Plan and manage an Azure AI solution 1.4 Implement responsible AI across generative AI and agentic systems
An internal creative-writing tool streams answers from an Azure OpenAI deployment, and…
Study guide skill: Configure safety filters, guardrails, risk detection, and content moderation
Question B8-010
An internal creative-writing tool streams answers from an Azure OpenAI deployment, and low latency matters more than real-time vetting. The team is considering the Asynchronous Filter streaming mode. Which two statements are correct? (Choose two.)
Select 2 answers.
Answer: B, E
With Asynchronous Filter, filters run asynchronously and content returns immediately with no buffering. The trade-off is that moderation messages and violation signals are delayed, so harmful content that would otherwise be filtered can be displayed, and Microsoft strongly recommends consuming annotations and adding controls such as redaction. Default streaming buffers content and runs the filter on it first, and it's the mode for regulated customer-facing chatbots.
Deep dive
How it works
By default, Azure OpenAI streaming doesn't send tokens the moment they're generated. The completion buffers, the guardrail system runs on the buffered content, and content that passes comes back in "content chunks" the size of the buffer. Content that violates the policy is blocked and a guardrail error comes back instead. Either way, content is fully vetted before the user sees it.
Asynchronous Filter is an extra option. Content filters still run, but asynchronously, so tokens stream back immediately, one by one, with no buffering. The trade-off is that moderation messages and policy violation signals arrive late, so some harmful content that would otherwise have been filtered can reach the user. When there's a violation, the error signal is sent as soon as it's available and the stream stops. The page guarantees that signal within about 1,000 characters of the violating content.
That's why the docs strongly recommend consuming the annotations returned during the stream and adding controls such as redacting content. There are two more caveats. Content flagged as protected material after it was shown might not be eligible for Customer Copyright Commitment coverage. And in both modes, a triggered filter still charges you for the prompt and for any completion tokens generated before filtering. You turn on Asynchronous Filter in a content filtering configuration in the Foundry portal, which applies at the deployment level and requires permission to modify content filtering policies.
The other options, and when you'd use them
- "Recommended for regulated customer-facing chatbots": The docs give that as the example for Default streaming. Default is the choice when you need maximum safety and compliance, need content filtered before it's displayed, or can't handle removing content after it's shown.
- "Default mode has zero filtering latency": That describes Asynchronous Filter. Default buffers content and returns it in chunks. Use Default when every chunk must be vetted before display.
- "Asynchronous Filter turns filtering off": It doesn't. Filters still run asynchronously, and the mode supports the default and any customer-defined filter settings. Use it when low latency is critical and you can redact on the client, for example in internal development tools or creative writing assistants.
Exam tip
"Fast, token-by-token, but harmful text might briefly appear" means Asynchronous Filter plus client-side redaction. "Nothing unvetted on screen" means Default.
References
Last reviewed: