Generative AI

A spam filter decides. A model that writes produces. That sentence is the distinction, and the engineering consequences of it run through everything else in this course.

A classifier picks from a set of options you defined. Four categories in, one of four categories out. A generative model produces new content one piece at a time, and the set of things it could produce is effectively unlimited. You cannot enumerate the possible outputs, which means you cannot test them all, and you cannot assume the shape of what comes back.

CATEGORIES = ["travel", "meals", "software", "equipment"]

def categorise(description: str) -> str:
    """Closed output. The return value is always one of four strings."""
    return classifier.predict([description])[0]

def draft_rejection_note(claim: dict, reason: str) -> str:
    """Open output. The return value is text that did not exist before."""
    response = client.chat.completions.create(
        model="gpt-4.1-mini",
        messages=[
            {
                "role": "system",
                "content": (
                    "You write short, plain, courteous notes to employees about "
                    "expense claims. Two sentences. State the reason given. "
                    "Do not invent policy."
                ),
            },
            {"role": "user", "content": f"Claim: {claim['description']}. Reason: {reason}"},
        ],
    )
    return response.choices[0].message.content

• categorise has a return type you can reason about. Four values are possible, so a test can cover the behaviour and a bug is visible as the wrong label.
• draft_rejection_note returns a string, and that is all its type tells you. It might be two sentences as asked. It might be five. It might contain a policy detail nobody supplied.
• The system message is where you constrain a generative call — length, tone, and an explicit instruction not to invent policy. Constraints in words are guidance, not enforcement, which is why later lessons validate the result in code.
• The full mechanics of this call — the client, the key, timeouts, retries — come in "Your First Model Integration". For now, notice only the shape: instructions plus data in, free text out.