feat(genai): support gemini 3 (#1365)

Co-authored-by: Mason Daugherty <[email protected]>

Author: ccurme

libs/genai/langchain_google_genai/_common.py

@@ -56,6 +56,12 @@ class _BaseGoogleGenerativeAI(BaseModel):
     """Run inference with this temperature.
 
     Must be within `[0.0, 2.0]`.
+
+    !!! warning "Gemini 3.0+ models"
+
+        Setting `temperature < 1.0` for Gemini 3.0+ models can cause infinite loops,
+        degraded reasoning performance, and failure on complex tasks.
+
     """
 
     top_p: float | None = None
@@ -163,7 +169,21 @@ class _BaseGoogleGenerativeAI(BaseModel):
     media_resolution: MediaResolution | None = Field(
         default=None,
     )
-    """Media resolution for the input media."""
+    """Media resolution for the input media.
+
+    May be defined at the individual part level, allowing for mixed-resolution requests
+    (e.g., images and videos of different resolutions in the same request).
+
+    May be `'low'`, `'medium'`, or `'high'`.
+
+    Can be set either per-part or globally for all media inputs in the request. To set
+    globally, set in the `generation_config`.
+
+    !!! warning "Model compatibility"
+
+        Setting per-part media resolution requests to Gemini 2.5 models is not
+        supported.
+    """
 
     thinking_budget: int | None = Field(
         default=None,

libs/genai/langchain_google_genai/chat_models.py

@@ -124,7 +124,7 @@ def _generate(
             generations.append(
                 [
                     Generation(
-                        text=g.message.content,
+                        text=g.message.text,
                         generation_info={
                             **g.generation_info,
                             "usage_metadata": g.message.usage_metadata,
@@ -148,7 +148,7 @@ def _stream(
             run_manager=run_manager,
             **kwargs,
         ):
-            chunk = GenerationChunk(text=stream_chunk.message.content)
+            chunk = GenerationChunk(text=stream_chunk.message.text)
             yield chunk
             if run_manager:
                 run_manager.on_llm_new_token(

libs/genai/langchain_google_genai/llms.py

@@ -9,11 +9,11 @@ license = {text = "MIT"}
 readme = "README.md"
 authors = []
 
-version = "3.0.3"
+version = "3.1.0"
 requires-python = ">=3.10.0,<4.0.0"
 dependencies = [
-    "langchain-core>=1.0.0,<2.0.0",
-    "google-ai-generativelanguage>=0.7.0,<1.0.0",
+    "langchain-core>=1.0.5,<2.0.0",
+    "google-ai-generativelanguage>=0.9.0,<1.0.0",
     "pydantic>=2.0.0,<3.0.0",
     "filetype>=1.2.0,<2.0.0",
 ]

libs/genai/pyproject.toml

@@ -75,8 +75,16 @@ def _check_usage_metadata(message: AIMessage) -> None:
 def _check_tool_calls(response: BaseMessage, expected_name: str) -> None:
     """Check tool calls are as expected."""
     assert isinstance(response, AIMessage)
-    assert isinstance(response.content, str)
-    assert response.content == ""
+    if isinstance(response.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in response.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert text_content == ""
+    else:
+        assert isinstance(response.content, str)
+        assert response.content == ""
 
     # function_call
     function_call = response.additional_kwargs.get("function_call")
@@ -120,7 +128,15 @@ async def test_chat_google_genai_batch(is_async: bool, with_tags: bool) -> None:
         result = llm.batch(cast("list", messages), config=config)
 
     for token in result:
-        assert isinstance(token.content, str)
+        if isinstance(token.content, list):
+            text_content = "".join(
+                block.get("text", "")
+                for block in token.content
+                if isinstance(block, dict) and block.get("type") == "text"
+            )
+            assert len(text_content) > 0
+        else:
+            assert isinstance(token.content, str)
 
 
 @pytest.mark.parametrize("is_async", [False, True])
@@ -140,9 +156,19 @@ async def test_chat_google_genai_invoke(is_async: bool) -> None:
             config={"tags": ["foo"]},
             generation_config={"top_k": 2, "top_p": 1, "temperature": 0.7},
         )
+
     assert isinstance(result, AIMessage)
-    assert isinstance(result.content, str)
-    assert not result.content.startswith(" ")
+    if isinstance(result.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in result.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert len(text_content) > 0
+        assert not text_content.startswith(" ")
+    else:
+        assert isinstance(result.content, str)
+        assert not result.content.startswith(" ")
     _check_usage_metadata(result)
 
