Multimodal API integration guide-AI Guardrails(AI Guardrails)-阿里云帮助中心

This document describes how to call the multimodal real-time moderation API of AI Guardrails.

Step 1: Activate the service

Go to the Activate AI Guardrails page to activate AI Guardrails.

Step 2: Grant permissions to a RAM user

Before you integrate the software development kit (SDK) or call the API, you must grant permissions to a Resource Access Management (RAM) user. You can create an AccessKey pair for your Alibaba Cloud account and RAM users. You must use the AccessKey pair to complete identity verification when you call Alibaba Cloud API operations. For more information about how to obtain an AccessKey pair, see Obtain an AccessKey pair.

Grant permissions to a RAM user

Log on to the RAM console using your Alibaba Cloud account.
Create a RAM user. For details, see Create a RAM user.
Grant the AliyunYundunGreenWebFullAccess system policy to the RAM user. This policy grants full access to Content Moderation. For details, see Manage RAM user permissions.
The RAM user can now call the Content Moderation API.

Step 3: Install and integrate the SDK

For more information about AI Guardrails SDKs, see Multimodal SDK reference

API description

Usage notes

You can call this operation to create a text Content Moderation task.

Operation:MultiModalGuard
Supported regions and endpoints:

Region	Public endpoint	Internal endpoint
China (Shanghai)	https://green-cip.cn-shanghai.aliyuncs.com	https://green-cip-vpc.cn-shanghai.aliyuncs.com
China (Beijing)	https://green-cip.cn-beijing.aliyuncs.com	https://green-cip-vpc.cn-beijing.aliyuncs.com
China (Hangzhou)	https://green-cip.cn-hangzhou.aliyuncs.com	https://green-cip-vpc.cn-hangzhou.aliyuncs.com
China (Shenzhen)	https://green-cip.cn-shenzhen.aliyuncs.com	https://green-cip-vpc.cn-shenzhen.aliyuncs.com
China (Chengdu)	https://green-cip.cn-chengdu.aliyuncs.com	None
China (Hong Kong)	https://green-cip.cn-hongkong.aliyuncs.com	None
China East 1 Finance (Hangzhou)	https://green-cip.cn-hangzhou-finance.aliyuncs.com	None
China East 2 Finance (Shanghai)	https://green-cip.cn-shanghai-finance-1.aliyuncs.com	None
Shenzhen Finance	https://green-cip.cn-shenzhen-finance-1.aliyuncs.com	None
Singapore	https://green-cip.ap-southeast-1.aliyuncs.com	https://green-cip-vpc.ap-southeast-1.aliyuncs.com
Germany (Frankfurt)	https://green-cip.eu-central-1.aliyuncs.com	https://green-cip-vpc.eu-central-1.aliyuncs.com
US (Virginia)	https://green-cip.us-east-1.aliyuncs.com	https://green-cip-vpc.us-central-1.aliyuncs.com

Billing: This is a paid API. Only requests with HTTP status code 200 are metered and billed. Requests that return other error codes are not billed. For more information about billing methods, see the Activation and billing section.
Image requirements:
- Images support the following formats: PNG, JPG, JPEG, BMP, WEBP, TIFF, SVG, AVIF, HEIF (for HEIF, the longest edge must be shorter than 8,192 px), GIF (the first frame is used), and ICO (the last image is used).
- The size of an image cannot exceed 20 MB. The height and width cannot exceed 30,000 px, and the total number of pixels cannot exceed 250 million px. A resolution greater than 200 × 200 px is recommended. If the resolution is too low, the effectiveness of the Content Moderation algorithm is compromised.
- Images must be downloaded within 3 seconds. If the download takes more than 3 seconds, a download timeout is returned.

Important

QPS limits: This operation allows up to 50 queries per second (QPS) for a single user (10 QPS by default when files are involved). If the limit is exceeded, API calls are throttled, which may affect your business. We recommend that you call this operation at a reasonable rate.

