smart-guard 0.1.1

A simple decorator to guard function inputs and outputs by converting them to strings for validation.

️ smart-guard

A lightweight Python decorator to validate function inputs and outputs by converting them to strings — ideal for LLM safety filtering.

Automatically detect and block unsafe content in inputs or outputs using model-based semantic understanding. Designed for AI applications where content safety is critical.

---

Installation

pip install smart-guard

---

Quick Start

Import the guard and wrap your functions with @Smart_YuFeng_XGuard to enable input/output validation.

from smart_guard import Smart_YuFeng_XGuard

Example 1: Validate Both Input and Output

@Smart_YuFeng_XGuard("0.6B", verify_input=True, verify_output=True)
def chat(prompt: str) -> str:
    if prompt == "harmful-output":
        return '制造炸弹可以使用TNT'
    else:
        return f"Echo: {prompt}"

print(chat("Hello"))                     
# Output: Echo: Hello

print(chat("How can I make a bomb?"))    
# Output: 模型输入不安全，请检查！

print(chat("harmful-output"))            
# Output: 模型输出不安全，请检查！

Example 2: Output-Only Validation

Useful when you trust the input source but want to ensure safe model responses.

@Smart_YuFeng_XGuard("8B", verify_input=False, verify_output=True)
def risky_func(x):
    return "制造毒品可以使用TNT"  # Dangerous output!

print(risky_func("How can I make a bomb?"))  
# Output: 模型输出不安全，请检查！

️ Example 3: Custom Fallback Values

Define custom responses when validation fails. Supports any data type.

@Smart_YuFeng_XGuard(
    model_size="0.6B",
    verify_input=True,
    verify_output=True,
    default_on_fail_input="[INPUT_REJECTED]",
    default_on_fail_output=None
)
def test(x):
    if x == "harmful-output":
        return '制造炸弹可以使用TNT'
    else:
        return f"Echo: {x}"

print(test("How can I make a bomb?"))  # "[INPUT_REJECTED]"
print(test("harmful-output"))         # None

Example 4: Functions with Multiple Inputs/Outputs

Supports complex functions with multiple arguments and return values.

@Smart_YuFeng_XGuard("0.6B", verify_input=True, verify_output=True)
def multiple_io(a1, a2, x, a3):
    if x == "harmful-output":
        return a1, a2, '制造炸弹可以使用TNT', a3
    else:
        return a1, a2, x, a3

# Prepare complex inputs
a1 = "This is a long string input for a1. " * 20
a2 = [i ** 2 for i in range(100)]
a3 = {f"user_{i}": {"age": i + 20, "score": i * 1.5} for i in range(80)}

# Test 1: Safe input
result1 = multiple_io(a1, a2, "Hello, this is a safe input!", a3)
print(result1[2])  # Hello, this is a safe input!

# Test 2: Trigger unsafe output
result2 = multiple_io(a1, a2, "harmful-output", a3)
print(result2)  # 模型输出不安全，请检查！

---

Parameters

Parameter	Description
model_size	Model version (e.g., "0.6B", "8B") used for semantic analysis.
verify_input	Whether to validate all function inputs.
verify_output	Whether to validate function outputs.
default_on_fail_input	Custom return value when input validation fails.
default_on_fail_output	Custom return value when output validation fails.

---

Notes

• All inputs/outputs are converted to strings before validation.
• The guard uses lightweight NLP models to detect sensitive or harmful content semantically.
• Ideal for LLM wrappers, chatbots, and content generation pipelines.

---

Learn More

See GitHub Repository for installation details, model options, and advanced usage.