Merge pull request #2001 from solliancenet/cp-langraph-react-workflow

Introduction of the LangGraph ReAct workflow

src/dotnet/Common/Models/ResourceProviders/Agent/AgentWorkflows/AgentWorkflowBase.cs

@@ -8,6 +8,7 @@ namespace FoundationaLLM.Common.Models.ResourceProviders.Agent.AgentWorkflows
     [JsonPolymorphic(TypeDiscriminatorPropertyName = "type")]
     [JsonDerivedType(typeof(AzureOpenAIAssistantsAgentWorkflow), AgentWorkflowTypes.AzureOpenAIAssistants)]
     [JsonDerivedType(typeof(LangChainExpressionLanguageAgentWorkflow), AgentWorkflowTypes.LangChainExpressionLanguage)]
+    [JsonDerivedType(typeof(LangGraphReactAgentWorkflow), AgentWorkflowTypes.LangGraphReactAgent)]
     public class AgentWorkflowBase
     {        
         /// <summary>

src/dotnet/Common/Models/ResourceProviders/Agent/AgentWorkflows/AgentWorkflowTypes.cs

@@ -14,5 +14,10 @@ public static class AgentWorkflowTypes
         /// The LangChain Expression Language agent workflow.
         /// </summary>
         public const string LangChainExpressionLanguage = "langchain-expression-language-workflow";
+
+        /// <summary>
+        /// The LangGraph ReAct agent workflow.
+        /// </summary>
+        public const string LangGraphReactAgent = "langgraph-react-agent-workflow";
     }
 }

src/dotnet/Common/Models/ResourceProviders/Agent/AgentWorkflows/LangGraphReactAgentWorkflow.cs

@@ -0,0 +1,14 @@
+using System.Text.Json.Serialization;
+
+namespace FoundationaLLM.Common.Models.ResourceProviders.Agent.AgentWorkflows
+{
+    /// <summary>
+    /// Provides an agent workflow configuration for a LangGraph ReAct Agent workflow.
+    /// </summary>
+    public class LangGraphReactAgentWorkflow : AgentWorkflowBase
+    {
+        /// <inheritdoc/>
+        [JsonIgnore]
+        public override string Type => AgentWorkflowTypes.LangGraphReactAgent;
+    }
+}

src/python/PythonSDK/PythonSDK.pyproj

@@ -34,6 +34,7 @@
     <Compile Include="foundationallm\models\agents\agent_workflows\agent_workflow_base.py" />
     <Compile Include="foundationallm\models\agents\agent_workflows\azure_openai_assistants_agent_workflow.py" />
     <Compile Include="foundationallm\models\agents\agent_workflows\langchain_expression_language_agent_workflow.py" />
+    <Compile Include="foundationallm\models\agents\agent_workflows\langgraph_react_agent_workflow.py" />
     <Compile Include="foundationallm\models\attachments\attachment_properties.py" />
     <Compile Include="foundationallm\models\attachments\attachment_providers.py" />
     <Compile Include="foundationallm\models\attachments\__init__.py" />

src/python/PythonSDK/foundationallm/langchain/agents/langchain_agent_base.py

@@ -4,6 +4,7 @@
 
 from langchain_core.language_models import BaseLanguageModel
 from langchain_aws import ChatBedrockConverse
+from langchain_core.messages import BaseMessage, AIMessage, HumanMessage
 from langchain_openai import AzureChatOpenAI, ChatOpenAI, OpenAI
 from openai import AsyncAzureOpenAI as async_aoi
 from foundationallm.config import Configuration, UserIdentity
@@ -173,6 +174,31 @@ def _build_conversation_history(self, messages:List[MessageHistoryItem]=None, me
         chat_history += "\n\n"
         return chat_history
 
