Scenarios

Updated at:
Copy as MD

Typical scenarios for ADA include metric anomaly investigation, natural language data retrieval, periodic reports, business monitoring, and cross-engine O&M. These scenarios are independent of each other, and the only common prerequisite is a data connection.

Key concept

Scenario selection is determined by three factors. The first is the data starting point: whether the data resides in local files, an already connected engine instance, or an engine instance that has not been connected yet determines whether you need to register a connector first. The second is the metric definition source: whether metric definitions are centrally maintained determines whether you need to build a knowledge base first. The third is the execution mode: whether the work is a one-time exploration or a periodic delivery determines whether you need to encapsulate it as a scheduled task.

The execution mode has the greatest impact on how capabilities are combined. The following table provides a detailed comparison.

Dimension

One-time exploration

Productized delivery

Trigger

Manually initiated in a conversation

Automatically triggered by the system on a schedule

Core capabilities

Conversation workbench, scenario agents, file and image upload

Skills, scheduled tasks, IM channel delivery

Metric definition requirements

Can be clarified on the fly during a conversation

Must be predefined in a knowledge base or skill

Output destination

Viewed and downloaded within the conversation

Delivered to IM; imported into the artifact library for long-term retention

When to use

The problem is not well-defined and requires multiple rounds of drill-down

The process is stable and the output format is determined

Use case

Metric anomaly investigation

When a business metric fluctuates, investigation typically requires layer-by-layer data retrieval and comparison across dimensions. Completing one round manually is time-consuming, and different people may apply inconsistent metric definitions. Describe the problem directly to ADA, and the Data Analysis Expert agent identifies the anomaly interval, breaks it down by dimension, and provides root cause clues. The query and data source of each step can be expanded and verified within the conversation.

The recommended combination is the conversation workbench plus the Data Analysis Expert agent, with a knowledge base associated to provide metric definitions. This combination is preferred over writing queries directly because the investigation process requires continuous drill-down, and the conversation retains context so that follow-up questions do not require repeating background information. For more information, see Start a conversation.

Natural language data retrieval and attribution

When data retrieval requests are scattered and ad hoc, scheduling mechanisms can cause even simple requests to wait for days. After metrics, dimensions, and statistical definitions are maintained in a knowledge base, the agent converts natural language into precise queries, retrieves data across connected engine instances such as MaxCompute, Hologres, and Serverless StarRocks, and outputs results with metric definition annotations.

The upfront investment for this scenario is in building metric definitions, and the return is that the same metric produces consistent results across different people and different times. The metric definition references attached to conclusions can be traced back to specific definition entries and the knowledge base version used. If metric definitions have not been centrally maintained, data retrieval accuracy cannot be guaranteed, and you should complete the knowledge base construction first. For more information, see Knowledge base management.

Automatic daily and weekly report generation

The data retrieval, analysis, and formatting process for recurring reports is stable, but each delivery still requires manual execution. The recommended implementation sequence is: first complete a report in a conversation to verify that the metric definitions and format are correct; then save the process as a skill with fixed input requirements and output format; finally create a scheduled task that binds to the skill and target agent, executes on a schedule, and delivers results to DingTalk, WeCom, or Lark through an IM channel.

The reason for following the sequence of "verify in a conversation first, then productize" rather than configuring a task directly is that unverified metric definitions cannot be caught in time during unattended execution. When reports need to be retained for the long term, import them into the artifact library for unified retrieval and sharing. For more information, see Scheduled tasks and Artifact management.

24/7 business monitoring

Anomalies such as data pipeline interruptions, metric threshold breaches, and data quality degradation are often discovered only during post-event reviews due to the lack of continuous monitoring. Configure a scheduled task to periodically execute monitoring logic that checks data completeness, consistency, and timeliness. When an anomaly is detected, the task generates an alert with a preliminary analysis and delivers it through an IM channel.

Compared with manual inspection, anomaly detection is advanced from post-event review to within the execution cycle, and alerts include context so that the person who takes over does not have to investigate from scratch. Monitoring tasks run in independent conversations, and historical execution records can be reviewed one by one. For more information, see Scheduled tasks.

