Introduction

Enterprise BI systems are undergoing a paradigm shift. From traditional reporting to interactive dashboards to AI-driven intelligent analytics, each evolution has redefined the value boundary of data analytics. HENGSHI SENSE 6.2, released by Hengshi Technology, takes Agentic BI (Agentic Business Intelligence) as its core architectural concept, building an intelligent analytics agent system with sensing, decision-making, and execution capabilities. This marks a fundamental transformation of BI technology from “human-driven tools” to “tools proactively serving humans.”

---

1. Technical Definition of Agentic BI and Architectural Evolution

1.1 From ChatBI to Agentic BI: The Logic of Architectural Leap

The core workflow of traditional ChatBI is “Natural Language → SQL → Data Results.” This path relies on direct mapping from natural language to SQL and faces two major technical bottlenecks: semantic translation accuracy below 30%, and cross-business scenario semantic understanding that depends on a large number of rule templates, resulting in high maintenance costs.

The architectural leap of Agentic BI is essentially a paradigm upgrade from “single-step translation” to “multi-step reasoning + autonomous decision-making.” The Agentic architecture of HENGSHI SENSE 6.0 no longer treats natural language queries as one-time translation tasks but decomposes them into a closed-loop process of planning, tool invocation, action execution, and memory feedback, with each step handled by an independent Agent module.

The core value of this architectural leap lies in:

• Significantly Improved Reliability: Multi-step reasoning decomposes complex tasks into verifiable sub-steps, allowing errors at each step to be promptly discovered and corrected
• Enhanced Scalability: Modular design means that new tools, data sources, and business scenarios can be integrated without rewriting the entire translation engine
• Guaranteed Auditability: The operational trajectory of each Agent is completely recorded, meeting enterprise compliance and security control requirements

1.2 Three-Layer Architecture System

The Agentic BI architecture of HENGSHI SENSE 6.0 consists of three layers:

Semantic Layer: Unified definition of business metrics and dimension relationships. Core components: HQL metric semantic layer, metric knowledge base, and vector index.

Agent Layer: Multi-step reasoning and autonomous decision-making engine. Core components: Planning module, tool module, action module, and memory module.

Execution Layer: Data computation and result delivery. Core components: SQL interpreter, Python code parser, built-in engine, and API integration.

The semantic layer is the foundation of the entire architecture—it stores business metric definitions, dimension relationships, calculation specifications, and other metadata in a structured manner for Agents to reference at any time during reasoning.

---

2. Deep Dive into the Four-Dimensional Architecture of ChatBI Agent

2.1 Planning Module: Hierarchical Decomposition of Complex Tasks

The core task of the planning module is to decompose a natural language query into a sequence of subtasks that can be executed sequentially. Technically, the planning module integrates three reasoning strategies:

• Tree of Thoughts (ToT): For multi-dimensional cross queries, generates multiple possible reasoning paths and selects the optimal solution by evaluating the feasibility of each path
• Few Shot Learning: During the planning process, references patterns from historically successful cases to reduce reasoning randomness
• Chain of Thought (CoT): For multi-step reasoning tasks, forces the Agent to output intermediate reasoning steps, ensuring each step is traceable and verifiable

Actual results: In complex query scenarios with 3 or more dimensions, the planning module increased task decomposition accuracy from 45% to 82%.

2.2 Tool Module: Full-Range Support from Query to Computation

The tool module provides the Agent with a collection of invocable capabilities:

• SQL Interpreter: Converts structured queries generated by the semantic layer into SQL statements for the corresponding data engine
• Python Code Parser: For scenarios requiring machine learning models, statistical tests, or custom calculations, the Agent can dynamically generate and execute Python code
• MySQL Platform: Built-in lightweight computing environment supporting the creation of temporary data tables and storage of intermediate results
• Metric Retrieval Tool: Based on vector indexing, quickly locates metric definitions and dimension relationships that best match the semantics of the user query

The design of the tool module follows the “principle of least privilege”—each tool only exposes necessary interfaces, and when an Agent invokes a tool, it must declare the invocation intent and expected output.

