Create services/graph_service.py

services/graph_service.py

@@ -0,0 +1,99 @@
+import os
+from typing import TypedDict, Annotated, Sequence
+from langchain_openai import ChatOpenAI
+from langchain_core.messages import BaseMessage, AIMessage, HumanMessage
+from langchain.agents import create_openai_tools_agent, AgentExecutor
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+from langgraph.graph import StateGraph, END
+from tools.analysis_tools import trigger_interview_analysis
+
+# --- 1. Définition de l'état du graphe ---
+class AgentState(TypedDict):
+    user_id: str
+    job_offer_id: str
+    cv_document: dict
+    job_offer: dict
+    messages: Annotated[Sequence[BaseMessage], lambda x, y: x + y]
+
+# --- 2. Configuration de l'agent ---
+def create_agent_executor():
+    # Le prompt instruit l'agent sur son rôle et quand utiliser l'outil
+    prompt = ChatPromptTemplate.from_messages([
+        ("system", """You are a friendly and professional HR recruiter named AIrh.
+        Your role is to conduct a job interview based on the provided resume (cv_document) and the job description (job_offer).
+        Engage with the user, ask relevant questions, and respond to their answers.
+        When you feel the interview is complete, you MUST use the 'trigger_interview_analysis' tool to end the conversation.
+        After calling the tool, say a final goodbye to the user."""),
+        MessagesPlaceholder(variable_name="messages"),
+    ])
+    
+    llm = ChatOpenAI(api_key=os.getenv("OPENAI_API_KEY"), model="gpt-4-turbo", temperature=0.7)
+    tools = [trigger_interview_analysis]
+    agent = create_openai_tools_agent(llm, tools, prompt)
+    return AgentExecutor(agent=agent, tools=tools, verbose=True)
+
+# --- 3. Définition des nœuds du graphe ---
+agent_executor = create_agent_executor()
+
+def agent_node(state: AgentState):
+    """Le nœud principal qui appelle l'agent pour répondre ou utiliser un outil."""
+    response = agent_executor.invoke({
+        "user_id": state["user_id"],
+        "job_offer_id": state["job_offer_id"],
+        "cv_document": state["cv_document"],
+        "job_offer": state["job_offer"],
+        "messages": state["messages"],
+        "conversation_history": state["messages"] # Pour la compatibilité si l'outil en a besoin
+    })
+    return {"messages": [AIMessage(content=response["output"])]}
+
+# --- 4. Définition du routeur (Conditional Edge) ---
+def router(state: AgentState):
+    """Décide du chemin à suivre après la réponse de l'agent."""
+    last_message = state["messages"][-1]
+    if hasattr(last_message, 'tool_calls') and last_message.tool_calls:
+        # Si l'agent a décidé d'appeler l'outil, on termine le graphe
+        return "end"
+    else:
+        # Sinon, on continue la conversation
+        return "continue"
+
+# --- 5. Construction du graphe ---
+def build_graph():
+    graph = StateGraph(AgentState)
+    graph.add_node("agent", agent_node)
+    graph.add_conditional_edges(
+        "agent",
+        router,
+        {"continue": "agent", "end": END}
+    )
+    graph.set_entry_point("agent")
+    return graph.compile()
+
+# --- Service principal à appeler depuis l'API ---
+class GraphConversationManager:
+    def __init__(self):
+        self.graph = build_graph()
+
+    def invoke(self, payload: dict):
+        # Prépare les messages pour le format LangChain
+        messages = [HumanMessage(content=m["content"]) if m["role"] == "user" else AIMessage(content=m["content"]) for m in payload["messages"]]
+        
+        state = {
+            "user_id": payload["user_id"],
+            "job_offer_id": payload["job_offer_id"],
+            "cv_document": payload["cv_document"],
+            "job_offer": payload["job_offer"],
+            "messages": messages,
+        }
+        
+        final_state = self.graph.invoke(state)
+        
+        # Détermine le statut final pour le front-end
+        last_message = final_state['messages'][-1]
+        status = "finished" if hasattr(last_message, 'tool_calls') and last_message.tool_calls else "interviewing"
+        
+        return {
+            "response": last_message.content,
+            "status": status
+        }