+    def _build_conversation_history_message_list(self, messages:List[MessageHistoryItem]=None, message_count:int=None) -> List[BaseMessage]:
+        """
+        Builds a LangChain Message chat history list from a list of MessageHistoryItem objects to
+        be added to the prompt template for the completion request.
+
+        Parameters
+        ----------
+        messages : List[MessageHistoryItem]
+            The list of messages from which to build the chat history.
+        message_count : int
+            The number of messages to include in the chat history.
+        """
+        if messages is None or len(messages)==0:
+            return []
+        if message_count is not None:
+            messages = messages[-message_count:]
+        history = []                   
+        for msg in messages:
+            # sender can be User (maps to HumanMessage) or Agent (maps to AIMessage)
+            if msg.sender == "User":
+                history.append(HumanMessage(content=msg.text))
+            else:
+                history.append(AIMessage(content=msg.text))        
+        return history
+
     def _record_full_prompt(self, prompt: str) -> str:
         """
         Records the full prompt for the completion request.

src/python/PythonSDK/foundationallm/langchain/agents/langchain_knowledge_management_agent.py

@@ -1,21 +1,21 @@
 import uuid
 from langchain_community.callbacks import get_openai_callback
-from langchain_community.callbacks.manager import get_bedrock_anthropic_callback
+from langchain_core.messages import HumanMessage
 from langchain_core.prompts import PromptTemplate
 from langchain_core.runnables import RunnablePassthrough, RunnableLambda
 from langchain_core.output_parsers import StrOutputParser
+from langgraph.prebuilt import create_react_agent
 from foundationallm.langchain.agents import LangChainAgentBase
 from foundationallm.langchain.exceptions import LangChainException
 from foundationallm.langchain.retrievers import RetrieverFactory, CitationRetrievalBase
-from foundationallm.models.agents.agent_workflows.azure_openai_assistants_agent_workflow import AzureOpenAIAssistantsAgentWorkflow
+from foundationallm.models.agents import AzureOpenAIAssistantsAgentWorkflow, LangGraphReactAgentWorkflow
 from foundationallm.models.constants import AgentCapabilityCategories
 from foundationallm.models.operations import OperationTypes
 from foundationallm.models.orchestration import (
     CompletionRequestObjectKeys,
     CompletionResponse,
     OpenAITextMessageContentItem
 )
-from foundationallm.models.resource_providers.ai_models import CompletionAIModel
 from foundationallm.models.resource_providers.configuration import APIEndpointConfiguration
 from foundationallm.models.agents import (
     AgentConversationHistorySettings,
@@ -44,7 +44,10 @@ class LangChainKnowledgeManagementAgent(LangChainAgentBase):
     """
     The LangChain Knowledge Management agent.
     """
-
+
+    MAIN_MODEL_KEY = "main_model"
+    MAIN_PROMPT_KEY = "main_prompt"
+
     def _get_document_retriever(
         self,
         request: KnowledgeManagementCompletionRequest,
@@ -174,12 +177,12 @@ def _validate_request(self, request: KnowledgeManagementCompletionRequest):
             raise LangChainException("The objects property on the completion request cannot be null.", 400)
 
         if request.agent.workflow is not None:
-            if request.agent.workflow.agent_workflow_ai_models["main_model"] is None:
+            if request.agent.workflow.agent_workflow_ai_models[self.MAIN_MODEL_KEY] is None:
                 raise LangChainException("The agent's workflow AI models requires a main_model.", 400)
-            if request.agent.workflow.prompt_object_ids["main_prompt"] is None:
+            if request.agent.workflow.prompt_object_ids[self.MAIN_PROMPT_KEY] is None:
                 raise LangChainException("The agent's workflow prompt object dictionary requires a main_prompt.", 400)
-            self.ai_model = self._get_ai_model_from_object_id(request.agent.workflow.agent_workflow_ai_models["main_model"].ai_model_object_id, request.objects)
-            self.prompt = self._get_prompt_from_object_id(request.agent.workflow.prompt_object_ids["main_prompt"], request.objects)
+            self.ai_model = self._get_ai_model_from_object_id(request.agent.workflow.agent_workflow_ai_models[self.MAIN_MODEL_KEY].ai_model_object_id, request.objects)
+            self.prompt = self._get_prompt_from_object_id(request.agent.workflow.prompt_object_ids[self.MAIN_PROMPT_KEY], request.objects)
         else:
             # Legacy code
             self.ai_model = self._get_ai_model_from_object_id(request.agent.ai_model_object_id, request.objects)
@@ -275,6 +278,7 @@ async def invoke_async(self, request: KnowledgeManagementCompletionRequest) -> C
             generated full prompt with context and token utilization and execution cost details.
         """
         self._validate_request(request)
+        llm = self._get_language_model()
 