Request parameters

Name

Type

Required

Example

Description

Service

String

Yes

query_security_check_pro

AI input Content Moderation Pro (query_security_check_pro)
AI-generated Content Moderation Pro (response_security_check_pro)
AI input Content Moderation (query_security_check)
AI-generated Content Moderation (response_security_check)

Important

query_security_check_pro and response_security_check_pro are officially released. They classify the detected labels at a finer granularity in terms of content compliance protection, and the total number of labels is significantly larger than that of the previous version.

AI input Content Moderation Global Edition (query_security_check_cb)
AI-generated Content Moderation Global Edition (response_security_check_cb)
AIGC input image moderation
(img_query_security_check)
AIGC output image moderation
(img_response_security_check)
Multimodal (text and image) Content Moderation
(text_img_security_check)
Real-time file moderation
(file_security_sync_check)
Real-time multimodal (text file) moderation
(text_file_sec_sync_check)
AIGC audio moderation (audio_security_check)

ServiceParameters

JSONString

Yes

The set of parameters required by the moderation service. The value is a JSON string. For the description of each parameter, see ServiceParameters.

Table 1. ServiceParameters

Name	Type	Required	Example	Description
content	String	At least one content parameter must be passed in	The text to be moderated	The text content to be moderated. Important A maximum of 2,000 characters can be entered in a single request.
imageUrls	JSONArray		http://xxxx123	Only one image is supported. Important
fileUrls	JSONArray		http://xxxx456	Only one file is supported. Important The file size cannot exceed 10 MB.
chatId	String	No. This parameter is required if streaming moderation is used.	ABC123	Uniquely identifies an interaction record of a round of "user input + LLM output".
sessionId	String	No	14****	The session ID. This parameter marks the content of the current request as part of the multi-turn conversation in the same session window.
done	Boolean	No. This parameter is recommended if streaming moderation is used.	true	true: indicates that the chunk is the end marker of the current conversation round false: indicates that the chunk is not the end marker of the current conversation round
dataId	String	No	id123******	Uniquely identifies the content to be moderated.
accountId	String	No	13****	The account ID that uniquely identifies an account. The context moderation feature must be applied for separately. To enable the feature, contact your business manager or submit a ticket. If the feature is not enabled, accountId is used only as an account identifier and does not accumulate the context of cross-request data for the same account. The complete content to be moderated must be sent in a single request. Sending the same content in multiple requests may cause risks to go undetected.
ip	String	No	192.168.1.***	The IP address of the account ID.
referer	String	No	www.aliyun.com	The Referer request header, used for hotlink protection and other scenarios. The length cannot exceed 256 characters.
referenceContent	String	No	Context content	The context content used for comparison with the content to be moderated to detect model hallucinations. Important Model hallucination detection depends on this parameter. If you enable the model hallucination detection capability but do not actually pass in this content, the capability does not take effect. In this case, the API does not return an error and returns the moderation result as normal (Code 200), but the returned Detail does not contain a detection item whose Type is modelHallucination. To check whether hallucination detection takes effect, see the "No hallucination detected (control)" example in the Examples section of this topic.

accountId, sessionId, and chatId all identify the ownership of a request, but they differ in scope. The following table describes the purposes and use scenarios of the three parameters.

Parameter	Purpose	Use scenario
accountId	Identifies multiple requests initiated by the same account	Account-level identifier. To accumulate context across requests of an account for moderation, you must apply to enable the context moderation feature.
sessionId	Marks the content of the current request as part of the streaming response of the same LLM output	Required when you need to chain multiple chunks of content for moderation by session
chatId	Uniquely identifies an interaction record of a round of "user input + LLM output"	Required when streaming moderation is used

Response parameters

Name	Type	Example	Description
Code	Integer	200	The status code. For more information, see Code description.
Data	JSONObject	{"Result":[...]}	The moderation result data. For more information, see Data.
Message	String	OK	The response message to the request.
RequestId	String	AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****	The request ID.

