Este artículo demuestra cómo construir un sistema de enrutamiento inteligente impulsado por Modelos Claude de Anthrope. Este sistema mejora la eficiencia y la calidad de la respuesta al clasificar automáticamente las solicitudes de los usuarios y dirigirlas a manejadores especializados. El flujo de trabajo analiza las consultas entrantes, determina su intención y las enruta a las tuberías de procesamiento apropiadas, ya sea para la atención al cliente, la asistencia técnica u otras respuestas específicas del dominio.
Paso 1: Instale los paquetes de Python requeridos
!pip install anthropic pandas scikit-learnPaso 2: Importar las bibliotecas necesarias para el proyecto
import os
import json
import time
import pandas as pd
import numpy as np
from anthropic import Anthropic
from IPython.display import display, Markdown
from sklearn.metrics import classification_reportPaso 3: Configure la autenticación de la API antrópica definiendo su clave API e inicializando el cliente antrópico
ANTHROPIC_API_KEY = "{Your API KEY}"
client = Anthropic(api_key=ANTHROPIC_API_KEY)Paso 4: Cree un conjunto de datos de muestra de consultas de clientes con categorías asociadas para capacitar y probar el sistema de enrutamiento.
customer_queries = [
{"id": 1, "query": "What are your business hours?", "category": "General Question"},
{"id": 2, "query": "How do I reset my password?", "category": "Technical Support"},
{"id": 3, "query": "I want a refund for my purchase.", "category": "Refund Request"},
{"id": 4, "query": "Where can I find your privacy policy?", "category": "General Question"},
{"id": 5, "query": "The app keeps crashing when I try to upload photos.", "category": "Technical Support"},
{"id": 6, "query": "I ordered the wrong size, can I get my money back?", "category": "Refund Request"},
{"id": 7, "query": "Do you ship internationally?", "category": "General Question"},
{"id": 8, "query": "My account is showing incorrect information.", "category": "Technical Support"},
{"id": 9, "query": "I was charged twice for my order.", "category": "Refund Request"},
{"id": 10, "query": "What payment methods do you accept?", "category": "General Question"}
]
Paso 5: Convierta la lista de consultas de clientes en un marco de datos PANDAS para una manipulación y análisis más fácil. Luego, muestre el marco de datos en el cuaderno para visualizar la estructura del conjunto de datos de capacitación.
df = pd.DataFrame(customer_queries)
display(df)Paso 6: Defina la función de enrutamiento central que utiliza el soneto Claude 3.7 para clasificar las consultas de los clientes en categorías predefinidas.
def route_query(query, client):
"""
Route a customer query to the appropriate category using Claude 3.5 Haiku.
Args:
query (str): The customer query to classify
client: Anthropic client
Returns:
str: The classified category
"""
system_prompt = """
You are a query classifier for a customer service system.
Your job is to categorize customer queries into exactly one of these categories:
1. General Question - Basic inquiries about the company, products, policies, etc.
2. Refund Request - Any query related to refunds, returns, or billing issues
3. Technical Support - Questions about technical problems, bugs, or how to use products
Respond with ONLY the category name, nothing else.
"""
try:
response = client.messages.create(
model="claude-3-7-sonnet-20250219",
max_tokens=1024,
system=system_prompt,
messages=[{"role": "user", "content": query}]
)
category = response.content[0].text.strip()
valid_categories = ["General Question", "Refund Request", "Technical Support"]
for valid_cat in valid_categories:
if valid_cat.lower() in category.lower():
return valid_cat
return "General Question"
except Exception as e:
print(f"Error in routing: {e}")
return "General Question"Paso 7: Defina tres funciones de controlador especializado para cada categoría de consulta, cada una utilizando el soneto Claude 3.5 con un mensaje del sistema específico de categoría.
def handle_general_question(query, client):
"""Handle general inquiries using Claude 3.5 Haiku."""
system_prompt = """
You are a customer service representative answering general questions about our company.
