POC
DataAgent数据智能平台
面向企业经营分析与智能问数场景，基于 Ontology 语义模型建设可追溯、可复用的数据分析 Agent 平台。
系统快照
架构、信号与证据会先于正文呈现。- Scenario
- NL2Analysis
- Model
- Ontology
- Validation
- POC
Problem
企业经营分析中的数据口径、跨表关系和指标定义往往很复杂。传统 NL2SQL 可以回答一部分查询，但在深度分析、归因诊断和可追溯决策支持上不够稳定。
DataAgent 面向智能问数与经营分析场景，基于 Ontology 语义模型，将自然语言问数进一步升级为可追溯的智能分析与决策能力。产品官网可以从 产品官网 查看。
Constraints
- 库表 Schema、指标口径、计算逻辑和分析维度需要沉淀成系统资产，而不是散落在提示词里。
- 生成的 SQL 与分析结论必须经过 SQL 真实执行与结果校验。
- 面向客户 POC 时，需要支持配置化接入、异常处理、敏感词过滤和请求校验。
System Design
平台后端围绕库表接入、Schema 元数据解析、指标管理、分析框架和任务执行进行模块拆分，明确服务边界，让数据分析能力可以平台化扩展和复用。
语义层通过 Ontology 语义模型沉淀业务指标、计算口径与分析维度。Agent 执行链路打通自然语言理解、分析框架匹配、SQL 生成与执行、结果校验和报告输出。
Key Decisions
最关键的判断是不要把模型生成的自由文本直接当成最终分析结果。系统通过 SQL 真实执行与结果校验构建 grounding loop，把模型输出约束在可验证的数据链路里。
归因分析、异常诊断等能力也不应只写成提示词，而应该抽象成可执行、可配置、可复用的分析框架。
Impact
- 完成 DataAgent 核心后端与数据分析链路建设。
- 将库表元数据、指标体系与分析框架沉淀为可配置、可复用的平台能力。
- 面向多家意向客户开展 POC 验证，覆盖智能问数与复杂经营分析场景，初步验证跨客户复制能力。
Reflection
下一步更值得投入的是指标治理、客户语义模型配置、评测集和分析师复核流程，让 DataAgent 从 POC 能力进一步走向可规模化交付的产品能力。
Problem
Enterprise analytics depends on complex metrics, table relationships, and business definitions. Traditional NL2SQL is often too unstable for deeper analysis.
The goal was to move from natural-language data querying toward traceable analysis and decision support.
Constraints
Database schemas, metric definitions, calculation logic, and analysis dimensions must be reusable system assets.
Generated SQL and analysis conclusions need validation through real execution and result checks.
The platform needs enough governance to handle request validation, sensitive-word filtering, exceptions, and customer-specific configuration.
System Design
The backend separates table access, schema metadata parsing, metric management, analysis frameworks, and task execution into clear service boundaries.
At the semantic layer, business metrics and dimensions are modeled through an ontology so the Agent can match user intent to reusable analysis methods.
The execution chain connects natural-language understanding, framework matching, SQL generation, SQL execution, result validation, and report output.
Key Decisions
The system should not trust free-form model output as the final answer. SQL execution and result checks form a grounding loop.
Attribution analysis and anomaly diagnosis are treated as executable analysis frameworks rather than one-off prompt templates.
Impact
The core backend and data-analysis workflow were built as configurable platform capabilities.
The project has been used for POC validation with multiple interested customers, covering intelligent data Q&A and complex business analysis.
Reflection
The next layer should strengthen metric governance, customer-specific semantic modeling, evaluation sets, and analyst review loops.