 - [langchain-google-genai](https://pypi.org/project/langchain-google-genai/) implements integrations of Google [Generative AI](https://ai.google.dev/) models.
 - [langchain-google-vertexai](https://pypi.org/project/langchain-google-vertexai/) implements integrations of Google Cloud [Generative AI on Vertex AI](https://cloud.google.com/vertex-ai/docs/generative-ai/learn/overview)

libs/genai/langchain_google_genai/chat_models.py

@@ -4,6 +4,7 @@
 import json
 import logging
 import os
+import warnings
 from io import BytesIO
 from typing import (
     Any,
@@ -300,27 +301,16 @@ def _convert_to_parts(
 
 def _parse_chat_history(
     input_messages: Sequence[BaseMessage], convert_system_message_to_human: bool = False
-) -> List[genai.types.ContentDict]:
+) -> Tuple[Optional[genai.types.ContentDict], List[genai.types.ContentDict]]:
     messages: List[genai.types.MessageDict] = []
 
-    raw_system_message: Optional[SystemMessage] = None
-    for i, message in enumerate(input_messages):
-        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:
+    if convert_system_message_to_human:
+        warnings.warn("Convert_system_message_to_human will be deprecated!")
 
-llm = ChatGoogleGenerativeAI(model="gemini-pro", convert_system_message_to_human=True)
-"""
-            )
-        elif i == 0 and isinstance(message, SystemMessage):
-            raw_system_message = message
+    system_instruction: Optional[genai.types.ContentDict] = None
+    for i, message in enumerate(input_messages):
+        if i == 0 and isinstance(message, SystemMessage):
+            system_instruction = _convert_to_parts(message.content)
             continue
         elif isinstance(message, AIMessage):
             role = "model"
@@ -365,16 +355,8 @@ def _parse_chat_history(
                 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.content) + parts
-            raw_system_message = None
         messages.append({"role": role, "parts": parts})
-    return messages
+    return system_instruction, messages
 
 
 def _parse_response_candidate(
@@ -659,11 +641,15 @@ def _prepare_chat(
             )
 
         params = self._prepare_params(stop, **kwargs)
-        history = _parse_chat_history(
+        system_instruction, history = _parse_chat_history(
             messages,
             convert_system_message_to_human=self.convert_system_message_to_human,
         )
         message = history.pop()
+        if self.client._system_instruction != system_instruction:
+            self.client = genai.GenerativeModel(
+                model_name=self.model, system_instruction=system_instruction
+            )
         chat = client.start_chat(history=history)
         return params, chat, message
 

libs/genai/pyproject.toml

@@ -12,8 +12,8 @@ license = "MIT"
 
 [tool.poetry.dependencies]
 python = ">=3.9,<4.0"
-langchain-core = "^0.1"
-google-generativeai = "^0.4.1"
+langchain-core = ">=0.1.27,<0.2"
+google-generativeai = "^0.5.0"
 pillow = { version = "^10.1.0", optional = true }
 
 [tool.poetry.extras]
@@ -30,6 +30,7 @@ syrupy = "^4.0.2"
 pytest-watcher = "^0.3.4"
 pytest-asyncio = "^0.21.1"
 numpy = "^1.26.2"
+langchain-core = {git = "https://github.com/langchain-ai/langchain.git", subdirectory = "libs/core"}
 
 [tool.poetry.group.codespell]
 optional = true
@@ -56,6 +57,7 @@ types-requests = "^2.28.11.5"
 types-google-cloud-ndb = "^2.2.0.1"
 types-pillow = "^10.1.0.2"
 types-protobuf = "^4.24.0.20240302"
+langchain-core = {git = "https://github.com/langchain-ai/langchain.git", subdirectory = "libs/core"}
 
 [tool.poetry.group.dev]
 optional = true
@@ -65,6 +67,7 @@ pillow = "^10.1.0"
 types-requests = "^2.31.0.10"
 types-pillow = "^10.1.0.2"
 types-google-cloud-ndb = "^2.2.0.1"
+langchain-core = {git = "https://github.com/langchain-ai/langchain.git", subdirectory = "libs/core"}
 
 [tool.ruff]
 select = [

libs/genai/tests/integration_tests/test_chat_models.py

@@ -112,6 +112,16 @@ def test_chat_google_genai_invoke_multimodal() -> None:
         assert len(chunk.content.strip()) > 0
 
 
+def test_system_message() -> None:
+    messages = [
+        SystemMessage(content="Be a helful assistant."),
+        HumanMessage(content="Hi, how are you?"),
+    ]
+    llm = ChatGoogleGenerativeAI(model="models/gemini-1.0-pro-latest")
+    answer = llm.invoke(messages)
+    assert isinstance(answer.content, str)
+
+
 def test_chat_google_genai_invoke_multimodal_too_many_messages() -> None:
     # Only supports 1 turn...
     messages: list = [
@@ -168,18 +178,6 @@ def test_chat_google_genai_single_call_with_history() -> None:
     assert isinstance(response.content, str)
 
 
-def test_chat_google_genai_system_message_error() -> None:
-    model = ChatGoogleGenerativeAI(model=_MODEL)
-    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])
-
-
 def test_chat_google_genai_system_message() -> None:
     model = ChatGoogleGenerativeAI(model=_MODEL, convert_system_message_to_human=True)
     text_question1, text_answer1 = "How much is 2+2?", "4"

libs/genai/tests/integration_tests/test_llms.py

@@ -75,7 +75,7 @@ def test_safety_settings_gemini() -> None:
     llm = GoogleGenerativeAI(temperature=0, model="gemini-pro")
     output = llm.generate(prompts=["how to make a bomb?"])
     assert isinstance(output, LLMResult)
-    assert len(output.generations[0]) == 0
+    assert len(output.generations[0]) > 0
 
     # safety filters
     safety_settings = {

libs/genai/tests/unit_tests/test_chat_models.py

@@ -61,13 +61,16 @@ def test_parse_history() -> None:
     message2 = AIMessage(content=text_answer1)
     message3 = HumanMessage(content=text_question2)
     messages = [system_message, message1, message2, message3]
-    history = _parse_chat_history(messages, convert_system_message_to_human=True)
+    system_instruction, history = _parse_chat_history(
+        messages, convert_system_message_to_human=True
+    )
     assert len(history) == 3
     assert history[0] == {
         "role": "user",
-        "parts": [{"text": system_input}, {"text": text_question1}],
+        "parts": [{"text": text_question1}],
     }
     assert history[1] == {"role": "model", "parts": [{"text": text_answer1}]}
+    assert system_instruction == [{"text": system_input}]
 
 
 @pytest.mark.parametrize("content", ['["a"]', '{"a":"b"}', "function output"])