 
@@ -248,7 +274,15 @@ def test_chat_google_genai_invoke_thinking(
     )
 
     assert isinstance(result, AIMessage)
-    assert isinstance(result.content, str)
+    if isinstance(result.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in result.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert len(text_content) > 0
+    else:
+        assert isinstance(result.content, str)
 
     _check_usage_metadata(result)
 
@@ -265,7 +299,10 @@ def _check_thinking_output(content: list, output_version: str) -> None:
     if output_version == "v0":
         thinking_key = "thinking"
         if content:
-            assert isinstance(content[-1], str)
+            if isinstance(content[-1], dict) and content[-1].get("type") == "text":
+                assert isinstance(content[-1].get("text"), str)
+            else:
+                assert isinstance(content[-1], str)
 
     else:  # v1
         thinking_key = "reasoning"
@@ -540,7 +577,7 @@ def simple_tool(query: str) -> str:
 
 def test_chat_google_genai_invoke_thinking_disabled() -> None:
     """Test invoking a thinking model with zero `thinking_budget`."""
-    llm = ChatGoogleGenerativeAI(model=_THINKING_MODEL, thinking_budget=0)
+    llm = ChatGoogleGenerativeAI(model="gemini-2.5-flash", thinking_budget=0)
 
     result = llm.invoke(
         "How many O's are in Google? Please tell me how you double checked the result",
@@ -566,8 +603,17 @@ def test_chat_google_genai_invoke_no_image_generation_without_modalities() -> No
         generation_config={"top_k": 2, "top_p": 1, "temperature": 0.7},
     )
     assert isinstance(result, AIMessage)
-    assert isinstance(result.content, str)
-    assert not result.content.startswith(" ")
+    if isinstance(result.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in result.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert len(text_content) > 0
+        assert not text_content.startswith(" ")
+    else:
+        assert isinstance(result.content, str)
+        assert not result.content.startswith(" ")
     _check_usage_metadata(result)
 
 
@@ -617,16 +663,33 @@ def test_chat_google_genai_multimodal(
         any_chunk = False
         for chunk in llm.stream(messages):
             print(chunk)  # noqa: T201
-            assert isinstance(chunk.content, str)
-            if chunk.content:
-                any_chunk = True
+            if isinstance(chunk.content, list):
+                text_content = "".join(
+                    block.get("text", "")
+                    for block in chunk.content
+                    if isinstance(block, dict) and block.get("type") == "text"
+                )
+                if text_content:
+                    any_chunk = True
+            else:
+                assert isinstance(chunk.content, str)
+                if chunk.content:
+                    any_chunk = True
         assert any_chunk
     else:
         # Test invoke
         response = llm.invoke(messages)
         assert isinstance(response, AIMessage)
-        assert isinstance(response.content, str)
-        assert len(response.content.strip()) > 0
+        if isinstance(response.content, list):
+            text_content = "".join(
+                block.get("text", "")
+                for block in response.content
+                if isinstance(block, dict) and block.get("type") == "text"
+            )
+            assert len(text_content.strip()) > 0
+        else:
+            assert isinstance(response.content, str)
+            assert len(response.content.strip()) > 0
 
 
 @pytest.mark.parametrize(
@@ -674,7 +737,15 @@ def test_chat_google_genai_single_call_with_history() -> None:
     message3 = HumanMessage(content=text_question2)
     response = model.invoke([message1, message2, message3])
     assert isinstance(response, AIMessage)
-    assert isinstance(response.content, str)
+    if isinstance(response.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in response.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert len(text_content) > 0
+    else:
+        assert isinstance(response.content, str)
 
 
 @pytest.mark.parametrize(
@@ -700,7 +771,15 @@ def test_chat_google_genai_system_message(
     message3 = HumanMessage(content=text_question2)
     response = model.invoke([system_message, message1, message2, message3])
     assert isinstance(response, AIMessage)
-    assert isinstance(response.content, str)
+    if isinstance(response.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in response.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert len(text_content) > 0
+    else:
+        assert isinstance(response.content, str)
 
 
 def test_generativeai_get_num_tokens_gemini() -> None:
@@ -788,13 +867,28 @@ def search(
     assert len(tool_messages) > 0
     assert len(response.tool_calls) == len(tool_messages)
 
