Catch langchain agent errors (#1539)

* Catch any uncaught/un-handled langchain agent errors, return an error string so that we don't end up with a missing result.
* Fix version conflicts for nemollm and dgl
* Expand existing tests for the `LangChainAgentNode`

## By Submitting this PR I confirm:
- I am familiar with the [Contributing Guidelines](https://github.com/nv-morpheus/Morpheus/blob/main/docs/source/developer_guide/contributing.md).
- When the PR is ready for review, new or existing tests cover these changes.
- When the PR is ready for review, the documentation is up to date with these changes.

Authors:
  - David Gardner (https://github.com/dagardner-nv)

Approvers:
  - Christopher Harris (https://github.com/cwharris)
  - Michael Demoret (https://github.com/mdemoret-nv)

URL: #1539

conda/environments/all_cuda-121_arch-x86_64.yaml

@@ -93,7 +93,9 @@ dependencies:
 - pytorch=*=*cuda*
 - rapidjson=1.1.0
 - rdma-core>=48
+- requests
 - requests-cache=1.1
+- requests-toolbelt
 - s3fs=2023.12.2
 - scikit-build=0.17.6
 - scikit-learn=1.3.2
@@ -117,7 +119,7 @@ dependencies:
   - --find-links https://data.dgl.ai/wheels/cu121/repo.html
   - PyMuPDF==1.23.21
   - databricks-connect
-  - dgl
+  - dgl==2.0.0
   - dglgo
   - google-search-results==2.4
   - langchain==0.1.9

conda/environments/dev_cuda-121_arch-x86_64.yaml

@@ -73,7 +73,9 @@ dependencies:
 - pytorch-cuda
 - pytorch=*=*cuda*
 - rapidjson=1.1.0
+- requests
 - requests-cache=1.1
+- requests-toolbelt
 - scikit-build=0.17.6
 - scikit-learn=1.3.2
 - sphinx

conda/environments/examples_cuda-121_arch-x86_64.yaml

@@ -46,7 +46,9 @@ dependencies:
 - python=3.10
 - pytorch-cuda
 - pytorch=*=*cuda*
+- requests
 - requests-cache=1.1
+- requests-toolbelt
 - s3fs=2023.12.2
 - scikit-learn=1.3.2
 - sentence-transformers
@@ -61,7 +63,7 @@ dependencies:
   - --find-links https://data.dgl.ai/wheels/cu121/repo.html
   - PyMuPDF==1.23.21
   - databricks-connect
-  - dgl
+  - dgl==2.0.0
   - dglgo
   - google-search-results==2.4
   - langchain==0.1.9

conda/environments/runtime_cuda-121_arch-x86_64.yaml

@@ -28,7 +28,9 @@ dependencies:
 - python=3.10
 - pytorch-cuda
 - pytorch=*=*cuda*
+- requests
 - requests-cache=1.1
+- requests-toolbelt
 - scikit-learn=1.3.2
 - sqlalchemy
 - tqdm=4

dependencies.yaml

@@ -261,7 +261,9 @@ dependencies:
           - python-graphviz
           - pytorch-cuda
           - pytorch=*=*cuda*
+          - requests
           - requests-cache=1.1
+          - requests-toolbelt # Transitive dep needed by nemollm, specified here to ensure we get a compatible version
           - sqlalchemy
           - tqdm=4
           - typing_utils=0.1
@@ -311,7 +313,7 @@ dependencies:
           - pip:
             - --find-links https://data.dgl.ai/wheels/cu121/repo.html
             - --find-links https://data.dgl.ai/wheels-test/repo.html
-            - dgl
+            - dgl==2.0.0
             - dglgo
 
   example-llm-agents:

morpheus/llm/nodes/langchain_agent_node.py

@@ -66,9 +66,14 @@ async def _run_single(self, **kwargs: dict[str, typing.Any]) -> dict[str, typing
             return results
 
         # We are not dealing with a list, so run single
-        return await self._agent_executor.arun(**kwargs)
-
-    async def execute(self, context: LLMContext) -> LLMContext:
+        try:
+            return await self._agent_executor.arun(**kwargs)
+        except Exception as e:
+            error_msg = f"Error running agent: {e}"
+            logger.exception(error_msg)
+            return error_msg
+
+    async def execute(self, context: LLMContext) -> LLMContext:  # pylint: disable=invalid-overridden-method
 
         input_dict = context.get_inputs()
 

tests/_utils/llm.py

@@ -87,7 +87,35 @@ def mk_mock_openai_response(messages: list[str]) -> mock.MagicMock:
     Creates a mocked openai.types.chat.chat_completion.ChatCompletion response with the given messages.
     """
     response = mock.MagicMock()
-    mock_choices = [_mk_mock_choice(message) for message in messages]
-    response.choices = mock_choices
+
+    response.choices = [_mk_mock_choice(message) for message in messages]
+    response.dict.return_value = {
+        "choices": [{
+            'message': {
+                'role': 'assistant', 'content': message
+            }
+        } for message in messages]
+    }
 
