Solution for stream-based content moderation of large model outputs

更新时间:
复制 MD 格式

Use this solution to dynamically moderate and control content that large models generate as a stream.

How it works

When a user submits a prompt, the application calls a large model, which returns generated text as a continuous stream. A guardrail system intercepts this stream and checks each segment for harmful content before the text is rendered to the user.

Segmentation methods

Stream-based content moderation triggers a check after content accumulates to a certain size. Two segmentation methods are commonly used:

Method 1: Segment by character count

When N characters accumulate, send those N characters as a batch for moderation.

Method 2: Segment using a sliding window

For every X new characters generated, send the most recently generated N characters for the check. Because the window retains surrounding context, semantic accuracy improves and perceived user wait time decreases.

High-performance moderation service

If you want to keep the moderation time for each segment under 100 ms, use our high-performance service.

  • response_security_check_hp

Important
  • This service is suitable for Method 1, which segments content by character count.

  • When you use the stream-based moderation solution, you must pass the following input parameters to associate moderation checks across multiple segments:

    • chatId: Uniquely identifies a single round of interaction, which includes the user input and the large model output.

    • done: Indicates that this segment is the final segment of the current conversation.

  • For more information, see Multi-modal API integration guide.

Handle moderation results

After each segment is checked, take action based on the result:

ResultAction
No threat detectedProgressively render the corresponding text to the user.
Threat detectedImmediately stop generating subsequent content. Revoke any harmful content already rendered, or replace it with a preset compliant response.

Stopping generation and revoking displayed content as soon as a threat is detected protects users and maintains content safety.