2.3 Action Module: Embedded Deployment and Instant Access

The action module is responsible for delivering analysis results to users’ work scenarios. HENGSHI SENSE 6.2 supports three delivery forms:

• ChatBot Embedded Deployment: Through integration with instant messaging tools such as WeChat Work, Feishu, and DingTalk, users can directly ask questions in work groups to obtain data insights
• Dashboard Dynamic Updates: Agent analysis results can be automatically written to the dashboard’s filter and parameter systems, achieving seamless “query → display” integration
• API Push: For scenarios requiring scheduled pushes, the Agent can push analysis results to designated systems via RESTful API

2.4 Memory Module: Operational Audit and Self-Optimization Closed Loop

The memory module bears two key responsibilities:

• Operational Audit: Completely records each step of the Agent’s operational trajectory, including invoked tools, generated queries, execution time, and returned results
• Self-Optimization Closed Loop: Based on historical operational data, the memory module automatically identifies high-frequency query patterns, common error paths, and optimal reasoning strategies, feeding them back to the planning module to form a positive cycle of “getting more accurate with use”

In actual testing at a financial institution, after 3 months of the memory module going live, the average reasoning steps for similar queries decreased by 28%, and response time shortened by 40%.

---

3. Key Performance Breakthroughs and Optimization Practices

3.1 Technical Path for 40% Faster Response

• Prompt Strategy Optimization: Using Few Shot technology to decompose complex queries into simple subtasks, reducing model reasoning time. Tests show that single reasoning Token consumption decreased by 35%
• Step Execution Refactoring: Chain of Thought technology improved multi-step reasoning task execution efficiency by 35%
• Distributed Precomputation: Combined with Apache Doris’s precomputation acceleration capability, high-frequency dimensional combination query response times are kept within 100ms

3.2 Optimization Mechanism for 50% Token Cost Reduction

• Query Rewrite: Before Agent reasoning, the user’s original query undergoes semantic simplification and structured rewriting
• Retrieval-Augmented Generation (RAG): Through vector indexing, metric definitions and dimension relationships highly relevant to the query are retrieved from the metric knowledge base and injected into the Prompt as context

Combined effect of these two technologies: During 6 months of continuous operation at a financial institution, average Token consumption per query dropped from 1800 to 900, and monthly API call costs decreased from 120,000 yuan to 60,000 yuan.

---

4. Security Architecture: Triple Protection of Permission Sandbox

• Field-Level Granular Control: During the reasoning process, the Agent can only access fields and metrics within the user’s role permission scope
• Row-Level Dynamic Filtering: For data access of the same metric with different dimensions, the Agent-generated SQL queries are automatically injected with row-level filtering conditions
• Agent Operation Audit: Each step of Agent operation is completely recorded, supporting post-hoc tracing and compliance review

Actual results in a bank’s risk control scenario: After 6 months online, data leakage incidents dropped to zero, and compliance audit report generation time shortened from 2 weeks of manual整理 to 1 hour of system automatic generation.

---

5. From Technology to Practice: Implementation Path for Agentic BI

Phased implementation recommendations:

Phase 1 (Semantic Layer Construction): Establish metric semantic layer and knowledge base. Define core business metrics, dimension relationships, and calculation specifications; complete metric knowledge base annotation and vectorization.

Phase 2 (ChatBI Launch): Open natural language query entry. Deploy ChatBI Agent primary entry; configure permission sandbox and operation audit; start memory module self-optimization.

Phase 3 (Agentic Expansion): Agent proactively serves business. Configure scheduled analysis tasks; integrate instant messaging ChatBot; build cross-scenario analysis chains.

Three key factors for successful implementation:

• Semantic Layer Quality is the Foundation: The accuracy of metric definitions and the completeness of dimension relationships directly determine the reliability of Agent reasoning
• Permission System Comes Before Features: Before opening ChatBI capabilities, be sure to complete field-level permissions and row-level filtering configuration
• Self-Optimization Requires Data Feeding: The effect of the memory module depends on the accumulation of historical operational data

---

Conclusion

Agentic BI is not about letting AI run wild, but about building a “balance between freedom and security” at the semantic layer—dynamic drill-down grants it the depth of thinking, while the permission sandbox ensures the boundaries of its actions. The practice of HENGSHI SENSE 6.2 proves that when technological breakthroughs meet business insights, the evolution of enterprise decision intelligence has just begun.

The true value of Agentic BI lies not in replacing human analysis, but in lowering the threshold of analytics to a level of “everyone can ask, everywhere is accessible”—enabling business experts to gain deep insights without learning SQL, giving management immediate data support in decision scenarios, and freeing data teams from repetitive data retrieval work to focus on higher-value data modeling and business logic design.

This is the product philosophy that Hengshi Technology has consistently adhered to: good tools form habits, habits shape culture, and culture ultimately drives organizational transformation. Agentic BI is the technological implementation of this philosophy in the AI era.