[wip] databricks-langchain: compatibility with gemini

integrations/langchain/src/databricks_langchain/chat_models.py

@@ -1055,6 +1055,7 @@ def _convert_message_to_dict(message: BaseMessage) -> dict:
         return {"role": "user", **message_dict}
     elif isinstance(message, AIMessage):
         if tool_calls := _get_tool_calls_from_ai_message(message):
+            print(tool_calls)
             message_dict["tool_calls"] = tool_calls  # type: ignore[assignment]
             # If tool calls present, content null value should be None not empty string.
             message_dict["content"] = message_dict["content"] or None  # type: ignore[assignment]
@@ -1196,15 +1197,37 @@ def _get_tool_calls_from_ai_message(message: AIMessage) -> List[Dict]:
         for tc in message.invalid_tool_calls
     ]
 
+    """
+    thought signature encodes model reasoning
+    it is required for each tool call to gemini 3 pro - https://arc.net/l/quote/jhoeoqbl
+    this means we need to encode this info in the responses events in order to fix this bug, in addition to the work on this PR
+
+    will have to change _langchain_message_stream_to_responses_stream
+    """
+
     if tool_calls or invalid_tool_calls:
-        return tool_calls + invalid_tool_calls
+        # Merge thoughtSignature from additional_kwargs if present
+        all_tool_calls = tool_calls + invalid_tool_calls
+        additional_tool_calls = message.additional_kwargs.get("tool_calls", [])
+        if additional_tool_calls:
+            # Create a mapping of tool call IDs to their thoughtSignature
+            thought_signatures = {
+                tc.get("id"): tc.get("thoughtSignature")
+                for tc in additional_tool_calls
+                if tc.get("thoughtSignature")
+            }
+            # Add thoughtSignature to matching tool calls
+            for tc in all_tool_calls:
+                if tc["id"] in thought_signatures:
+                    tc["thoughtSignature"] = thought_signatures[tc["id"]]
+        return all_tool_calls
 
     # Get tool calls from additional kwargs if present.
     return [
         {
             k: v
             for k, v in tool_call.items()  # type: ignore[union-attr]
-            if k in {"id", "type", "function"}
+            if k in {"id", "type", "function", "thoughtSignature"}
         }
         for tool_call in message.additional_kwargs.get("tool_calls", [])
     ]

