Use Large Language Model (LLM) moderation when traditional rule-based systems struggle with nuanced, context-dependent, or implicit content. AI Guardrails Enhanced Edition provides Qwen-customized LLM moderation services across text, images, videos, voice, and documents, delivering contextual analysis, detailed risk labels, strong generalization, and improved dialect and multilingual recognition. This topic describes when to use each LLM moderation service and how to integrate and configure them.
Prerequisites
-
Activate AI Guardrails Enhanced Edition.
-
Integrate AI Guardrails Enhanced Edition - Text Moderation LLM, AI Guardrails Enhanced Edition - Image Moderation LLM, AI Guardrails Enhanced Edition - Video Detection LLM, AI Guardrails Enhanced Edition - Voice Moderation LLM, and AI Guardrails Enhanced Edition - Document Detection LLM.
LLM overview
-
Alibaba Cloud AI Guardrails currently provides Image Moderation LLM, Text Moderation LLM, Video Moderation LLM, Voice Moderation LLM, and Document Moderation LLM services. The specific services are as follows:
|
LLM |
Service |
Detection Content |
Use cases |
|
Image Moderation LLM |
Image Moderation for Large and Small Model Integration (postImageCheckByVL) |
Combines the capabilities of the image moderation Large Language Model (LLM) and expert models to comprehensively detect violations in images, including pornography, sexually suggestive content, politically sensitive material, violence, terrorism, prohibited items, religious content, promotional ads, and harmful content. Returns detailed tags. |
Optimized for image moderation scenarios where the highest detection accuracy is required. Use this service when high performance is critical. For more information, see Enhanced Image Moderation Service Based on Large Language Models. |
|
Image Moderation for Large and Small Model Integration_For regions outside the Chinese mainland (postImageCheckByVL_cb) |
Built for outside China scenarios, this service combines the image moderation LLM and expert models to comprehensively detect violations in images, including pornography, sexually suggestive content, politically sensitive material, violence, terrorism, prohibited items, religious content, flags, promotional ads, harmful content, and abusive language. Returns detailed tags. |
Optimized for image moderation scenarios where the highest detection accuracy is required. Use this service when high performance is critical. For more information, see Enhanced Image Moderation Service Based on Large Language Models. |
|
|
Image Moderation Service Based on LLMs (baselineCheckByVL) |
A moderation Large Language Model fine-tuned for image moderation scenarios. Detects risks in images such as pornography, politically sensitive content, violence, terrorism, prohibited items, harmful content, abusive language, and ads. Returns broad-category tags. |
A detection service primarily based on the image moderation LLM. Use this service to experience LLM-powered moderation capabilities. For more information, see Enhanced Image Moderation Service Based on Large Language Models. |
|
|
Image Ad Detection LLM (adCheckByVL) |
Uses a custom-trained LLM for deep semantic understanding of images to accurately identify various highly evasive and adversarial ad and spam behaviors, such as disguised contact information, misleading redirects, QR code steganography, and AI-generated spam ads. |
Managing and combating spam ads on social media, content communities, and forums. Preventing non-compliant ads in e-commerce product reviews and Q&A sections. Detecting spam in images within comments and live comments in live streaming and short video scenarios. |
|
|
Text Moderation LLM |
UGC Text Moderation Large Model Service_Professional Edition (ugc_moderation_byllm_pro) |
Moderates all types of user-generated content (UGC) text. Detects various violations with strong contextual analysis and understanding. Excels at identifying positive/negative sentiment, metaphors, and ideological content. The Professional Edition provides more granular risk tags. |
A text moderation service powered by large language model capabilities. Efficiently and accurately identifies various text violations. For details, see Text Moderation with LLMs. |
|
LLM-based Text Moderation Service in UGC Scenarios (ugc_moderation_byllm) |
Moderates all types of user-generated content (UGC) text. Detects various violations with strong contextual analysis and understanding. Excels at identifying positive/negative sentiment, metaphors, and ideological content. |
A text moderation service powered by large language model capabilities. Efficiently and accurately identifies various text violations. For details, see Text Moderation with LLMs. |
|
|
Cross-border UGC text moderation (LLM) (ugc_moderation_byllm_cb) |
Built for outside China scenarios, this UGC text moderation service leverages large language model capabilities to efficiently and accurately identify various text violations. |
Use this service for moderating all types of UGC text in outside China scenarios. For more information, see Text Moderation with LLMs. |
|
|
LLM-based Text Moderation Service in AIGC Scenarios (aigc_moderation_byllm) |
Moderates all types of AIGC text. Detects various violations with strong contextual analysis and understanding. Excels at identifying positive/negative sentiment, metaphors, and ideological content. |
A text moderation service powered by large language model capabilities. Efficiently and accurately identifies various text violations. For more information, see Text Moderation with LLMs. |
|
|
Video Moderation LLM |
Video File Moderation_LLM-based Version (videoDetectionByVL) |
Uses the image moderation Large Language Model service to analyze video files for visual or audio violations. |
Detects prohibited or inappropriate content in video files. You can configure image LLM moderation rules. The default limit is 10 concurrent ingest endpoints. Control your usage accordingly. For more information, see Video File Detection – Large Language Model Edition. |
|
Video Audit LLM (International) (videoDetectionByVL_cb) |
