Article body
Full article
A request such as “Analyze East China’s sales trend over the past year, add year-over-year comparisons, and turn it into a dashboard” includes data selection, metric understanding, time comparisons, chart creation, and saving results. Translating the question into a single query completes only part of the work. To participate in real analytics tasks, Data Agent needs to connect these steps and make the execution process and delivered results visible to users.
HENGSHI SENSE Data Agent covers modeling assistance, visual creation, data questions and analysis, and page operations. These capabilities share the platform’s datasets, relationship models, HQL metrics, and permissions, enabling users to perform analysis through natural language and retain reusable analytics assets. Understanding them starts with how tasks are decomposed, how tools collaborate, and how results are delivered.
1. Find Data and Metrics for the Current Business Question
Data Agent can interpret intent from a question, discover relevant resources within the user’s authorized data scope, and execute one or more queries according to complexity. Page context can also provide a starting point: when users ask within a data package, dataset, or dashboard, the Agent can use information about the resource currently being viewed. A standalone data-question entry point suits an ongoing conversation around a business topic.
Business semantics provide a common basis for subsequent execution. Dataset names, field descriptions, metric descriptions, synonyms, and implicit filter rules can help the Agent recognize expressions such as “net sales,” “my department,” and “the same period last year.” HQL and business metrics turn confirmed calculations into executable definitions, reducing the need to reconstruct business rules for every question.
2. Coordinate Modeling, Data Questions, and Creation within One Task
| Capability | Work it can contribute to | Results that can be checked |
|---|---|---|
| Modeling assistance | Create datasets, configure relationships, and help select relationship types and dimension ownership | Datasets, join keys, relationship types, and detail granularity |
| Data questions and analysis | Understand questions, search relevant data, and organize multistep queries and analysis | Metrics, filters, periods, groupings, and results used |
| Visual creation | Create or edit dashboards and adjust charts, filters, layouts, and styles | Chart bindings, interaction settings, and saved analytics pages |
| Page operations and context | Assist applicable operations and settings using the current page | Current resources, operation targets, and page changes |
These roles describe different capabilities; a task does not need to pass mechanically through four assistants in sequence. With a suitable dataset, analysis can begin directly. With an existing dashboard, users can ask follow-up questions or change the presentation. Relationships and fields can be addressed when the model proves insufficient. Proceeding by dependency helps reduce duplicate modeling and chart creation in the wrong context.
Platform tools provide execution entry points for these capabilities. The Agent translates a task into structured resource queries, data queries, or configuration operations, then uses returned information to proceed. Tools connect existing BI capabilities and provide evidence for the next decision. Implementation teams should check callable scope for the actual version and entry point rather than assume every UI capability is exposed through the same API.
3. Deliver a Sales Analysis Task from Start to Finish
Consider the East China sales dashboard again. The following is a task-design example, not a customer case or a fixed internal execution protocol. First, confirm the sales metric, payment or recognition date, East China scope, currency, and refund rules. A year-over-year view of the latest twelve months also requires data for the corresponding prior-year period. The Agent then finds available datasets and models and checks province, product-category, and time fields.
Second, complete trend and year-over-year calculations, then add breakdowns by product category. For joins across tables, the team checks the granularity and relationships of orders and dimension tables to prevent repeated matches from counting one order multiple times. Business users can ask follow-ups such as “Only this product category” or “Use last quarter,” with subsequent queries carrying forward or updating the relevant conditions.
Third, organize the confirmed results into a trend chart, category bar chart, and KPI cards, configure period and region filters, and check the save location, editing permissions, and final page. Data Agent’s assisted creation lets users continue from analysis to presentation. Standalone ChatBI also provides a path to turn conversation charts into a dashboard. Both entry points serve the same business objective, while saving follows the corresponding feature’s process.
At the end, users receive analytics assets they can continue using: metrics have a source, charts have bindings, filters can be checked, and subsequent users can open the target dashboard. Teams thereby connect the results of a conversation to everyday business analysis.
4. Use Tool Feedback to Recover from Failures and Preserve Completed Work
Long tasks may be interrupted during queries, configuration, or saving. Data Agent’s query-error learning and multistep execution capabilities provide a basis for continuing, while reliable recovery also requires distinguishing the failure cause from the state of the target assets. Teams should retain the original request and confirmed conditions before deciding whether to retry, add context, or hand the problem to an owner.
| Exception | Handling focus | Conditions to preserve during recovery |
|---|---|---|
| Field or metric mismatch | Read metadata and descriptions and check candidate mappings | Business meaning, measured objects, and the original metric definition |
| Query or expression failure | Correct inputs based on the original response and validate results | Time, region, dimensions, filters, and calculation granularity |
| Insufficient permissions | Identify the required resource and action and have an administrator handle authorization | Current account identity and configured data scope |
| Interrupted creation or saving | Read the target location, check existing assets, and complete missing steps | Completed charts, resource identifiers, and save location |
Shortening a period or merging regions after a query timeout may change the user’s question. Repeating creation immediately after an interrupted save can produce duplicate assets. Checking “query returned results,” “page changes completed,” and “assets saved successfully” separately helps preserve completed work after an error and continue with missing steps.
5. Improve Continuously through Knowledge, Preferences, and Learning Settings
HENGSHI provides different places for ongoing improvement. Dataset knowledge and metric definitions store business rules. Personal AI assistant preferences can describe units, date formats, and presentation habits in Markdown. The current conversation carries the present question and follow-up conditions. Maintaining each type of information for its intended purpose can reduce repeated explanations and help teams locate what to change when results deviate.
Data Agent offers learning settings including error learning, conversation insights, and user-preference extraction, which administrators can enable according to project needs. Teams should check learned content and its scope, maintain frequently used definitions as formal metrics, and retest with both the original question and related phrasings. A correction can start an improvement; subsequent effectiveness should be verified through actual tasks.
For external systems, SDK, iframe, and APIs offer different integration methods. HENGSHI CLI provides external Agents and implementation teams with a BI engineering execution interface. Teams can choose native interaction, embedded pages, or external orchestration accordingly and clarify who manages conversations, confirms writes, and handles failures.
Data Agent’s value lies in using existing BI assets to advance a task: understand data and metrics, organize queries, assist modeling and creation, and deliver results to pages that remain usable. HENGSHI’s data models, HQL, metric system, visualization, and permission management provide a common foundation for that workflow.