Cross-engine intelligent O&M

When a big data platform spans offline data warehouses, real-time data warehouses, search services, and machine learning platforms, job diagnostics, query optimization, and cluster inspections are scattered across the consoles of individual engines, and O&M expertise becomes fragmented. Subscribe to official agents by engine in the Capability Marketplace to consolidate O&M operations into a single conversation entry point.

The capability scope of each official engine agent is as follows.

Official agent

Capabilities

MaxCompute Agent

SQL development, job tuning, resource management, and partition lifecycle governance for MaxCompute

DataWorks Data Agent

SQL completion, lineage analysis, schedule settings, and data quality rule recommendations

Hologres Agent

Query optimization, table management, and performance diagnostics with support for hybrid row-column analytics

Flink Agent

Flink SQL development, job diagnostics, checkpoint optimization, and real-time data pipeline construction

EMR AI Assistant

Product consultation, task analysis, exception diagnostics, and operation assistance for EMR

Elasticsearch Intelligent Assistant

Cluster management, index lifecycle, search query tuning, and log analysis

PAI Agent

Model training, experiment management, hyperparameter tuning, and inference deployment

The reason to subscribe to official agents instead of building your own is that engine diagnostics depend on compute resources and job states on the engine side. Official agents are released and continuously maintained by the engine product teams. After ADA is connected, it can observe these states and provide diagnostic recommendations. The diagnostic capabilities are updated as the publishers release new versions. For more information, see AI Capability Center.

Enterprise-specific business scenarios

Decision rules for scenarios such as finance, risk control, and operations belong to internal enterprise knowledge that general-purpose agents cannot cover. You can clone a built-in agent and then customize the system prompt, associate enterprise knowledge bases, and bind dedicated skills to create an enterprise-customized agent that can be shared across your team. The ownership of a customized agent is specified at creation time. Set it to be visible to all members within the tenant to make it available to the team. For more information, see Agent Management and Team Collaboration & Resource Sharing.

Scenarios and capabilities

Scenario

Primary capability

Supplementary capability

Metric anomaly investigation

Conversation workbench + Data Analysis Expert

Knowledge base provides metric definitions and context

Natural language data retrieval and root cause analysis

Knowledge base + data connection

Semantic model and reference tracing

Automatic daily/weekly report generation

Skill + scheduled task

IM channel delivery and artifact management

24/7 business monitoring

Scheduled task + data quality inspection skill

IM channel alert delivery

Cross-engine intelligent OM

Official agents for each engine

Multiple agents collaborate to complete a single task

Enterprise-specific business scenarios

Enterprise-customized agent + enterprise knowledge base

Custom skill encapsulation

FAQ

  • Q: A scenario involves multiple capabilities. Where do I start?

    A: First, manually complete an end-to-end workflow in the conversation workbench to verify that data access and metric definitions are correct. Then, gradually save the results as skills and scheduled tasks. For a step-by-step example, see Quick Start: Complete Your First Conversational Analysis. Subscribe to the required official agents from the Capability Marketplace as needed. You do not need to activate all of them at the beginning.

  • Q: Metric definitions are not centrally maintained. Can I start conversational data retrieval directly?

    A: Yes, but result consistency cannot be guaranteed. When metric definitions are scattered across team documents, different expressions lead to different data retrieval logic. We recommend that you first add the definitions of frequently used metrics to the knowledge base before expanding usage.

  • Q: Which industries are these scenarios applicable to?

    A: The scenarios are industry-agnostic. They depend on data formats and collaboration methods. Use cases such as data analysis, report delivery, and O&M diagnostics can all adopt the preceding combinations.

  • Q: Do I need to implement multiple scenarios at the same time?

    A: No. Scenarios are independent of each other and can be implemented separately based on business priorities. The only shared prerequisite is data connectivity. The depth of metric definition development depends on the accuracy requirements of each scenario.