Añadir instrucciones al modelo

Estás creando un sistema de atención al cliente con IA para una empresa de servicios financieros. Por motivos de seguridad y cumplimiento, la IA debe tener límites claros para responder únicamente a preguntas sobre saldos de cuentas e historial de transacciones. Debe rechazar educadamente cualquier otra solicitud, como restablecimientos de contraseña.

El cliente de OpenAI ya está disponible como client. Lo tendrás disponible durante el resto del curso.

Este ejercicio forma parte del curso

Trabajar con la API Responses de OpenAI

Instrucciones del ejercicio

Crea una solicitud a la Responses API para "gpt-5-mini" que incluya instrucciones claras para responder solo a preguntas sobre saldos de cuentas e historial de transacciones, y no a otras solicitudes como restablecimientos de contraseña.

Ejercicio interactivo práctico

Prueba este ejercicio y completa el código de muestra.

# Create a guardrailed AI response
response = client.responses.create(
    model=____,
    instructions="You are a customer support AI that only answers questions about ____. Politely decline any other requests such as ____.",
    input="Can you help me reset my password?"
)

print(response.output_text)

Editar y ejecutar código

Este ejercicio forma parte del curso

Trabajar con la API Responses de OpenAI

IntermedioNivel de habilidad

4.8+

Comienza el curso gratis

Building AI applications has never been easier with the Responses API! Learn how to prompt models using a simple interface, best practices for working with reasoning models, and how to build back-and-forth conversations with LLMs.

Exercise 1: La Responses API: la mejor aliada de cualquier desarrollador Exercise 2: Tu primera llamada a la Responses API Exercise 3: Añadir instrucciones al modelo

Ejercicio actual

Exercise 4: Extrayendo información de la respuesta Exercise 5: Domina los parámetros de respuesta Exercise 6: Experimentar con modelos más potentes Exercise 7: Razonar sobre el razonamiento Exercise 8: Encadenado de conversaciones Exercise 9: De un mensaje a otro Exercise 10: Habilitar conversaciones de chat

LLMs are powerful, but they're even more powerful once tools are introduced! Learn how to integrate web search with your LLM systems with no additional API keys or credentials. Discover that tools can be, well... pretty much anything! Build your own tools for LLMs to call to go beyond the model's base capabilities.

Exercise 1: Response Items: The Building Blocks Exercise 2: Extracting Output Items and Their Contents Exercise 3: Using Items for Custom Handling Exercise 4: Web Search That Just Works Exercise 5: Combining LLMs and Web Search Exercise 6: Building Confidence with LLM Sources Exercise 7: Function-Calling Tools: Defining the Function Exercise 8: The Function-Calling Tool Workflow Exercise 9: Defining a Function for Converting Timezones Exercise 10: Function-Calling Tools: The Full Workflow Exercise 11: Building the OpenAI Tool Definition Exercise 12: Integrating Function-Calling Tools

Move your LLM systems from proof-of-concept to production with three key design paradigms. Learn how to generate consistent and reliable structured outputs, like JSON, using LLMs. You'll also use semantic events to stream updates to your users, creating interactive and enjoyable experiences. Finally, go multi-modal and learn how to combine text and images with OpenAI's GPT-5 models.

Exercise 1: Generating Structured Outputs Exercise 2: Consistent Outputs, Every Time!Exercise 3: Nesting Pydantic Classes Exercise 4: Streaming with Semantic Events Exercise 5: Streaming OpenAI Responses Exercise 6: Streaming Semantic Events Exercise 7: Streaming with Function Calls Exercise 8: Input Flexibility and Multimodality Exercise 9: Role-Based Inputs Exercise 10: Prompting with Images from URLs Exercise 11: Prompting with Local Image Files Exercise 12: Congratulations!