Bias Detection

Note

This capability is provided by the “Bias Detection Guardrail” plugin, which you need to install. Please see Installing plugins.

This plugin is Not supported

The Bias Detection guardrail check that questions to a LLM or responses do not exhibit bias or stereotyped statements, notably racial, religious, gender-based, sexuality-based or disability-based.

The Bias Detection guardrail first uses a small specialized model to provide a first level of filtering. This small specialized model however has a significant false positive rate. Optionally, the Bias Detection guardrail can then use an auxiliary LLM to confirm or infirm the initial assumption (“LLM as a judge” approach).

It can be configured to either:

• block the request (with an error)

• simply audit the biased conversation

• ask the guarded LLM to rewrite its answer