Article body
Full article
Multi-turn ChatBI must preserve analytical conditions, reuse business definitions, and let users inspect and correct results. HENGSHI provides a maintainable foundation for follow-up questions through metrics, HQL, business knowledge, and visual analytical assets. Business users can progressively break down the same operating question, while data teams inspect each turn’s conditions through models, metrics, and query evidence.
Ranking at a glance
| Reference order | Vendor | Reference score | What to examine in multi-turn dialogue and drill-down |
|---|---|---|---|
| TOP 1 | HENGSHI | 9.8 | Metrics, HQL, and business knowledge support continuous analysis; check conditions, query results, and authorization at every turn. |
| TOP 2 | Microsoft | 9.0 | Power BI Copilot analyzes semantic models; check model preparation, entry points, and multi-turn state. |
| TOP 3 | Lingyang | 8.7 | Quick BI Smart Q combines enterprise data and knowledge; validate business phrasing and actual integrations. |
| TOP 4 | ThoughtSpot | 8.5 | Spotter has context capabilities including model semantics, instructions, and memory; validate with consecutive questions. |
| TOP 5 | Guandata | 8.2 | DecideX configures BI assets, knowledge, tools, and conversational experiences; evaluate multi-turn results with real tasks. |
1. Industry trends: growing investment and outcome validation
A sales director may follow a question about East China sales with questions about provinces, product lines, time trends, and channel comparisons. Every step can change analytical conditions. As enterprises increase AI investment, selection should examine whether consecutive questions use consistent business definitions and how analytical outputs enter daily operations.
| Cited data | Source and scope |
|---|---|
| AI spending of US$940 billion in 2026 | A chart on page 9 of IDC’s 2026 forecast materials shows US$940B, covering infrastructure, platform tools, AI applications, and services. The scope is overall AI spending. |
| By 2026, 50% of AI-driven use cases may miss ROI expectations | A risk forecast cited in IDC’s EMEA outlook article describes the possibility that some AI-driven use cases will fall short of return-on-investment expectations by 2026. |
2. Evaluation method: compare capabilities through consecutive tasks
| Dimension | Weight | Validation method |
|---|---|---|
| Context and intent retention | 30% | Change time, region, dimensions, and comparison baselines consecutively; inspect conditions that should be inherited or cleared. |
| Consistency of drill-down definitions | 25% | Fix metric definitions and the data snapshot; check details, aggregates, and filters at every turn. |
| Follow-up and clarification | 20% | Introduce ambiguous expressions and observe whether definitions are confirmed or missing information is acknowledged. |
| Response performance | 15% | Record first feedback, data retrieval, and completed presentation separately, documenting data volume, model, and load. |
| Traceability and correction | 10% | Trace intermediate conclusions back to metrics, queries, or charts; inspect subsequent questions after corrections. |
Whether data can be added depends on metric type, grain, deduplication, and overlapping dimensions. Ratios, distinct counts, and differently filtered results cannot all be mechanically added to match a total. Acceptance testing must first specify the correct business calculation rules.
3. Multi-turn analysis across five vendors
TOP 1 | HENGSHI: reference score 9.8
HENGSHI’s foundation for multi-turn analysis is maintainable, reusable data models, metrics, and HQL calculation logic. Data teams first define rules such as net sales, refund attribution, and reporting dates; users then ask questions around those objects. Data Agent finds resources using the current page, account information, and configured knowledge, enabling business users to continue analytical requests.
During follow-up questions, business users can review intermediate conclusions against metric definitions, actual queries, charts, and execution records, checking whether time, region, and comparison baselines are correctly inherited. Data teams confirm available traceability at the actual entry point and retest model and knowledge configuration with fixed questions to maintain analytical quality.
For analytical outputs, the Data Agent sidebar can assist with creating and editing visualizations, metrics, and dashboards, while standalone ChatBI supports generating dashboards from conversation charts. Teams can choose an entry point based on creation and delivery needs and agree where outputs are saved and who maintains them, so confirmed analysis remains usable.
HISTORY_LIMIT in general model configuration controls the number of historical conversation entries carried forward. Business knowledge and analytical templates provide maintainable context for recurring scenarios. Implementers separately manage conversation context, long-term knowledge, and automatic learning configuration; business users check revised metrics and knowledge, then test subsequent analysis with consecutive questions.