Table 2. Data

Name

Type

Example

Description

Detail

JSONArray

The content compliance risk labels, confidence scores, and other results of the moderated content. For more information, see Detail.

Suggestion

String

Pass

The moderation suggestion.

block: block is recommended
pass: pass is recommended
watch: watch is recommended
mask: mask is recommended

Note

Currently, only sensitive content detection supports the watch and mask suggestions. Other detection dimensions support only block or pass.
If you moderate multiple dimensions, we merge the Suggestion results of all dimensions. The merge priority in descending order is block, mask, watch, pass.

Table 3. Detail

Name	Type	Example	Description
Suggestion	String	pass	The moderation suggestion. block: block is recommended pass: pass is recommended watch: watch is recommended mask: mask is recommended Note Currently, only sensitive content detection supports the watch and mask suggestions. Other detection dimensions support only block or pass. If you moderate multiple dimensions, we merge the Suggestion results of all dimensions. The merge priority in descending order is block, mask, watch, pass.
Type	String	contentSecurity	The protection dimension. contentModeration: content compliance detection promptAttack: prompt attack detection sensitiveData: sensitive content detection modelHallucination: model hallucination maliciousFile: malicious file detection maliciousUrl: malicious URL detection waterMark: digital watermarking identifier customLabel: custom detection agent
Level	String	high	The risk level, returned based on the configured high- and low-risk scores. Valid values: high: high risk. If a custom word library is hit, the risk level is high by default. medium: medium risk low: low risk none: no risk detected Note We recommend that you handle high-risk content directly, manually review medium-risk content, and handle low-risk content only when high recall is required. In routine operations, handle low-risk content in the same way as content with no detected risk. You can configure the risk scores in the AI Guardrails console. The sensitivity level (for SensitiveData). Valid values: S0, S1, S2, S3 S0 indicates that no sensitive content is detected. A larger number indicates a higher sensitivity level.
Result	JSONArray		The content compliance risk labels, confidence scores, and other results of the moderated content. For more information, see Result.

Table 4. Result

Name	Type	Example	Description
Description	String	Suspected political entity	The description of the Label field. Important This field describes the Label field and may be changed. We recommend that you process results based on the Label field instead of this field.
Confidence	Float	81.22	The confidence score, ranging from 0 to 100 and rounded to two decimal places. Some labels do not have confidence scores.
Label	String	political_xxx	The label returned after text content detection. Multiple labels and scores may be returned.
Level	String	high	The risk level, returned based on the configured high- and low-risk scores. Valid values: high: high risk. If a custom word library is hit, the risk level is high by default. medium: medium risk low: low risk none: no risk detected Note We recommend that you handle high-risk content directly, manually review medium-risk content, and handle low-risk content only when high recall is required. In routine operations, handle low-risk content in the same way as content with no detected risk. You can configure the risk scores in the AI Guardrails console. The sensitivity level (for SensitiveData). Valid values: S0, S1, S2, S3 S0 indicates that no sensitive content is detected. A larger number indicates a higher sensitivity level.
Ext	JSONObject		Some protection dimensions return the corresponding extension information. For more information, see Ext.

Table 5. Ext

Name	Type	Example	Description
Riskwords	String	AA,BB,CC	Applicable protection dimension: contentModeration: content compliance detection The detected sensitive words, separated by commas (,). Sensitive words are not returned for some labels.
CustomizedHit	JSONArray	[{"LibName":"...","Keywords":"..."}]	Applicable protection dimension: contentModeration: content compliance detection If a custom word library is hit, the value of Label is customized, and the name of the custom library and the custom words are returned.
SensitiveData	JSONArray	["6201112223455"]	Applicable protection dimension: sensitiveData: sensitive content detection The detected sensitive data samples.
Desensitization	String	...[Mobile phone number] is my contact information...	Applicable protection dimension: sensitiveData: sensitive content detection The masked content.
FileUrl	String	https://sase-public-server-files.oss-cn-hangzhou.aliyuncs.com/saas-XXX	Applicable protection dimension: waterMark: digital watermarking identifier The download URL of the watermarked file.
OutFileSize	String	152357	Applicable protection dimension: waterMark: digital watermarking identifier The file size.
FileUrlExp	String	1754135551	Applicable protection dimension: waterMark: digital watermarking identifier The time when the download URL of the watermarked file expires.
Filename	String	B7VKehJ4gZR.png	Applicable protection dimension: waterMark: digital watermarking identifier The file name.
OutFileHashMd5	String	8b96ff73e8d8060016bb41b16d337871	Applicable protection dimension: waterMark: digital watermarking identifier The MD5 hash of the file.

