Manual review 2.0-AI Guardrails(AI Guardrails)-阿里云帮助中心

Use cases

When your business encounters the following scenarios, you can use the Alibaba Cloud Human-machine Review Enhanced Edition console to customize detection rules and the detection scope, configure the human-machine review process, query detection results, and view detection statistics.

Use case	Description
Integrate a pure human review service	If you only need a pure human review service, integrate the service by calling the manual review API. Before you start, contact your account manager to confirm your detection rule configuration.
Configure a human-machine review service	If you use the Content Moderation Machine Auditing 2.0 service and want to send specific content for manual review after machine review, you can configure human-machine review rules. This automatically forwards content that meets your criteria to the manual review stage. For example, you use the Profile Photo Check (profilePhotoCheck) service of Image Moderation and want to manually review all content identified as medium-risk. You can enable human-machine review for the service and configure it to forward medium-risk content to the manual review queue. After the manual review is complete, you receive the result through a callback.
Set different risk detection scopes for multiple business scenarios	If you have multiple business scenarios that use the same service but require different risk detection scopes, you can copy the service to create separate rule configurations for each scenario. For example, you have three business scenarios (A, B, and C) that all use the generic baseline check (baselineCheck) service, but each has different moderation standards. You can copy the `baselineCheck` service in the console to create new services, such as `baselineCheck_01` and `baselineCheck_02`. By setting different risk detection scopes for these three services, you can meet the unique requirements of each business scenario.
Query detailed manual review results and re-review them	If you need to view details of recent manual review results or re-review them, you can use the Result Query page to find and act on specific results. Batch re-review operations are also supported on this page.
View recent manual review statistics	If you need to view statistics on the risk distribution and review volume of recent manual reviews, you can use the Usage Statistics page.

Prerequisites

Visit the Manual Review - Enhanced Edition page to activate the Manual Review - Enhanced Edition service.

Note

Before you activate the Manual Review - Enhanced Edition service, please understand the billing rules for Manual Review - Enhanced Edition. For more information, see Introduction and billing for Manual Review - Enhanced Edition.

Integrate a pure human review service

If you only need a pure human review service, you can call the Manual Review 2.0 detection API to submit content for review.

Activate pay-as-you-go for Manual Review 2.0.
Integrate with the asynchronous detection API for Manual Review 2.0. We recommend using the Manual Review 2.0 SDK.
Manual Review 2.0 supports services for images, text, audio, and video. You can view the supported services and labels in the Content Moderation console.
1. Log on to Content Moderation console.
2. In the navigation pane on the left, choose Manual Moderation Enhanced Edition > Rule Configuration.
3. On this page, you can view all Manual Review Services for Images, Text, Audio, and Video, or click 查看标签 to view the labels supported by the services.
After integrating the manual review API, call it with the required data to use the service.

Human-machine review configuration

If you use the Content Moderation Machine Auditing 2.0 API and want to automatically send content that meets specific criteria for manual review, you can configure a human-machine review workflow.

Activate pay-as-you-go for Manual Review 2.0.
Log on to Content Moderation console.
In the navigation pane on the left, choose Machine Auditing 2.0 > Image Moderation.
On the Rule Management tab, find the Profile Photo Check (profilePhotoCheck) service and click Settings in the Actions column.
Select the Human-machine review configuration tab to configure the workflow.
1. Turn on the Enable human-machine review switch.
2. For Content for manual review, select the content to be forwarded. For this example, select the checkbox for medium-risk content.
3. Select a Manual review rule and a Callback notification scheme.
  - All manual review results are returned through this callback.
4. Click Save.
  The new configuration takes effect in 2 to 5 minutes. Once effective, machine review results that meet the criteria are automatically sent for manual review. You can query the manual review results by using the manual review TaskId or receive them through a callback.
  
  Note
  When content triggers a human-machine review rule, it is automatically sent for manual review. The machine review result includes a ManualTaskId field for querying the manual review result. You can receive manual review results through a callback or query them by using the Manual Review 2.0 API.

Configure differentiated risk scopes

To meet the different moderation requirements of your business scenarios, you can copy a service and configure a unique risk detection scope for each copy.

Log on to the Content Moderation console.
In the navigation pane on the left, choose Manual Review 2.0 > Rule Configuration.
On the Rule Management tab, copy the service for which you want to create different detection scopes. This example uses the Image Manual Review Service (imageManualCheck).
1. In the service list, find the Image Manual Review Service and click Copy in the Actions column.
2. In the Copy Service Panel that appears, enter a Service Name and Service Description.
3. Click Create. The information of the copied service is saved and the service is created successfully. The new service can be invoked in 1 to 2 minutes.
4. After the service is created, you can modify its configuration by using the Settings and Edit actions. For example, after creating imageManualCheck_01, you can configure it differently from the original imageManualCheck service. By calling these separate services for different business needs, you can apply distinct risk detection scopes.

Result query and re-review

You can query detailed results for both manual review and human-machine review. You can filter the results by submission time, manual review type, manual review result, violation type, TaskId, DataId, re-review status, and other criteria.

Log on to Content Moderation console.
In the navigation pane on the left, choose Manual Review 2.0 > Result Query.
On the Query Results page, enter search criteria to query the detection results.
- Query results: You can use the following criteria to filter results: Submission time, Manual review type, Manual review result, Violation type, Service, TaskId, Machine review request ID, DataId, Re-review status, and Re-review violation type.
Note
By default, results are displayed in descending chronological order, with a maximum of 50,000 entries. The console provides query results for the last 7 days. We recommend that you store the data or logs from each API response for long-term analysis and record-keeping.
- Manual re-review: If you need to re-review a manual review result, you can perform a No Risk or High Risk re-review in the Re-review Actions column of the results list. Batch operations on multiple items are supported.
- Detailed results: You can click the Detailed Results button for specific content to view detailed information, including machine review information, manual review information, and re-review information.

Manual review statistics

You can view recent manual review statistics to help you refine your content moderation policies.

Log on to the Content Moderation console.
In the navigation pane on the left, choose Manual Review 2.0 > Usage Statistics.

On the Usage Statistics page, select a time range to view usage or export usage data.

Query Usage: You can view usage statistics on a daily and monthly basis. The data is retained for one year and can be queried for a period of up to two months.

Export usage: Click the Download icon in the upper-right corner of the page to export usage details by day or by month.

The exported report is in Excel format and includes only data for periods in which API calls were made. The following table describes the fields in the report.

Parameter	Description	Unit
Account UID	The UID of the account that exported the report.	N/A
Service	The detection service that was called.	N/A
Usage	The total number of calls.	Calls
Date	The date of the statistics.	Day/Month

View risk summary: The Usage Statistics page displays an overview of risk levels, risk trends, and risk distribution.
- Risk overview: Displays the total volume of content classified as No Risk and High Risk.
- Trend statistics: Displays the daily or monthly volume for each risk level within the selected time range.
- Risk distribution: Displays the distribution of the top five risk labels and all risk labels within the selected time range.