Implementation:Protectai Llm guard Output BanTopics

Knowledge Sources	Protectai_Llm_guard
Domains	Content_Filtering, Zero_Shot_Classification
Last Updated	2026-02-14 12:00 GMT

Overview

BanTopics is an output scanner that detects and blocks LLM responses covering banned topics using zero-shot classification, delegating to the input-side InputBanTopics scanner.

Description

The BanTopics output scanner is a thin wrapper around the corresponding input scanner InputBanTopics. It uses a zero-shot classification model to determine whether the LLM output falls under any of the specified banned topics. The scanner accepts a list of topic labels and a threshold parameter that controls the minimum confidence score required for a topic to be considered a match. When the model classifies the output as belonging to a banned topic with confidence above the threshold, the output is flagged as invalid. This approach does not require topic-specific training data, making it flexible for a wide range of content policies.

Usage

Use this scanner when you need to enforce topic-based content policies on LLM outputs. Common scenarios include preventing discussions of violence, politics, religion, adult content, or any other topics that violate your application's content guidelines. The zero-shot approach allows you to add new banned topics without retraining any models.

Code Reference

Source Location

• Repository: Protectai_Llm_guard
• File: llm_guard/output_scanners/ban_topics.py
• Lines: 1-45

Signature

class BanTopics(Scanner):
    def __init__(
        self,
        topics: list[str],
        *,
        threshold: float = 0.75,
        model: Model | None = None,
        use_onnx: bool = False,
    ) -> None: ...

    def scan(self, prompt: str, output: str) -> tuple[str, bool, float]: ...

Import

from llm_guard.output_scanners import BanTopics

I/O Contract

Inputs

Name	Type	Required	Description
prompt	str	Yes	The input prompt
output	str	Yes	The LLM output to scan for banned topics

Constructor Parameters

Name	Type	Required	Default	Description
topics	list[str]	Yes	N/A	List of topic labels to ban
threshold	float	No	0.75	Minimum classification confidence to trigger detection
model	None	No	None	Custom zero-shot classification model
use_onnx	bool	No	False	Whether to use ONNX runtime for inference

Outputs

Name	Type	Description
sanitized_output	str	The output (potentially modified)
is_valid	bool	Whether the output passed the scan (True if no banned topics detected)
risk_score	float	Risk score (-1.0 to 1.0)

Usage Examples

Basic Usage

from llm_guard.output_scanners import BanTopics

scanner = BanTopics(
    topics=["violence", "politics", "religion"],
    threshold=0.75,
)

prompt = "Tell me about world history"
output = "The French Revolution was a period of major political upheaval in France."

sanitized_output, is_valid, risk_score = scanner.scan(prompt, output)

if not is_valid:
    print(f"Banned topic detected (risk: {risk_score})")
else:
    print("Output topic is acceptable")

Related Pages

• Principle:Protectai_Llm_guard_Topic_Filtering