integrations/langchain/tests/integration_tests/agent.py

@@ -0,0 +1,200 @@
+from typing import Annotated, Any, Generator, Optional, Sequence, TypedDict, Union
+
+import mlflow
+from langchain.messages import AIMessage, AIMessageChunk, AnyMessage
+from langchain_core.runnables import RunnableConfig, RunnableLambda
+from langchain_core.tools import BaseTool
+from langgraph.graph import END, StateGraph
+from langgraph.graph.message import add_messages
+from langgraph.prebuilt.tool_node import ToolNode
+from mlflow.pyfunc import ResponsesAgent
+from mlflow.types.responses import (
+    ResponsesAgentRequest,
+    ResponsesAgentResponse,
+    ResponsesAgentStreamEvent,
+    output_to_responses_items_stream,
+    to_chat_completions_input,
+)
+
+from databricks_langchain import (
+    ChatDatabricks,
+    UCFunctionToolkit,
+)
+
+############################################
+# Define your LLM endpoint and system prompt
+############################################
+# TODO: Replace with your model serving endpoint
+# LLM_ENDPOINT_NAME = "databricks-claude-sonnet-4-5"
+LLM_ENDPOINT_NAME = "databricks-gemini-3-pro"
+llm = ChatDatabricks(endpoint=LLM_ENDPOINT_NAME)
+
+# TODO: Update with your system prompt
+system_prompt = "You are a helpful assistant that can run Python code."
+
+###############################################################################
+## Define tools for your agent, enabling it to retrieve data or take actions
+## beyond text generation
+## To create and see usage examples of more tools, see
+## https://docs.databricks.com/en/generative-ai/agent-framework/agent-tool.html
+###############################################################################
+tools = []
+
+# You can use UDFs in Unity Catalog as agent tools
+# Below, we add the `system.ai.python_exec` UDF, which provides
+# a python code interpreter tool to our agent
+# You can also add local LangChain python tools. See https://python.langchain.com/docs/concepts/tools
+
+# TODO: Add additional tools
+UC_TOOL_NAMES = ["system.ai.python_exec"]
+uc_toolkit = UCFunctionToolkit(function_names=UC_TOOL_NAMES)
+tools.extend(uc_toolkit.tools)
+
+# Use Databricks vector search indexes as tools
+# See https://docs.databricks.com/en/generative-ai/agent-framework/unstructured-retrieval-tools.html#locally-develop-vector-search-retriever-tools-with-ai-bridge
+# List to store vector search tool instances for unstructured retrieval.
+VECTOR_SEARCH_TOOLS = []
+
+# To add vector search retriever tools,
+# use VectorSearchRetrieverTool and create_tool_info,
+# then append the result to TOOL_INFOS.
+# Example:
+# VECTOR_SEARCH_TOOLS.append(
+#     VectorSearchRetrieverTool(
+#         index_name="",
+#         # filters="..."
+#     )
+# )
+
+tools.extend(VECTOR_SEARCH_TOOLS)
+
+#####################
+## Define agent logic
+#####################
+
+
+class AgentState(TypedDict):
+    messages: Annotated[Sequence[AnyMessage], add_messages]
+    custom_inputs: Optional[dict[str, Any]]
+    custom_outputs: Optional[dict[str, Any]]
+
+
+def create_tool_calling_agent(
+    model: ChatDatabricks,
+    tools: Union[ToolNode, Sequence[BaseTool]],
+    system_prompt: Optional[str] = None,
+):
+    model = model.bind_tools(tools)
+
+    # Define the function that determines which node to go to
+    def should_continue(state: AgentState):
+        messages = state["messages"]
+        last_message = messages[-1]
+        # If there are function calls, continue. else, end
+        if isinstance(last_message, AIMessage) and last_message.tool_calls:
+            return "continue"
+        else:
+            return "end"
+
+    if system_prompt:
+        preprocessor = RunnableLambda(
+            lambda state: [{"role": "system", "content": system_prompt}] + state["messages"]
+        )
+    else:
+        preprocessor = RunnableLambda(lambda state: state["messages"])
+    model_runnable = preprocessor | model
+
+    def call_model(
+        state: AgentState,
+        config: RunnableConfig,
+    ):
+        response = model_runnable.invoke(state, config)
+
+        return {"messages": [response]}
+
+    workflow = StateGraph(AgentState)
+
+    workflow.add_node("agent", RunnableLambda(call_model))
+    workflow.add_node("tools", ToolNode(tools))
+
+    workflow.set_entry_point("agent")
+    workflow.add_conditional_edges(
+        "agent",
+        should_continue,
+        {
+            "continue": "tools",
+            "end": END,
+        },
+    )
+    workflow.add_edge("tools", "agent")
+
+    return workflow.compile()
+
+
+class LangGraphResponsesAgent(ResponsesAgent):
+    def __init__(self, agent):
+        self.agent = agent
+
+    def predict(self, request: ResponsesAgentRequest) -> ResponsesAgentResponse:
+        session_id = None
+        if request.custom_inputs and "session_id" in request.custom_inputs:
+            session_id = request.custom_inputs.get("session_id")
+        elif request.context and request.context.conversation_id:
+            session_id = request.context.conversation_id
+
+        if session_id:
+            mlflow.update_current_trace(
+                metadata={
+                    "mlflow.trace.session": session_id,
+                }
+            )
+
+        outputs = [
+            event.item
+            for event in self.predict_stream(request)
+            if event.type == "response.output_item.done"
+        ]
+        return ResponsesAgentResponse(output=outputs, custom_outputs=request.custom_inputs)
+
+    def predict_stream(
+        self,
+        request: ResponsesAgentRequest,
+    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
+        session_id = None
+        if request.custom_inputs and "session_id" in request.custom_inputs:
+            session_id = request.custom_inputs.get("session_id")
+        elif request.context and request.context.conversation_id:
+            session_id = request.context.conversation_id
+
+        if session_id:
+            mlflow.update_current_trace(
+                metadata={
+                    "mlflow.trace.session": session_id,
+                }
+            )
+
+        cc_msgs = to_chat_completions_input([i.model_dump() for i in request.input])
+
+        for event in self.agent.stream({"messages": cc_msgs}, stream_mode=["updates", "messages"]):
+            if event[0] == "updates":
+                for node_data in event[1].values():
+                    if len(node_data.get("messages", [])) > 0:
+                        yield from output_to_responses_items_stream(node_data["messages"])
+            # filter the streamed messages to just the generated text messages
+            elif event[0] == "messages":
+                try:
+                    chunk = event[1][0]
+                    if isinstance(chunk, AIMessageChunk) and (content := chunk.content):
+                        yield ResponsesAgentStreamEvent(
+                            **self.create_text_delta(delta=content, item_id=chunk.id),
+                        )
+                except Exception as e:
+                    print(e)
+
+
+# Create the agent object, and specify it as the agent object to use when
+# loading the agent back for inference via mlflow.models.set_model()
+mlflow.langchain.autolog()
+agent = create_tool_calling_agent(llm, tools, system_prompt)
+AGENT = LangGraphResponsesAgent(agent)
+mlflow.models.set_model(AGENT)

integrations/langchain/tests/integration_tests/test_chat_models.py

@@ -243,9 +243,9 @@ class GetWeather(BaseModel):
         return
 
     # Models should make at least one tool call when tool_choice is not "none"
-    assert len(response.tool_calls) >= 1, (
-        f"Expected at least 1 tool call, got {len(response.tool_calls)}"
-    )
+    assert (
+        len(response.tool_calls) >= 1
+    ), f"Expected at least 1 tool call, got {len(response.tool_calls)}"
 
