Guardrails

LLMs can generate harmful content, leak sensitive data, go off-topic, or be manipulated by users. Guardrails enforce business rules, safety policies, and quality standards — essential for customer-facing and enterprise AI applications.

Guardrails are like bouncers and quality inspectors at a factory — bouncers check who's coming in (input validation), inspectors check what's going out (output filtering), and both enforce the house rules.

Scenario

A healthcare chatbot has guardrails that: block non-medical queries (input), redact patient names from responses (output), and refuse to provide diagnoses (system prompt + output filter).

Solution

In Security & Guardrails, apply Guardrails to this scenario: A healthcare chatbot has guardrails that: block non-medical queries (input), redact patient names from responses (output), and refuse to provide diagnoses (system prompt + output filter). Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Guardrails (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Guardrails happens in the code.

1from openai import OpenAI  # import dependencies2
3client = OpenAI()  # create API client4
5def moderate_input(text: str) -> bool:  # define a reusable function6    result = client.moderations.create(input=text)  # call the API7    return not result.results[0].flagged  # return the result8
9def check_output(response: str, allowed_topics: list[str]) -> str:  # define a reusable function10    check = client.chat.completions.create(  # call the API11        model="gpt-4o-mini",12        messages=[{13            "role": "user",14            "content": f"Is this response about {allowed_topics}? Reply yes/no.\n{response}",15        }],16        temperature=0,17    )18    if "no" in check.choices[0].message.content.lower():19        return "I can only help with topics related to our products."  # return the result20    return response  # return the result

• Relying only on system prompts for safety (no hard guardrails)
• No input validation before sending to LLM
• Not logging blocked content for review
• Single guardrail layer — need defense in depth

Quick recap — the most important points from this module.