guardrails-ai-restrict-to-topic 0.1.0

Checks that a text stays on a set of valid topics and avoids invalid ones.

guardrails-ai-restrict-to-topic

Checks that a text's topic is within a set of valid topics and, optionally, that it avoids a set of invalid topics. By default it runs a local zero-shot classifier (facebook/bart-large-mnli) and falls back to an LLM when the classifier is not confident.

Registered as tryolabs/restricttotopic.

Installation

pip install guardrails-ai-restrict-to-topic

Usage

from guardrails import Guard
from guardrails_ai.restricttotopic import RestrictToTopic

guard = Guard().use(
    RestrictToTopic(
        valid_topics=["sports", "politics"],
        invalid_topics=["music"],
        disable_llm=True,  # zero-shot classifier only (no OpenAI key needed)
        on_fail="exception",
    )
)

guard.validate("The Chiefs won the Super Bowl.")  # passes

The default ensemble mode also uses an LLM fallback and requires OPENAI_API_KEY.

License

Apache-2.0. Originally authored by Tryolabs.