         agent = request.agent
         image_analysis_token_usage = CompletionUsage(prompt_tokens=0, completion_tokens=0, total_tokens=0)
@@ -373,7 +377,7 @@ async def invoke_async(self, request: KnowledgeManagementCompletionRequest) -> C
             image_service = None
             if "dalle-image-generation" in request.agent.tools:
                 dalle_tool = request.agent.tools["dalle-image-generation"]
-                model_object_id = dalle_tool["ai_model_object_ids"]["main_model"]
+                model_object_id = dalle_tool["ai_model_object_ids"][self.MAIN_MODEL_KEY]
                 image_generation_deployment_model = request.objects[model_object_id]["deployment_name"]
                 api_endpoint_object_id = request.objects[model_object_id]["endpoint_object_id"]
                 image_generation_client = self._get_image_gen_language_model(api_endpoint_object_id=api_endpoint_object_id, objects=request.objects)
@@ -414,6 +418,48 @@ async def invoke_async(self, request: KnowledgeManagementCompletionRequest) -> C
             )
         # End Assistants API implementation
 
+        # Start LangGraph ReAct Agent workflow implementation
+        if (agent.workflow is not None and isinstance(agent.workflow, LangGraphReactAgentWorkflow)):            
+            # Temporary placeholder
+            from typing import Literal
+            from langchain_core.tools import tool
+            @tool
+            def get_weather(city: Literal["nyc", "sf"]):
+                """Use this to get weather information."""
+                if city == "nyc":
+                    return "It might be cloudy in nyc"
+                elif city == "sf":
+                    return "It's always sunny in sf"
+                else:
+                    raise AssertionError("Unknown city")
+            tools = [get_weather]
+            # End temporary placeholder
+
+            # Define the graph          
+            graph = create_react_agent(llm, tools=tools, state_modifier=self.prompt.prefix)
+            messages = self._build_conversation_history_message_list(request.message_history, agent.conversation_history_settings.max_history)
+            messages.append(HumanMessage(content=request.user_prompt))
+            response = await graph.ainvoke({'messages': messages})
+            # TODO: process tool messages with analysis results AIMessage with content='' but has addition_kwargs={'tool_calls';[...]}
+            # print(response)
+            final_message = response["messages"][-1]
+            response_content = OpenAITextMessageContentItem(
+                value = final_message.content,
+                agent_capability_category = AgentCapabilityCategories.FOUNDATIONALLM_KNOWLEDGE_MANAGEMENT
+            )
+            return CompletionResponse(
+                        operation_id = request.operation_id,
+                        content = [response_content],
+                        citations = [],
+                        user_prompt = request.user_prompt,
+                        full_prompt = self.prompt.prefix,
+                        completion_tokens = final_message.usage_metadata["output_tokens"] or 0,
+                        prompt_tokens = final_message.usage_metadata["input_tokens"] or 0,
+                        total_tokens = final_message.usage_metadata["total_tokens"] or 0,
+                        total_cost = 0
+                    )
+        # End LangGraph ReAct Agent workflow implementation            
+
         # Start LangChain Expression Language (LCEL) implementation
 
