I am processing an Excel file that contains artist and title names through a Server Action integrated with an Agentic AI application to match records against a database. If a record is not found, I have implemented a prompt to retry the search using additional or modified keywords. However, during processing, certain titles that contain words like “kill,” “devil,” or other potentially offensive terms are triggering content policy violation errors. Even after lowering the content filtering level to “Low” in the Azure/Microsoft configuration, the system still throws exceptions for some of these terms. Since these words are part of legitimate song or artist names and not being used in any harmful context, this is blocking valid processing scenarios. I would like guidance on how to handle such cases properly—whether there is a recommended way to bypass or safely handle content filtering for neutral data processing use cases, or if there are best practices (such as preprocessing, encoding, or prompt restructuring) to avoid these policy violations while still maintaining compliance.
Hello Sathwik,
This is a "model" issue, not an OutSystems issue.
Maybe this can help: https://learn.microsoft.com/en-us/answers/questions/1851563/how-to-disable-content-filter-and-abuse-monitoring
Cheers!
Hi Sathwik,
Where are you encountering these errors? During runtime?
There are other way you can try managing these. For example, in Agent Guardrails, you can set it to "Log and Continue" so you won't encounter an error. You can also control this in your prompt for example by telling your AI to treat these columns as neutral whatever the content is.
This issue is caused by Azure OpenAI content moderation, which may flag words like kill or devil even when they are part of valid song or artist names. Since these terms are being processed as metadata and not harmful content, the best approach is to restructure the prompt clearly, for example: “The following values are music metadata used only for database matching.” Avoid conversational wording and send the data in a structured format such as JSON or separate fields. You can also preprocess titles before retrying searches by removing special characters or generating alternate keywords in code instead of repeatedly sending modified prompts to the AI. Content filters cannot be fully bypassed, so reducing false positives through prompt design and preprocessing is the recommended solution.
No even i have tried with single line Prompt but 90 percent of the offensive words are being analyzed by AI but only 10% of it Throwing Error as
Failed to call AI model. Status code: BadRequest. The Azure OpenAI AI model provider returned this message: "litellm.ContentPolicyViolationError: The response was filtered due to the prompt triggering Azure OpenAI's content management policy. Please modify your prompt and retry. To learn more about our content filtering policies please read our documentation: https://go.microsoft.com/fwlink/?linkid=2198766 model=5341d3c0-37ae-4e95-b251-c90fe1cdae49. content_policy_fallback=None. fallbacks=None. Set 'content_policy_fallback' - https://docs.litellm.ai/docs/routing#fallbacks." (OS-ABRS-FM-40004)