Be helpful, concise, and friendly. Provide direct answers to customer queries.
"""
try:
response = client.messages.create(
model="claude-3-7-sonnet-20250219",
max_tokens=1024,
system=system_prompt,
messages=[{"role": "user", "content": query}]
)
return response.content[0].text.strip()
except Exception as e:
print(f"Error in general question handler: {e}")
return "I apologize, but I'm having trouble processing your request. Please try again later."def handle_refund_request(query, client):
"""Handle refund requests using Claude 3.5 Sonnet for more nuanced responses."""
system_prompt = """
You are a customer service representative specializing in refunds and billing issues.
Respond to refund requests professionally and helpfully.
For any refund request, explain the refund policy clearly and provide next steps.
Be empathetic but follow company policy.
"""
try:
response = client.messages.create(
model="claude-3-7-sonnet-20250219",
max_tokens=1024,
system=system_prompt,
messages=[{"role": "user", "content": query}]
)
return response.content[0].text.strip()
except Exception as e:
print(f"Error in refund request handler: {e}")
return "I apologize, but I'm having trouble processing your refund request. Please contact our support team directly."def handle_technical_support(query, client):
"""Handle technical support queries using Claude 3.5 Sonnet for more detailed technical responses."""
system_prompt = """
You are a technical support specialist.
Provide clear, step-by-step solutions to technical problems.
If you need more information to resolve an issue, specify what information you need.
Prioritize simple solutions first before suggesting complex troubleshooting.
"""
try:
response = client.messages.create(
model="claude-3-7-sonnet-20250219",
max_tokens=1024,
system=system_prompt,
messages=[{"role": "user", "content": query}]
)
return response.content[0].text.strip()
except Exception as e:
print(f"Error in technical support handler: {e}")
return "I apologize, but I'm having trouble processing your technical support request. Please try our knowledge base or contact our support team."Paso 8: Cree la función de flujo de trabajo principal que orquesta todo el proceso de enrutamiento. Esta función primero clasifica una consulta, rastrea las métricas de tiempo, la dirige al controlador especializado apropiado en función de la categoría y devuelve un diccionario de resultados integral con estadísticas de rendimiento.
def process_customer_query(query, client):
"""
Process a customer query through the complete routing workflow.
Args:
query (str): The customer query
client: Anthropic client
Returns:
dict: Information about the query processing, including category and response
"""
start_time = time.time()
category = route_query(query, client)
routing_time = time.time() - start_time
start_time = time.time()
if category == "General Question":
response = handle_general_question(query, client)
model_used = "claude-3-5-haiku-20240307"
elif category == "Refund Request":
response = handle_refund_request(query, client)
model_used = "claude-3-5-sonnet-20240620"
elif category == "Technical Support":
response = handle_technical_support(query, client)
model_used = "claude-3-5-sonnet-20240620"
else:
response = handle_general_question(query, client)
model_used = "claude-3-5-haiku-20240307"
handling_time = time.time() - start_time
total_time = routing_time + handling_time
return {
"query": query,
"routed_category": category,
"response": response,
"model_used": model_used,
"routing_time": routing_time,
"handling_time": handling_time,
"total_time": total_time
}Paso 9: Procese cada consulta en el conjunto de datos de muestra a través del flujo de trabajo de enrutamiento, recopile los resultados con categorías reales versus predichas y evalúe el rendimiento del sistema.
results = []
for _, row in df.iterrows():
query = row['query']
result = process_customer_query(query, client)
result["actual_category"] = row['category']
results.append(result)
results_df = pd.DataFrame(results)
display(results_df[["query", "actual_category", "routed_category", "model_used", "total_time"]])
accuracy = (results_df["actual_category"] == results_df["routed_category"]).mean()
print(f"Routing Accuracy: {accuracy:.2%}")
from sklearn.metrics import classification_report
print(classification_report(results_df["actual_category"], results_df["routed_category"]))Paso 10: Resultados simulados.