     # The first tool call should be for GetWeather
     first_call = response.tool_calls[0]
@@ -267,9 +267,9 @@ class GetWeather(BaseModel):
         ]
     )
     # Should call GetWeather tool for the followup question
-    assert len(response.tool_calls) >= 1, (
-        f"Expected at least 1 tool call, got {len(response.tool_calls)}"
-    )
+    assert (
+        len(response.tool_calls) >= 1
+    ), f"Expected at least 1 tool call, got {len(response.tool_calls)}"
     tool_call = response.tool_calls[0]
     assert tool_call["name"] == "GetWeather", f"Expected GetWeather tool, got {tool_call['name']}"
     assert "location" in tool_call["args"], f"Expected location in args, got {tool_call['args']}"
@@ -559,12 +559,8 @@ def test_chat_databricks_chatagent_invoke():
                     ):
                         python_tool_used = True
 
-    assert has_tool_calls, (
-        f"Expected ChatAgent to use tool calls for fibonacci computation. Content: {response.content}"
-    )
-    assert python_tool_used, (
-        f"Expected ChatAgent to use python execution tool for fibonacci computation. Content: {response.content}"
-    )
+    assert has_tool_calls, f"Expected ChatAgent to use tool calls for fibonacci computation. Content: {response.content}"
+    assert python_tool_used, f"Expected ChatAgent to use python execution tool for fibonacci computation. Content: {response.content}"
 
 
 @pytest.mark.st_endpoints
@@ -847,9 +843,9 @@ def test_chat_databricks_gpt5_stream_with_usage():
     ]
 
     # Should have exactly ONE usage chunk from the final usage-only chunk
-    assert len(usage_chunks) == 1, (
-        f"Expected exactly 1 usage chunk from GPT-5 final chunk, got {len(usage_chunks)}"
-    )
+    assert (
+        len(usage_chunks) == 1
+    ), f"Expected exactly 1 usage chunk from GPT-5 final chunk, got {len(usage_chunks)}"
 
     # Verify usage chunk has correct metadata structure
     usage_chunk = usage_chunks[0]
@@ -860,12 +856,12 @@ def test_chat_databricks_gpt5_stream_with_usage():
     assert "total_tokens" in usage_chunk.usage_metadata
 
     # Verify token counts are positive
-    assert usage_chunk.usage_metadata["input_tokens"] > 0, (
-        f"Expected positive input_tokens, got {usage_chunk.usage_metadata['input_tokens']}"
-    )
-    assert usage_chunk.usage_metadata["output_tokens"] > 0, (
-        f"Expected positive output_tokens, got {usage_chunk.usage_metadata['output_tokens']}"
-    )
+    assert (
+        usage_chunk.usage_metadata["input_tokens"] > 0
+    ), f"Expected positive input_tokens, got {usage_chunk.usage_metadata['input_tokens']}"
+    assert (
+        usage_chunk.usage_metadata["output_tokens"] > 0
+    ), f"Expected positive output_tokens, got {usage_chunk.usage_metadata['output_tokens']}"
 
     # Verify total_tokens equals sum of input and output
     expected_total = (
@@ -875,3 +871,40 @@ def test_chat_databricks_gpt5_stream_with_usage():
         f"Expected total_tokens ({usage_chunk.usage_metadata['total_tokens']}) "
         f"to equal input_tokens + output_tokens ({expected_total})"
     )
+
+
+def test_chat_databricks_with_gemini():
+    os.environ["DATABRICKS_CONFIG_PROFILE"] = "dogfood"
+    from .agent import AGENT
+
+    result = AGENT.predict(
+        {
+            "input": [
+                {"role": "user", "content": "What is 6*7 in Python?"},
+                {
+                    "type": "function_call",
+                    "id": "lc_run--e58dec26-ce5d-4597-b4f8-28e6db62cd49",
+                    "call_id": "system__ai__python_exec",
+                    "name": "system__ai__python_exec",
+                    "arguments": '{"code": "print(6 * 7)"}',
+                },
+                {
+                    "type": "function_call_output",
+                    "call_id": "system__ai__python_exec",
+                    "output": '{"format": "SCALAR", "value": "42\\n"}',
+                },
+                {
+                    "type": "message",
+                    "id": "lc_run--dd658def-dfdc-4bc7-b0d9-b6e25d1ecc48",
+                    "content": [
+                        {"text": "The result of `6 * 7` in Python is 42.", "type": "output_text"}
+                    ],
+                    "role": "assistant",
+                },
+            ]
+        }
+    )
+    assert result is not None
+    assert result.output is not None
+    print(result.model_dump())
+    assert False