Pythonguardrails.py
Copy
1
from pydantic import BaseModel
2
from openai import OpenAI
3
 
4
client = OpenAI()
5
 
6
class SafeResponse(BaseModel):
7
    answer: str
8
    confidence: str  # "high" | "medium" | "low"
9
 
10
def input_guard(user_message: str) -> str | None:
11
    """Returns None if safe, or an error message string if flagged."""
12
    result = client.moderations.create(input=user_message)
13
    if result.results[0].flagged:
14
        return "I can't help with that request."
15
    return None
16
 
17
def safe_chat(user_message: str) -> str:
18
    # Input guard
19
    guard_result = input_guard(user_message)
20
    if guard_result:
21
        return guard_result
22
 
23
    # LLM call with structured output
24
    resp = client.beta.chat.completions.parse(
25
        model="gpt-4o-mini",
26
        messages=[
27
            {"role": "system", "content": "Answer helpfully and concisely."},
28
            {"role": "user", "content": user_message},
29
        ],
30
        response_format=SafeResponse,
31
    )
32
    parsed = resp.choices[0].message.parsed
33
    if parsed is None:
34
        return "Unable to generate a response."
35
 
36
    return f"{parsed.answer} (confidence: {parsed.confidence})"
37
 
38
# Test
39
print(safe_chat("What is machine learning?"))
40
print(safe_chat("How do I make a bomb?"))  # Should be blocked
41