Setting the Granite Guardian input moderation threshold
You can enable, disable, or adjust the sensitivity of Granite Guardian input moderation by setting the GuardianInputModerationThreshold value in the Gen AI Advanced Options field.
About this task
Granite Guardian input moderation is enabled by default with a threshold value of 0.5. The threshold determines the minimum confidence level required before a query is flagged and blocked. The score threshold ranges from 0.0 to 1.0:
- A lower value, such as 0.1 or 0.2, applies stricter filtering. The system blocks queries even when it is only moderately confident they are unsafe. Harmful content is more likely to be detected, but safe content might occasionally be flagged.
- A higher value, such as 0.8 or 0.9, applies more lenient filtering. The system only blocks queries it is highly confident about, letting borderline cases through. Content that is flagged is more likely to be genuinely harmful, but some harmful content might be missed.
- Setting the value to 0.0 disables Granite Guardian filtering. User queries are still submitted for moderation, but no query receives a score that meets the threshold, so no queries are blocked.
You set the threshold by editing the GenAI Configuration object for the object store and specifying the value as an advanced configuration option.
Procedure
Results
Granite Guardian moderation is updated for the object store. The system applies the new threshold to all subsequent user queries.