Table 6. CustomizedHit

Name	Type	Example	Description
LibName	String	Custom library 1	The name of the custom library.
Keywords	String	Custom word 1,Custom word 2	The custom words, separated by commas (,).

Examples

Request example

{
  "Service": "XXX",
  "ServiceParameters": {
    "content": "This thermos cup supports IP67 water resistance and can be used underwater.",
    "chatId": "ABC123",
    "dataId": "id123******",
    "accountId":"abc",
    "sessionId":"abc",
    "imageUrls": ["http://xxxx"], # Only one image is supported
    "fileUrls": ["http://xxxx"], # Only one file is supported
    "referer": "http://www.aliyun.com",
    "referenceContent": "Product name: Smart thermos cup. Material: 316 stainless steel. Heat preservation duration: 12 hours. Capacity: 500 ml. Packing list: 1 thermos cup and 1 user manual."
  }
}

Correct usage: send the complete content to be moderated in a single request. If the same content is sent in multiple requests, each request is moderated independently, and risks may not be detected. To accumulate context across requests of an account for moderation, apply to enable the context moderation feature first.

Response examples:

Moderation by query_security_check_pro, with a system policy hit:

{
  "Code": 200,
  "RequestId": "AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****",
  "Message": "OK",
  "Data": {
    "Suggestion": "block",
    "DataId": "id123******",
    "Detail": [
      {
        "Suggestion": "block",
        "Type": "maliciousUrl",
        "Level": "high",
        "Result": [
          {
            "Description": "Information theft",
            "Label": "Info Stealer",
            "Level": "high"
          }
        ]
      },
      {
        "Suggestion": "mask",
        "Type": "sensitiveData",
        "Level": "S2",
        "Result": [
          {
            "Ext": {
              "Desensitization": "...[Mobile phone number] is my contact information...",
              "SensitiveData": [
                "136********"
              ]
            },
            "Description": "Mobile phone number (Chinese mainland)",
            "Label": "1814",
            "Level": "S2"
          },
          {
            "Ext": {
              "SensitiveData": [
                "**City"
              ]
            },
            "Description": "City (Chinese mainland)",
            "Label": "1739",
            "Level": "S0"
          }
        ]
      },
      {
        "Suggestion": "block",
        "Type": "promptAttack",
        "Level": "high",
        "Result": [
          {
            "Description": "Refusal suppression jailbreak",
            "Confidence": 100,
            "Label": "Refusal Supression Jailbreak",
            "Level": "high"
          }
        ]
      },
      {
        "Suggestion": "block",
        "Type": "customLabel",
        "Level": "high",
        "Result": [
          {
            "Description": "Custom label hit",
            "Label": "*** Business risk expression",
            "Level": "high"
          }
        ]
      },
      {
        "Suggestion": "block",
        "Type": "contentModeration",
        "Level": "high",
        "Result": [
          {
            "Ext": {
              "RiskWords": "The name of a political entity",
              "Advice": [
                {
                  "answer": "test",
                  "hitLabel": "political_entity",
                  "hitLibName": "pornographic_adult-0827"
                }
              ]
            },
            "Description": "Suspected political entity",
            "Confidence": 100,
            "Label": "political_entity",
            "Level": "high"
          },
          {
            "Ext": {
              "CustomizedHit": [
                {
                  "LibName": "To be blacklisted and blocked",
                  "KeyWords": "word a,word b,word c"
                }
              ]
            },
            "Description": "Custom library hit",
            "Confidence": 100,
            "Label": "customized",
            "Level": "high"
          }
        ]
      },
      {
        "Result": [
          {
            "Description": "Intrinsic hallucination",
            "Confidence": 95,
            "Label": "Intrinsic Hallucination",
            "Level": "medium"
          }
        ],
        "Type": "modelHallucination",
        "Suggestion": "block",
        "Level": "medium"
      }
    ]
  }
}