simulated_results = []
for _, row in df.iterrows():
query = row['query']
actual_category = row['category']
if "hours" in query.lower() or "policy" in query.lower() or "ship" in query.lower() or "payment" in query.lower():
routed_category = "General Question"
model_used = "claude-3-5-haiku-20240307"
elif "refund" in query.lower() or "money back" in query.lower() or "charged" in query.lower():
routed_category = "Refund Request"
model_used = "claude-3-5-sonnet-20240620"
else:
routed_category = "Technical Support"
model_used = "claude-3-5-sonnet-20240620"
simulated_results.append({
"query": query,
"actual_category": actual_category,
"routed_category": routed_category,
"model_used": model_used,
"routing_time": np.random.uniform(0.2, 0.5),
"handling_time": np.random.uniform(0.5, 2.0)
})
simulated_df = pd.DataFrame(simulated_results)
simulated_df["total_time"] = simulated_df["routing_time"] + simulated_df["handling_time"]
display(simulated_df[["query", "actual_category", "routed_category", "model_used", "total_time"]])
Paso 11: Calcule y muestre la precisión del sistema de enrutamiento simulado comparando categorías predichas con categorías reales.
accuracy = (simulated_df["actual_category"] == simulated_df["routed_category"]).mean()
print(f"Simulated Routing Accuracy: {accuracy:.2%}")
print(classification_report(simulated_df["actual_category"], simulated_df["routed_category"]))Paso 12: Cree una interfaz de demostración interactiva usando widgets de iPython.
from IPython.display import HTML, display, clear_output
from ipywidgets import widgets
def create_demo_interface():
query_input = widgets.Textarea(
value="",
placeholder="Enter your customer service query here...",
description='Query:',
disabled=False,
layout=widgets.Layout(width="80%", height="100px")
)
output = widgets.Output()
button = widgets.Button(
description='Process Query',
disabled=False,
button_style="primary",
tooltip='Click to process the query',
icon='check'
)
def on_button_clicked(b):
with output:
clear_output()
query = query_input.value
if not query.strip():
print("Please enter a query.")
return
if "hours" in query.lower() or "policy" in query.lower() or "ship" in query.lower() or "payment" in query.lower():
category = "General Question"
model = "claude-3-5-haiku-20240307"
response = "Our standard business hours are Monday through Friday, 9 AM to 6 PM Eastern Time. Our customer service team is available during these hours to assist you."
elif "refund" in query.lower() or "money back" in query.lower() or "charged" in query.lower():
category = "Refund Request"
model = "claude-3-5-sonnet-20240620"
response = "I understand you're looking for a refund. Our refund policy allows returns within 30 days of purchase with a valid receipt. To initiate your refund, please provide your order number and the reason for the return."
else:
category = "Technical Support"
model = "claude-3-5-sonnet-20240620"
response = "I'm sorry to hear you're experiencing technical issues. Let's troubleshoot this step by step. First, try restarting the application. If that doesn't work, please check if the app is updated to the latest version."
print(f"Routed to: {category}")
print(f"Using model: {model}")
print("\nResponse:")
print(response)
button.on_click(on_button_clicked)
return widgets.VBox([query_input, button, output])Paso 13: Implemente una función de enrutamiento avanzada que no solo clasifique consultas, sino que también proporciona puntajes de confianza y razonamiento para cada clasificación.
def advanced_route_query(query, client):
"""
An advanced routing function that includes confidence scores and fallback mechanisms.
Args:
query (str): The customer query to classify
client: Anthropic client
Returns:
dict: Classification result with category and confidence
"""
system_prompt = """
You are a query classifier for a customer service system.
Your job is to categorize customer queries into exactly one of these categories:
1. General Question - Basic inquiries about the company, products, policies, etc.
2. Refund Request - Any query related to refunds, returns, or billing issues
3. Technical Support - Questions about technical problems, bugs, or how to use products
Respond in JSON format with:
1. "category": The most likely category
2. "confidence": A confidence score between 0 and 1
3. "reasoning": A brief explanation of your classification
Example response:
{
"category": "General Question",
"confidence": 0.85,
"reasoning": "The query asks about business hours, which is basic company information."
