app.py

import os
from dotenv import load_dotenv
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from typing import TypedDict, Annotated, Literal
from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
import gradio as gr

load_dotenv()

# Configuración
max_tokens = 2000
num_iterations = 2
quality_threshold = 8

# Base de datos simulada de destinos turísticos
travel_database = {
    "paris": {"destination": "Paris", "price": 1500, "features": ["romantic", "cultural", "historic"]},
    "bali": {"destination": "Bali", "price": 1200, "features": ["beach", "relaxing", "adventurous"]},
    "new_york": {"destination": "New York", "price": 2000, "features": ["urban", "shopping", "nightlife"]},
    "tokyo": {"destination": "Tokyo", "price": 1800, "features": ["modern", "cultural", "tech-savvy"]},
}

# Modelos estructurados para la salida de cada nodo
class GenerateRecommendation(BaseModel):
    destination: str = Field(description="El destino turístico recomendado")
    explanation: str = Field(description="Explicación breve de la recomendación")

class RecommendationQualityScore(BaseModel):
    score: int = Field(description="Puntuación de la recomendación entre 1-10")
    comment: str = Field(description="Comentario sobre la calidad de la recomendación")

# Estado del grafo
class GraphState(TypedDict):
    messages: Annotated[list, add_messages]
    quality: Annotated[int, 0]  # Valor inicial 0
    iterations: Annotated[int, 0]  # Valor inicial 0

# Inicialización del grafo
builder = StateGraph(GraphState)

llm = ChatOpenAI(
    model="gpt-4o-mini",
    temperature=0,
    max_tokens=max_tokens,
    api_key=os.getenv("OPENAI_API_KEY")
)
developer_structure_llm = llm.with_structured_output(GenerateRecommendation, method="json_mode")
reviewer_structure_llm = llm.with_structured_output(RecommendationQualityScore, method="json_mode")

def travel_recommender(state):
    # Se asume que el último mensaje contiene las preferencias del usuario
    user_requirements = state["messages"][-1].content
    system_prompt = f"""
    Eres un experto en recomendaciones de viajes.
    Con base en las siguientes preferencias del usuario: {user_requirements},
    selecciona el mejor destino de la siguiente base de datos: {travel_database}.
    Responde en JSON con la clave `destination` para el destino recomendado y `explanation` con una breve razón de la recomendación.
    """
    human_messages = [msg for msg in state["messages"] if isinstance(msg, HumanMessage)]
    ai_messages = [msg for msg in state["messages"] if isinstance(msg, AIMessage)]
    system_messages = [SystemMessage(content=system_prompt)]
    
    messages = system_messages + human_messages + ai_messages
    message = developer_structure_llm.invoke(messages)
    
    recommendation_output = f"Destination: {message.destination}\nExplanation: {message.explanation}"
    # Se agrega la recomendación a los mensajes para los siguientes nodos
    state["messages"].append(AIMessage(content=recommendation_output))
    state["iterations"] += 1
    return state

def recommendation_review(state):
    system_prompt = """
    Eres un revisor de recomendaciones con altos estándares.
    Revisa la recomendación proporcionada y asigna una puntuación de calidad entre 1-10.
    Evalúa la relevancia, precisión y alineación con las necesidades del cliente.
    Responde en JSON con las claves `score` y `comment`.
    """
    human_messages = [msg for msg in state["messages"] if isinstance(msg, HumanMessage)]
    ai_messages = [msg for msg in state["messages"] if isinstance(msg, AIMessage)]
    system_messages = [SystemMessage(content=system_prompt)]
    
    messages = system_messages + human_messages + ai_messages
    message = reviewer_structure_llm.invoke(messages)
    
    review_comment = f"Review Score: {message.score}\nComment: {message.comment}"
    state["messages"].append(AIMessage(content=review_comment))
    state["quality"] = message.score
    return state

def final_recommendation(state):
    system_prompt = "Revisa la recomendación final y proporciona una respuesta final para el usuario."
    human_messages = [msg for msg in state["messages"] if isinstance(msg, HumanMessage)]
    ai_messages = [msg for msg in state["messages"] if isinstance(msg, AIMessage)]
    system_messages = [SystemMessage(content=system_prompt)]
    
    messages = system_messages + human_messages + ai_messages
    final_message = llm.invoke(messages)
    
    # Se guarda la recomendación final en el estado para mostrarla
    state["final_recommendation"] = final_message.content
    state["messages"].append(AIMessage(content=f"Final Recommendation: {final_message.content}"))
    return state

# Función para definir la condición de la bifurcación en el grafo
def quality_gate_condition(state) -> Literal["travel_recommender", "final_recommendation"]:
    if state["iterations"] >= num_iterations:
        return "final_recommendation"
    if state["quality"] < quality_threshold:
        return "travel_recommender"
    else:
        return "final_recommendation"

# Agregar nodos al grafo
builder.add_node("travel_recommender", travel_recommender)
builder.add_node("recommendation_review", recommendation_review)
builder.add_node("final_recommendation", final_recommendation)

# Conectar nodos
builder.add_edge(START, "travel_recommender")
builder.add_edge("travel_recommender", "recommendation_review")
builder.add_edge("final_recommendation", END)

builder.add_conditional_edges("recommendation_review", quality_gate_condition)

graph = builder.compile()

# Función que ejecuta el grafo con la entrada del usuario
def run_graph(user_input: str) -> str:
    initial_state = {"messages": [HumanMessage(content=user_input)], "quality": 0, "iterations": 0}
    final_state = graph.invoke(initial_state)
    return final_state.get("final_recommendation", "No se generó una recomendación final.")

# Interfaz de Gradio
iface = gr.Interface(
    fn=run_graph,
    inputs=gr.Textbox(label="Ingrese sus preferencias de viaje"),
    outputs=gr.Textbox(label="Recomendación Final"),
    title="Sistema de Recomendación de Viajes"
)

if __name__ == "__main__":
    iface.launch()