Moderation by img_response_security_check, with a system policy hit:

  {
  "Code": 200,
  "Data": {
    "Detail": [
      {
        "Level": "none",
        "Result": [
          {
            "Confidence": 0.0,
            "Description": "No risk detected",
            "Ext": {
              "FileUrl": "https://sase-public-server-files.oss-cn-hangzhou.aliyuncs.com/saas-XXX",
              "OutFileSize": 527918,
              "FileUrlExp": "1754200240",
              "Filename": "wJGz6kmZ1Ce.jpg",
              "OutFileHashMd5": "02f5129f606027c7a87b84377ec98f8e"
            },
            "Label": "nonLabel",
            "Level": "none"
          }
        ],
        "Suggestion": "pass",
        "Type": "waterMark"
      },
      {
        "Level": "high",
        "Result": [
          {
            "Confidence": 90,
            "Description": "Advertising Law violation - extreme word",
            "Label": "ad_Compliance_WordLimit_Tii",
            "Level": "high"
          }
        ],
        "Suggestion": "block",
        "Type": "contentModeration"
      }
    ],
    "Suggestion": "block"
  },
  "Msg": "OK"
}

Moderation by text_img_security_check, with a system policy hit:

{
  "Code": 200,
  "Data": {
    "Detail": [
      {
        "Ext": {},
        "Level": "high",
        "Result": [
          {
            "Confidence": 98.34,
            "Description": "Female cleavage",
            "Label": "sexual_Cleavage",
            "Level": "high"
          }
        ],
        "Suggestion": "block",
        "Type": "contentModeration"
      }
    ],
    "Suggestion": "block"
  },
  "Msg": "OK"
}

Moderation by file_security_sync_check, with a system policy hit:

{
  "Code": 200,
  "Data": {
    "Detail": [
      {
        "Ext": {},
        "Level": "high",
        "Result": [
          {
            "Confidence": 100,
            "Description": "Website backdoor",
            "Label": "WebShell",
            "Level": "high"
          }
        ],
        "Suggestion": "block",
        "Type": "maliciousFile"
      },
      {
        "Ext": {
          "PageSum": 1
        },
        "Level": "none",
        "Result": [
          {
            "Description": "No risk detected",
            "Label": "nonLabel",
            "Level": "none"
          }
        ],
        "Suggestion": "pass",
        "Type": "contentModeration"
      }
    ],
    "Suggestion": "block"
  },
  "Msg": "OK"
}

Moderation by text_file_sec_sync_check, with a system policy hit:

{
  "Code": 200,
  "Data": {
    "Detail": [
      {
        "Ext": {
          "PageSum": 4
        },
        "Level": "none",
        "Result": [
          {
            "Description": "No risk detected",
            "Label": "nonLabel",
            "Level": "none"
          }
        ],
        "Suggestion": "pass",
        "Type": "contentModeration"
      }
    ],
    "Suggestion": "pass"
  },
  "Msg": "OK"
}

Model hallucination detection examples

Model hallucination detection (in public preview) identifies false or inaccurate information in model output by comparing the content generated by the LLM with the reference content passed in referenceContent. The following examples cover three typical hallucination scenarios and one control scenario in which no hallucination is detected. Each example includes the scenario description, the input content (the content to be moderated + referenceContent), the moderation result, and the result explanation. You can copy the example content to verify the integration effect on your own. Prerequisite: in Configure check items, you have enabled the Model Hallucination dimension for the Service, and in the ServiceParameters of the request, you have passed in referenceContent.