}
"""
try:
response = client.messages.create(
model="claude-3-5-sonnet-20240620",
max_tokens=150,
system=system_prompt,
messages=[{"role": "user", "content": query}]
)
response_text = response.content[0].text.strip()
try:
result = json.loads(response_text)
if "category" not in result or "confidence" not in result:
raise ValueError("Incomplete classification result")
return result
except json.JSONDecodeError:
print("Failed to parse JSON response. Using simple classification.")
if "general" in response_text.lower():
return {"category": "General Question", "confidence": 0.6, "reasoning": "Fallback classification"}
elif "refund" in response_text.lower():
return {"category": "Refund Request", "confidence": 0.6, "reasoning": "Fallback classification"}
else:
return {"category": "Technical Support", "confidence": 0.6, "reasoning": "Fallback classification"}
except Exception as e:
print(f"Error in advanced routing: {e}")
return {"category": "General Question", "confidence": 0.3, "reasoning": "Error fallback"}Paso 14: Cree un flujo de trabajo de procesamiento de consultas mejorado con un enrutamiento basado en la confianza que aumente las consultas de baja confianza al manejo especializado, incorporando clasificación simulada para fines de demostración.
def advanced_process_customer_query(query, client, confidence_threshold=0.7):
"""
Process a customer query with confidence-based routing.
Args:
query (str): The customer query
client: Anthropic client
confidence_threshold (float): Minimum confidence score for automated routing
Returns:
dict: Information about the query processing
"""
start_time = time.time()
if "hours" in query.lower() or "policy" in query.lower() or "ship" in query.lower() or "payment" in query.lower():
classification = {
"category": "General Question",
"confidence": np.random.uniform(0.7, 0.95),
"reasoning": "Query related to business information"
}
elif "refund" in query.lower() or "money back" in query.lower() or "charged" in query.lower():
classification = {
"category": "Refund Request",
"confidence": np.random.uniform(0.7, 0.95),
"reasoning": "Query mentions refunds or billing issues"
}
elif "password" in query.lower() or "crash" in query.lower() or "account" in query.lower():
classification = {
"category": "Technical Support",
"confidence": np.random.uniform(0.7, 0.95),
"reasoning": "Query mentions technical problems"
}
else:
categories = ["General Question", "Refund Request", "Technical Support"]
classification = {
"category": np.random.choice(categories),
"confidence": np.random.uniform(0.4, 0.65),
"reasoning": "Uncertain classification"
}
routing_time = time.time() - start_time
start_time = time.time()
if classification["confidence"] >= confidence_threshold:
category = classification["category"]
if category == "General Question":
response = "SIMULATED GENERAL QUESTION RESPONSE: I'd be happy to help with your question about our business."
model_used = "claude-3-5-haiku-20240307"
elif category == "Refund Request":
response = "SIMULATED REFUND REQUEST RESPONSE: I understand you're looking for a refund. Let me help you with that process."
model_used = "claude-3-5-sonnet-20240620"
elif category == "Technical Support":
response = "SIMULATED TECHNICAL SUPPORT RESPONSE: I see you're having a technical issue. Let's troubleshoot this together."
model_used = "claude-3-5-sonnet-20240620"
else:
response = "I apologize, but I'm not sure how to categorize your request."
model_used = "claude-3-5-sonnet-20240620"
else:
response = "SIMULATED ESCALATION RESPONSE: Your query requires special attention. I'll have our advanced support system help you with this complex request."