         # Get the vector document retriever, if it exists.
@@ -452,7 +498,7 @@ async def invoke_async(self, request: KnowledgeManagementCompletionRequest) -> C
             chain_context
             | prompt_template
             | RunnableLambda(self._record_full_prompt)
-            | self._get_language_model()
+            | llm
         )
 
         retvalue = None

src/python/PythonSDK/foundationallm/models/agents/__init__.py

@@ -1,11 +1,13 @@
 from .agent_conversation_history_settings import AgentConversationHistorySettings
 from .agent_gatekeeper_settings import AgentGatekeeperSettings
 from .agent_orchestration_settings import AgentOrchestrationSettings
+from .agent_tool import AgentTool
 from .agent_vectorization_settings import AgentVectorizationSettings
 from .agent_workflows.agent_workflow_ai_model import AgentWorkflowAIModel
 from .agent_workflows.agent_workflow_base import AgentWorkflowBase
 from .agent_workflows.azure_openai_assistants_agent_workflow import AzureOpenAIAssistantsAgentWorkflow
 from .agent_workflows.langchain_expression_language_agent_workflow import LangChainExpressionLanguageAgentWorkflow
+from .agent_workflows.langgraph_react_agent_workflow import LangGraphReactAgentWorkflow
 from .agent_base import AgentBase
 from .knowledge_management_agent import KnowledgeManagementAgent
 from .knowledge_management_completion_request import KnowledgeManagementCompletionRequest

src/python/PythonSDK/foundationallm/models/agents/agent_base.py

@@ -8,7 +8,8 @@
     AgentGatekeeperSettings,
     AgentOrchestrationSettings,
     AzureOpenAIAssistantsAgentWorkflow,
-    LangChainExpressionLanguageAgentWorkflow
+    LangChainExpressionLanguageAgentWorkflow,
+    LangGraphReactAgentWorkflow
 )
 from foundationallm.models.resource_providers import ResourceBase
 
@@ -24,7 +25,7 @@ class AgentBase(ResourceBase):
     tools: Optional[dict] = Field(default=[], description="A dictionary object with assigned agent tools.")
     workflow: Optional[
         Annotated [
-            Union[AzureOpenAIAssistantsAgentWorkflow, LangChainExpressionLanguageAgentWorkflow],
+            Union[AzureOpenAIAssistantsAgentWorkflow, LangChainExpressionLanguageAgentWorkflow, LangGraphReactAgentWorkflow],
             Field(discriminator='type')
         ]
     ]= Field(default=None, description="The workflow configuration for the agent.")

src/python/PythonSDK/foundationallm/models/agents/agent_workflows/langchain_expression_language_agent_workflow.py

@@ -1,4 +1,3 @@
-from pydantic import Field
 from typing import Any, Self, Literal
 from foundationallm.langchain.exceptions import LangChainException
 from foundationallm.utils import object_utils

src/python/PythonSDK/foundationallm/models/agents/agent_workflows/langgraph_react_agent_workflow.py

@@ -0,0 +1,25 @@
+from typing import Any, Self, Literal
+from foundationallm.langchain.exceptions import LangChainException
+from foundationallm.utils import object_utils
+from .agent_workflow_base import AgentWorkflowBase
+
+class LangGraphReactAgentWorkflow(AgentWorkflowBase):
+    """
+    The configuration for a LangGraph ReAct agent workflow.
+    """
+    type: Literal["langgraph-react-agent-workflow"] = "langgraph-react-agent-workflow"    
+
+    @staticmethod
+    def from_object(obj: Any) -> Self:
+
+        workflow: LangGraphReactAgentWorkflow = None
+
+        try:
+            workflow = LangGraphReactAgentWorkflow(**object_utils.translate_keys(obj))
+        except Exception as e:
+            raise LangChainException(f"The LangGraph ReAct Agent Workflow object provided is invalid. {str(e)}", 400)
+
+        if workflow is None:
+            raise LangChainException("The LangGraph ReAct Agent Workflow object provided is invalid.", 400)
+
+        return workflow

src/python/PythonSDK/requirements.txt

@@ -10,6 +10,7 @@ langchain==0.3.7
 langchain-aws==0.2.7
 langchain-experimental==0.3.3
 langchain-openai==0.2.9
+langgraph==0.2.53
 openai==1.55.0
 opentelemetry-api==1.27.0
 opentelemetry-sdk==1.27.0