Built for outside China scenarios, this service checks video files for visual or audio violations using the image moderation Large Language Model service. |
Detects prohibited or inappropriate content in video files distributed outside China. You can configure image LLM moderation rules. The default limit is 10 concurrent ingest endpoints. Control your usage accordingly. For more information, see Video File Detection – Large Language Model Edition. |
|
|
Live Stream Moderation_LLM-based Version (liveStreamDetectionByVL) |
Uses the image moderation Large Language Model service to analyze live video streams for visual or audio violations. |
Detects prohibited or inappropriate content in live video streams. You can configure image LLM moderation rules. The default limit is 5 concurrent ingest endpoints. Control your usage accordingly. For more information, see Live Video Stream Moderation – Large Language Model Edition. |
|
|
Live Stream Detection LLM (International) (liveStreamDetectionByVL_cb) |
Built for outside China scenarios, this service checks live video streams for visual or audio violations using the image moderation Large Language Model service. |
Detects prohibited or inappropriate content in live video streams distributed outside China. You can configure image LLM moderation rules. The default limit is 10 concurrent ingest endpoints. Control your usage accordingly. For more information, see Live Video Stream Moderation – Large Language Model Edition. |
|
|
Voice Moderation LLM |
Audio and Video Media Detection_LLM Edition (audio_detection_byllm) |
Leverages Tongyi large language model capabilities to detect risks in audio, including politically sensitive content, violence, terrorism, prohibited items, and pornography. Returns detailed tags. |
Use this service for audio and video media moderation scenarios requiring high accuracy. For more information, see LLM Voice Moderation. |
|
Audio and Video Moderation (LLM Global) (audio_detection_byllm_cb) |
Built for outside China scenarios, this service adds overseas risk controls to detect various violations in audio. |
Use this service for audio and video media moderation in outside China business scenarios. For more information, see LLM Voice Moderation. |
|
|
Social & Entertainment Live Stream Detection_LLM Edition (stream_detection_byllm) |
Delivers the best overall performance. Ideal for high-accuracy live stream moderation. Detects various violations in audio streams. |
Use this service for social and entertainment live stream moderation. It offers the best overall performance. For more information, see LLM Voice Moderation. |
|
|
Social Live Streaming Moderation (LLM Global) (stream_detection_byllm_cb) |
Built for outside China scenarios, this live stream moderation service adds overseas risk controls. |
Use this service for social and entertainment live stream moderation in outside China scenarios. For more information, see LLM Voice Moderation. |
|
|
Multimodal Text-and-Image Moderation LLM |
Multimodal Post Text-and-Image Detection (post_text_image_detection) |
Analyzes both main post content and comments to assess overall risk. Detects politically sensitive content, pornography, sexual innuendo, harmful content, ads, abusive language, violence, terrorism, and prohibited items. |
Designed for communities and forums with large volumes of posts. Performs comprehensive risk assessment by combining post images, text, and context. For more information, see Multimodal content moderation. |
|
Multimodal Profile Picture and Nickname Detection (profile_text_image_detection) |
Performs integrated risk analysis of profile pictures and nicknames. Detects politically sensitive content, pornography, sexual innuendo, harmful content, ads, abusive language, violence, terrorism, and prohibited items. |
Designed for social platforms, instant messaging, games, e-commerce, education, and other scenarios involving user profiles. For more information, see Multimodal content moderation. |
|
|
Document Moderation LLM |
General Document Moderation (Large Model Edition) (document_detection_byvl) |
Uses the image moderation Large Language Model service for document images. Detects image or text violations in documents, including pornography, sexually suggestive content, politically sensitive material, violence, terrorism, and prohibited items. |
Use this service for general document content detection. Supports files in formats such as Word, PowerPoint, Excel, and PDF. For more information, see Document Moderation Large Language Model Edition. |
|
Custom Agent |
Custom Moderation Agent |
Lets you create and configure custom moderation agents. Powered by large language models, this feature enables rapid implementation of business-specific detection and filtering through flexible user-defined interaction configurations. |
Use this service when your platform requires custom detection categories and filtering. For more information, see Build custom detection with moderation agents. |
-
APIs to integrate and configure:
|
LLM |
API |
API service configuration |
|
Image Moderation LLM |
ImageModeration |
|
|
Text Moderation LLM |
TextModerationPlus |
|
|
Video Moderation LLM |
VideoModeration |
|
|
Voice Moderation LLM |
VoiceModeration |
|
|
Image-Text Multi-modal Moderation LLM |
MultimodalAsyncModeration |
post_text_image_detection profile_text_image_detection |
|
Document Moderation LLM |
FileModeration |
Adjust detection rules
Alibaba Cloud AI Guardrails supports adjusting detection rules through the console. For more information, see AI Guardrails console. For example, using the ImageModeration API operation:
-
In the navigation pane on the left, choose .
-
On the Rules Management tab, using the Image Moderation for Large and Small Model Integration (postImageCheckByVL) scenario as an example, click Set Thesaurus in the Operation column.
-
Select the detection type to adjust. For example, undesirable content detection.
-
Click Edit to enter edit mode. Modify the corresponding detection status.
-
Click Save to save the newly configured detection scope. The new configuration takes effect in approximately 2 to 5 minutes and goes live.
-