model_used = "claude-3-5-sonnet-20240620"
category = "Escalated (Low Confidence)"
handling_time = time.time() - start_time
total_time = routing_time + handling_time
return {
"query": query,
"routed_category": classification["category"],
"confidence": classification["confidence"],
"reasoning": classification["reasoning"],
"final_category": category,
"response": response,
"model_used": model_used,
"routing_time": routing_time,
"handling_time": handling_time,
"total_time": total_time
}Paso 15: Pruebe el sistema de enrutamiento avanzado con diversas consultas de muestra.
test_queries = [
"What are your business hours?",
"I need a refund for my order #12345",
"My app keeps crashing when I try to save photos",
"I received the wrong item in my shipment",
"How do I change my shipping address?",
"I'm not sure if my payment went through",
"The product description was misleading"
]
advanced_results = []
for query in test_queries:
result = advanced_process_customer_query(query, None, 0.7)
advanced_results.append(result)
advanced_df = pd.DataFrame(advanced_results)
display(advanced_df[["query", "routed_category", "confidence", "final_category", "model_used"]])
print("\nRouting Distribution:")
print(advanced_df["final_category"].value_counts())
print(f"\nAverage Confidence: {advanced_df['confidence'].mean():.2f}")
escalated = (advanced_df["final_category"] == "Escalated (Low Confidence)").sum()
print(f"Escalated Queries: {escalated} ({escalated/len(advanced_df):.1%})")Paso 16: Defina una función de utilidad para calcular las métricas clave de rendimiento para el sistema de enrutamiento, incluidos los tiempos de procesamiento, los niveles de confianza, las tasas de escalada y las estadísticas de distribución de categorías.
def calculate_routing_metrics(results_df):
"""
Calculate key metrics for routing performance.
Args:
results_df (DataFrame): Results of routing tests
Returns:
dict: Key performance metrics
"""
metrics = {
"total_queries": len(results_df),
"avg_routing_time": results_df["routing_time"].mean(),
"avg_handling_time": results_df["handling_time"].mean(),
"avg_total_time": results_df["total_time"].mean(),
"avg_confidence": results_df["confidence"].mean(),
"escalation_rate": (results_df["final_category"] == "Escalated (Low Confidence)").mean(),
}
category_distribution = results_df["routed_category"].value_counts(normalize=True).to_dict()
metrics["category_distribution"] = category_distribution
return metricsPaso 17: Genere y muestre un informe de rendimiento integral para el sistema de enrutamiento.
metrics = calculate_routing_metrics(advanced_df)
print("Routing System Performance Metrics:")
print(f"Total Queries: {metrics['total_queries']}")
print(f"Average Routing Time: {metrics['avg_routing_time']:.3f} seconds")
print(f"Average Handling Time: {metrics['avg_handling_time']:.3f} seconds")
print(f"Average Total Time: {metrics['avg_total_time']:.3f} seconds")
print(f"Average Confidence: {metrics['avg_confidence']:.2f}")
print(f"Escalation Rate: {metrics['escalation_rate']:.1%}")
print("\nCategory Distribution:")
for category, percentage in metrics["category_distribution"].items():
print(f" {category}: {percentage:.1%}")Este Sistema de enrutamiento de solicitudes inteligentes Demuestra cómo Modelos Claude puede clasificar y manejar eficientemente las diversas consultas de los clientes. Al implementar manejadores específicos de categoría con la selección de modelos apropiada, el sistema ofrece respuestas personalizadas mientras mantiene una alta precisión. El enrutamiento basado en la confianza con rutas de escalada garantiza que consultas complejas reciban atención especializada, creando una solución de servicio al cliente robusta y escalable.
Mira el Cuaderno de colab aquí. Además, no olvides seguirnos Gorjeo.
Aquí hay una breve descripción de lo que estamos construyendo en MarkTechPost:
Asjad es consultor interno en MarktechPost. Está persiguiendo B.Tech en Ingeniería Mecánica en el Instituto de Tecnología Indio, Kharagpur. Asjad es un entusiasta de aprendizaje automático y aprendizaje profundo que siempre está investigando las aplicaciones del aprendizaje automático en la atención médica.