Restricted topics

The Restricted topics guardrail can be applied to both user messages (Input) and the model's responses (Output).

• Restricted topics (Input) - Identifies and filters out user messages that contain restricted topics, preventing the model from engaging in those discussions.
• Restricted topics (Output) - Ensures that any model-generated response does not include prohibited topics, keeping interactions appropriate and within guidelines

Example:

In the following example, we've set the restricted topic to be "UFOs":

Configuration:

Parameters required to set up a Restricted topics guardrail:

• List of Topics: Specify the topics you want to restrict from discussions.
• OpenAI-Compatible Model: An OpenAI-compatible model is used to enforce the restricted topics test.

Using the UI

Using the SDK

Create Guardrail

from superwise_api.models.application.application import ApplicationRestrictedTopicsGuard
from superwise_api.models.application.application import OpenAIModel

restricted_topics_guard = ApplicationRestrictedTopicsGuard(
    name="Rule name",
    tag= "input",
    topics=["politics", "religion", "adult_content"],
    model=OpenAIModel(version="gpt-4o", api_token="REPLACE")
)

Add guards to application

app = sw.application.put(
    str(app.id),
    additional_config=AdvancedAgentConfig(tools=[]),
    llm_model=model, 
    prompt=None, 
    name="App name",
    show_cites = True,
    guards=[restricted_topics_guard]
)