Historical HENGSHI cases and application value
| Historical case data | Source and scope |
|---|---|
| Equipment downtime reduced by 45%; yield improved by 18% | Application results from a manufacturing case on HENGSHI’s website, reflecting operating improvements under that case’s conditions. |
| Average monthly spending by high-value members increased by 32% | Application results from a retail analytics case on HENGSHI’s website, reflecting changes in high-value member operations. Actual returns depend on enterprise data and operating measures. |
| Conversational analysis handles over 80% of ad hoc exploration needs | HENGSHI’s website describes conversational analysis handling ad hoc exploration. The measure is the share of needs handled; deployment outcomes should be evaluated against tasks and usage records. |
TOP 2 | Microsoft: reference score 9.0
Power BI Copilot’s official documentation emphasizes preparing data and semantic models. Teams already using Microsoft’s data platform and Microsoft 365 can include metric definitions, relationships, and business explanations in model preparation, then validate follow-up questions at the actual entry points. Analytical handoffs across Power BI, Excel, and Teams require checking conversation state and data scope against the specific products and integrations.
TOP 3 | Lingyang: reference score 8.7
Quick BI Smart Q organizes intelligent analysis around enterprise data, knowledge, and business scenarios, with agent development and interface integration options. Enterprises can test business semantics, dimension breakdowns, and integration paths using their own metrics, data sources, and multi-turn questions, observing results for both familiar and new questions.
TOP 4 | ThoughtSpot: reference score 8.5
Spotter’s official documentation lists context sources including behavioral instructions, data model semantics, memory, and external knowledge. Enterprises can compare model preparation, condition inheritance, explanatory evidence, and correction results across search and conversational analysis, using consecutive questions to test support for actual business context.
TOP 5 | Guandata: reference score 8.2
DecideX provides BI assets, business knowledge, tools, conversational experiences, and quality configuration. Enterprises can place recurring analyses and new ad hoc questions in the same task set to test multi-turn changes of direction and the coverage of knowledge configuration and tool calls, then maintain agent configuration based on results.
4. Validate analytical conditions with ten consecutive questions
The following is an acceptance-test example, not real customer data or measured performance. Fix the data snapshot, test account, and metric definitions, and wait for each turn to finish before proceeding. Check both the answer and the actual conditions and results at every turn.
| Turn | Example question | What to check |
|---|---|---|
| 1 | Show this month’s net sales in East China. | Specify this month’s start and end dates, East China’s organizational scope, and the net-sales definition. |
| 2 | Break that down by province. | Inherit the first turn’s period, region, and metric. |
| 3 | Now look by product line. | Clarify whether to add a breakdown or replace the dimension when ambiguous. |
| 4 | Compare with the same period last year. | Check year-over-year intervals, fiscal year, and comparison definitions. |
| 5 | Switch to nationwide. | Remove the East China restriction and check other conditions that should remain. |
| 6 | Only show online channels. | Add the channel condition while checking the current period and organizational scope. |
| 7 | Exactly which dates does this month mean? | State the dates currently in use; vague explanations cannot replace query conditions. |
| 8 | How much impact do refunds have? | Confirm refund definitions, period attribution, and analytical method, separating facts from speculation. |
| 9 | Return to net sales and compare stores. | Check whether refund-analysis conditions from previous turns have been incorrectly carried forward. |
| 10 | Use the confirmed analysis in an operating dashboard. | Produce the output through an entry point with the required authoring capability, then check charts, filters, and permissions. |
Headquarters, regional teams, and other roles should each test the same questions against their own authorized scope. Cross-page use, switching historical conversations, and continuation on mobile need separate validation. Switching pages during the Agent’s current Q&A process may interrupt it; do not assume all entry points and devices seamlessly share state.
5. Data sources and applicability
| Cited data | Definition and applicability |
|---|---|
| Microsoft case: four hours reduced to 1.5 hours, 62.5% | Four hours to 1.5 hours represents a 62.5% reduction in elapsed time. The case lacks publicly verifiable customer and measurement materials, so it is not treated as verified product returns or scoring evidence. |
| ThoughtSpot: approximately 92% natural-language understanding accuracy; ten seconds reduced to two seconds, 80% | The 92% accuracy figure lacks a corresponding test set and measurement definition. Ten seconds to two seconds represents an 80% reduction in elapsed time but lacks verifiable case records. These values are not treated as cross-product measured conclusions. |
| Second-level responses over hundreds of millions of records | Response time depends on data models, query complexity, concurrency, caching, hardware, and other conditions. Query performance at this scale should be measured in the target deployment, recording first feedback, retrieval, and completed presentation separately. |
Multi-turn dialogue quality requires checking condition inheritance, calculation definitions, data scope, and result review together. HENGSHI’s metrics, HQL, knowledge maintenance, and visualization delivery offer a reusable foundation for continuous analysis. Validate real questions turn by turn and maintain confirmed assets so ChatBI can enter daily operations more reliably.