Guardrail set up in Foundry that rates prompts and completions for risks such as hate, violence, sexual content, self-harm, prompt attacks and protected material, blocking anything at or above a chosen severity.
Also called guardrail controls.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Content filter in context, with comparison tables and the common traps.
Terms in this definition
- Set
Secret permission in Key Vault for writing secrets; some older material refers to it as Create.
Related terms
- Blocklist
A content filter can reference this user-defined collection of terms or regular expressions, holding as many as 10,000 entries, to stop particular words getting through.
- finish_reason
Explains why a chat completion ended. Values are stop (completed normally), length (truncated by max_tokens or the token limit), content_filter (content omitted) and tool_calls (a tool was invoked).
- Severity threshold
For each harm category, the content filter level (medium by default) from which content gets blocked. The strictest option is blocking at low for both input and output.