     return response
+
+
+def mk_mock_langchain_tool(responses: list[str]) -> mock.MagicMock:
+    """
+    Creates a mocked LangChainTestTool with the given responses.
+    """
+
+    # Langchain will call inspect.signature on the tool methods, typically mock objects don't have a signature,
+    # explicitly providing one here
+    async def _arun_spec(*_, **__):
+        pass
+
+    def run_spec(*_, **__):
+        pass
+
+    tool = mock.MagicMock()
+    tool.arun = mock.create_autospec(spec=_arun_spec)
+    tool.arun.side_effect = responses
+    tool.run = mock.create_autospec(run_spec)
+    tool.run.side_effect = responses
+    return tool

tests/llm/nodes/test_langchain_agent_node.py

@@ -16,8 +16,14 @@
 from unittest import mock
 
 import pytest
+from langchain.agents import AgentType
+from langchain.agents import Tool
+from langchain.agents import initialize_agent
+from langchain.chat_models import ChatOpenAI  # pylint: disable=no-name-in-module
 
 from _utils.llm import execute_node
+from _utils.llm import mk_mock_langchain_tool
+from _utils.llm import mk_mock_openai_response
 from morpheus.llm import LLMNodeBase
 from morpheus.llm.nodes.langchain_agent_node import LangChainAgentNode
 
@@ -50,8 +56,90 @@ def test_execute(
     expected_output: list,
     expected_calls: list[mock.call],
 ):
+    # Tests the execute method of the LangChainAgentNode with a mocked agent_executor
     mock_agent_executor.arun.return_value = arun_return
 
     node = LangChainAgentNode(agent_executor=mock_agent_executor)
     assert execute_node(node, **values) == expected_output
     mock_agent_executor.arun.assert_has_calls(expected_calls)
+
+
+def test_execute_tools(mock_chat_completion: tuple[mock.MagicMock, mock.MagicMock]):
+    # Tests the execute method of the LangChainAgentNode with a a mocked tools and chat completion
+    (_, mock_async_client) = mock_chat_completion
+    chat_responses = [
+        'I should check Tool1\nAction: Tool1\nAction Input: "name a reptile"',
+        'I should check Tool2\nAction: Tool2\nAction Input: "name of a day of the week"',
+        'I should check Tool1\nAction: Tool1\nAction Input: "name a reptile"',
+        'I should check Tool2\nAction: Tool2\nAction Input: "name of a day of the week"',
+        'Observation: Answer: Yes!\nI now know the final answer.\nFinal Answer: Yes!'
+    ]
+    mock_responses = [mk_mock_openai_response([response]) for response in chat_responses]
+    mock_async_client.chat.completions.create.side_effect = mock_responses
+
+    llm_chat = ChatOpenAI(model="fake-model", openai_api_key="fake-key")
+
+    mock_tool1 = mk_mock_langchain_tool(["lizard", "frog"])
+    mock_tool2 = mk_mock_langchain_tool(["Tuesday", "Thursday"])
+
+    tools = [
+        Tool(name="Tool1",
+             func=mock_tool1.run,
+             coroutine=mock_tool1.arun,
+             description="useful for when you need to know the name of a reptile"),
+        Tool(name="Tool2",
+             func=mock_tool2.run,
+             coroutine=mock_tool2.arun,
+             description="useful for when you need to know the day of the week")
+    ]
+
+    agent = initialize_agent(tools,
+                             llm_chat,
+                             agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
+                             verbose=True,
+                             handle_parsing_errors=True,
+                             early_stopping_method="generate",
+                             return_intermediate_steps=False)
+
+    node = LangChainAgentNode(agent_executor=agent)
+
+    assert execute_node(node, input="input1") == "Yes!"
+
+
+def test_execute_error(mock_chat_completion: tuple[mock.MagicMock, mock.MagicMock]):
+    # Tests the execute method of the LangChainAgentNode with a a mocked tools and chat completion
+    (_, mock_async_client) = mock_chat_completion
+    chat_responses = [
+        'I should check Tool1\nAction: Tool1\nAction Input: "name a reptile"',
+        'I should check Tool2\nAction: Tool2\nAction Input: "name of a day of the week"',
+        'Observation: Answer: Yes!\nI now know the final answer.\nFinal Answer: Yes!'
+    ]
+    mock_responses = [mk_mock_openai_response([response]) for response in chat_responses]
+    mock_async_client.chat.completions.create.side_effect = mock_responses
+
+    llm_chat = ChatOpenAI(model="fake-model", openai_api_key="fake-key")
+
+    mock_tool1 = mk_mock_langchain_tool(["lizard"])
+    mock_tool2 = mk_mock_langchain_tool(RuntimeError("unittest"))
+
+    tools = [
+        Tool(name="Tool1",
+             func=mock_tool1.run,
+             coroutine=mock_tool1.arun,
+             description="useful for when you need to know the name of a reptile"),
+        Tool(name="Tool2",
+             func=mock_tool2.run,
+             coroutine=mock_tool2.arun,
+             description="useful for when you need to test tool errors")
+    ]
+
+    agent = initialize_agent(tools,
+                             llm_chat,
+                             agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
+                             verbose=True,
+                             handle_parsing_errors=True,
+                             early_stopping_method="generate",
+                             return_intermediate_steps=False)
+
+    node = LangChainAgentNode(agent_executor=agent)
+    assert execute_node(node, input="input1") == "Error running agent: unittest"