-    result = llm_with_search.invoke([request, response, *tool_messages])
+    follow_up = HumanMessage(
+        content=(
+            "Based on the search results above, what did you find about the bird "
+            "colors?"
+        )
+    )
+    result = llm_with_search.invoke([request, response, *tool_messages, follow_up])
 
     assert isinstance(result, AIMessage)
-    content_str = (
-        result.content if isinstance(result.content, str) else str(result.content)
-    )
-    assert "brown" in content_str.lower()
+
+    if isinstance(result.content, list):
+        text_content = "".join(
+            block.get("text", "")
+            for block in result.content
+            if isinstance(block, dict) and block.get("type") == "text"
+        )
+        assert "brown" in text_content.lower()
+    else:
+        content_str = (
+            result.content if isinstance(result.content, str) else str(result.content)
+        )
+        assert "brown" in content_str.lower()
 
 
 def test_chat_vertexai_gemini_function_calling() -> None:
@@ -862,9 +956,8 @@ def my_tool(name: str, age: int, likes: list[str]) -> None:
     # Test .content_blocks property
     content_blocks = response.content_blocks
     assert isinstance(content_blocks, list)
-    assert len(content_blocks) == 1
-    assert isinstance(content_blocks[0], dict)
-    assert content_blocks[0].get("type") == "tool_call"
+    tool_call_blocks = [b for b in content_blocks if b.get("type") == "tool_call"]
+    assert len(tool_call_blocks) == 1
 
 
 @pytest.mark.flaky(retries=3, delay=1)
@@ -1027,9 +1120,11 @@ def _check_web_search_output(message: AIMessage, output_version: str) -> None:
     # Lazy parsing
     content_blocks = message.content_blocks
     text_blocks = [block for block in content_blocks if block["type"] == "text"]
-    assert len(text_blocks) == 1
-    text_block = text_blocks[0]
-    assert text_block.get("annotations")
+    assert len(text_blocks) >= 1
+
+    # Check that at least one block has annotations
+    text_block = next((b for b in text_blocks if b.get("annotations")), None)
+    assert text_block is not None
 
     if output_version == "v1":
         text_blocks = [block for block in message.content if block["type"] == "text"]  # type: ignore[misc,index]
@@ -1186,8 +1281,19 @@ def test_search_builtin_with_citations(use_streaming: bool) -> None:
 def _check_code_execution_output(message: AIMessage, output_version: str) -> None:
     if output_version == "v0":
         blocks = [block for block in message.content if isinstance(block, dict)]
+        # Find code execution blocks
+        code_blocks = [
+            block
+            for block in blocks
+            if block.get("type") in {"executable_code", "code_execution_result"}
+        ]
+        # For integration test, code execution must happen
+        assert code_blocks, (
+            f"No code execution blocks found in content: "
+            f"{[block.get('type') for block in blocks]}"
+        )
         expected_block_types = {"executable_code", "code_execution_result"}
-        assert {block.get("type") for block in blocks} == expected_block_types
+        assert {block.get("type") for block in code_blocks} == expected_block_types
 
     else:
         # v1
@@ -1207,7 +1313,7 @@ def test_code_execution_builtin(output_version: str) -> None:
     ).bind_tools([{"code_execution": {}}])
     input_message = {
         "role": "user",
-        "content": "What is 3^3?",
+        "content": "Calculate the value of 3^3 using Python code execution.",
     }
 
     full: BaseMessageChunk | None = None

libs/genai/tests/integration_tests/test_chat_models.py

@@ -61,8 +61,6 @@ def search_tool(query: str) -> str:
     llm = ChatGoogleGenerativeAI(model=model_name).bind(functions=[search_tool])
     response = llm.invoke("weather in san francisco")
     assert isinstance(response, AIMessage)
-    assert isinstance(response.content, str)
-    assert response.content == ""
     function_call = response.additional_kwargs.get("function_call")
     assert function_call
     assert function_call["name"] == "search_tool"
@@ -85,8 +83,6 @@ def test_pydantic_call(model_name: str) -> None:
     llm = ChatGoogleGenerativeAI(model=model_name).bind(functions=[MyModel])
     response = llm.invoke("my name is Erick and I am 27 years old")
     assert isinstance(response, AIMessage)
-    assert isinstance(response.content, str)
-    assert response.content == ""
     function_call = response.additional_kwargs.get("function_call")
     assert function_call
     assert function_call["name"] == "MyModel"

libs/genai/tests/integration_tests/test_function_call.py

@@ -102,7 +102,7 @@ def test_safety_settings_gemini(model_name: str) -> None:
     streamed_messages = list(output_stream)
     assert len(streamed_messages) > 0
 
-    # test  with safety filters on instantiation
+    # test with safety filters on instantiation
     llm = GoogleGenerativeAI(
         model=model_name,
         safety_settings=safety_settings,