Scenario 1: Fabricating information that does not exist in the reference content

Scenario description: The reference content does not mention water resistance, but the model output fabricates "IP67 water resistance". This is fabricating information that does not exist.

Input content:

{
  "Service": "query_security_check_pro",
  "ServiceParameters": {
    "content": "This thermos cup supports IP67 water resistance and can be used underwater.",
    "referenceContent": "Product name: Smart thermos cup. Material: 316 stainless steel. Heat preservation duration: 12 hours. Capacity: 500 ml. Packing list: 1 thermos cup and 1 user manual."
  }
}

Moderation result:

{
  "Code": 200,
  "Message": "OK",
  "RequestId": "AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****",
  "Data": {
    "Suggestion": "block",
    "Detail": [
      {
        "Type": "modelHallucination",
        "Suggestion": "block",
        "Level": "medium",
        "Result": [
          {
            "Description": "Intrinsic hallucination",
            "Confidence": 95,
            "Label": "Intrinsic Hallucination",
            "Level": "medium"
          }
        ]
      }
    ]
  }
}

Result explanation: "IP67 water resistance" in the content to be moderated is not supported by the reference content and is determined to be a model hallucination. The returned Detail contains a detection item whose Type is modelHallucination, the Label is Intrinsic Hallucination, the Confidence is the confidence score of the hallucination determination, and the moderation suggestion is block.

Scenario 2: Fabricating information that contradicts the reference content

Scenario description: The reference content explicitly states that a 7-day no-reason return is supported, but the model output claims that "returns are not supported". This is fabricating information that contradicts the reference content.

Input content:

{
  "Service": "query_security_check_pro",
  "ServiceParameters": {
    "content": "This product does not support returns, and no refund can be requested after the order is placed.",
    "referenceContent": "After-sales service: This product supports a 7-day no-reason return, and the return shipping fee is borne by the buyer."
  }
}

Moderation result:

{
  "Code": 200,
  "Message": "OK",
  "RequestId": "AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****",
  "Data": {
    "Suggestion": "block",
    "Detail": [
      {
        "Type": "modelHallucination",
        "Suggestion": "block",
        "Level": "medium",
        "Result": [
          {
            "Description": "Intrinsic hallucination",
            "Confidence": 95,
            "Label": "Intrinsic Hallucination",
            "Level": "medium"
          }
        ]
      }
    ]
  }
}

Result explanation: The content to be moderated directly contradicts the return policy in the reference content and is determined to be a model hallucination (Intrinsic Hallucination). The moderation suggestion is block.

No hallucination detected (control)

Scenario description: The model output is consistent with the reference content, and no hallucination is detected. The API returns a normal result (Code 200), but the Detail does not contain a detection item whose Type is modelHallucination.

Input content:

{
  "Service": "query_security_check_pro",
  "ServiceParameters": {
    "content": "The heat preservation duration of this thermos cup is 12 hours, and its capacity is 500 ml.",
    "referenceContent": "Product name: Smart thermos cup. Material: 316 stainless steel. Heat preservation duration: 12 hours. Capacity: 500 ml. Packing list: 1 thermos cup and 1 user manual."
  }
}

Moderation result:

{
  "Code": 200,
  "Message": "OK",
  "RequestId": "AAAAAA-BBBB-CCCCC-DDDD-EEEEEEEE****",
  "Data": {
    "Suggestion": "pass",
    "Detail": []
  }
}

Result explanation: The content to be moderated is faithful to the reference content and is not determined to be a hallucination, so the Detail does not contain a modelHallucination detection item. This is also the response behavior when the model hallucination dimension is enabled but referenceContent is not passed in: the API does not report an error, but does not return hallucination detection results. First check whether the request passes in referenceContent.

Note

The values of fields such as Label, Description, and Confidence in the moderation result are subject to the actual response of the API.
The model hallucination detection capability is in public preview.

Code description

Code	Status code	Description
200	OK	Cause: The request is processed successfully. Solution: Parse the Data field in the response as usual, and handle the content based on the moderation suggestion (Suggestion) returned in Detail/Result. Troubleshooting: No action is required.
400	BAD_REQUEST	Cause: The request is invalid, usually because the request parameters are incorrect. Solution: Check the request body item by item against the "Request parameters" section of this topic: make sure that Service is a service code supported in this topic, ServiceParameters is a valid JSON string, and required fields (such as content) are correctly passed in. Troubleshooting: Modify the parameters and send the request again. If 400 is still returned after you confirm that the parameters are correct, submit a ticket and provide the RequestId.
408	PERMISSION_DENY	Cause: The request failed authorization. Possible causes include an unauthorized account, an account with overdue payments, an account for which the service is not activated, or a disabled account. Solution: Make sure that the AI Guardrails service is activated, that the RAM user corresponding to the AccessKey used for the call is granted the Content Moderation permissions (such as AliyunYundunGreenWebFullAccess), and that the account status is normal and has no overdue payments. Troubleshooting: Check the following items in order: (1) whether "Step 1: Activate the service" is complete; (2) whether "Step 2: Grant permissions to a RAM user" is complete and the RAM user has the required permissions; (3) whether the Alibaba Cloud account has overdue payments or is disabled. Retry after ruling out the issues one by one. If the issue persists, submit a ticket.
500	GENERAL_ERROR	Cause: A temporary error occurred on the server. Solution: Retry later. We recommend that you add an automatic retry mechanism on the business side, such as exponential backoff retry. Troubleshooting: If this error code is returned persistently, record the RequestId and contact us via Online Service.
581	TIMEOUT	Cause: The request timed out. Solution: Retry later. Check whether the timeout setting on the business side is too short, and properly control concurrency during peak calling hours. Troubleshooting: After confirming that the client timeout is properly set and network connectivity is normal, retry. If this error code is returned persistently, record the RequestId and contact us via Online Service.
588	EXCEED_QUOTA	Cause: The request frequency exceeds the quota. This operation allows up to 50 queries per second (QPS) for a single user (10 QPS by default when files are involved). Solution: Reduce the call frequency. We recommend that you implement throttling and retry logic on the business side to prevent throttling triggered by traffic bursts. Troubleshooting: Check whether the current call QPS exceeds the preceding quota. If your business requires a higher quota, submit a ticket to apply for an adjustment.

What do I do if the model hallucination dimension is enabled but no model hallucination detection item is returned?

When the model hallucination (in public preview) dimension does not take effect in the returned result, no error code is returned: the API returns a normal result (Code 200), but the Detail does not contain a detection item whose Type is modelHallucination. In this case, troubleshoot in the following order:

Make sure that the ServiceParameters of the request contains the referenceContent parameter. Model hallucination detection relies on this parameter to compare against the content to be moderated. If the parameter is not passed in, no hallucination detection result is returned. For the parameter description and self-check examples, see the "Request parameters" and "Examples" sections of this topic.
Make sure that the Service is a text-based service (such as AI input Content Moderation and AI-generated Content Moderation). The model hallucination dimension is supported only by text-based Services; image, file, video, and audio services do not support this dimension.
Log on to the AI Guardrails console, choose Protection Config > Model protection, select the corresponding Service, go to Configuration management > Protection dimensions, and make sure that the Model Hallucination card is enabled.
In the Model Hallucination card, in Configuration management, check Hallucination Degree: this setting adjusts the interception score of model hallucination detection. Content is determined to be a model hallucination only when the score returned by the model exceeds the configured value. Value range: 60 to 100. Default value: 60. If the determination score of the content does not reach the configured interception score, no hallucination detection result is returned.

If all the preceding items are confirmed but the hallucination detection result is still not returned, submit a ticket and provide the RequestId.