Research Papers 论文研究 7h ago Updated 2h ago 更新于 2小时前 47

GROUND: Reducing Hallucinations in LLM-Based Enterprise Analytics Through Governed Semantic Definitions GROUND:通过治理语义定义减少LLM企业分析中的幻觉

GROUND (Governed Retrieval Over Unified Normalized Definitions) is a framework that constrains LLM-generated analytics to a governed semantic layer, addressing hallucinations in enterprise text-to-SQL systems. The system validates generated SQL against six rule categories—schema, metric, join, grain, filter, and row-level security—before execution, retrying or abstaining on violations. In a 100-question synthetic benchmark, GROUND was the only system free of hallucinations across all six evaluat GROUND框架通过治理语义层约束LLM生成的企业分析SQL,解决幻觉指标、无效连接、错误粒度、不安全数据访问等生产级问题 在100题合成基准测试中,GROUND是唯一在所有六个评估类别中零幻觉的系统,而未治理系统多次违反行级安全策略 仅依赖精确指标定义无法替代完整治理,缺少访问控制策略仍会导致数据泄露 在真实NHTSA车辆安全数据和对抗性测试中,GROUND的过滤器和行级安全保证在所有四个模型上均实现零违规

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • GROUND (Governed Retrieval Over Unified Normalized Definitions) is a framework that constrains LLM-generated analytics to a governed semantic layer, addressing hallucinations in enterprise text-to-SQL systems.
  • The system validates generated SQL against six rule categories—schema, metric, join, grain, filter, and row-level security—before execution, retrying or abstaining on violations.
  • In a 100-question synthetic benchmark, GROUND was the only system free of hallucinations across all six evaluated categories, while ungoverned systems violated row-level security.
  • Semantic-only grounding (exact metric definitions without access policies) still leaked data, proving that governance cannot be replaced by metric fidelity alone.
  • Zero violations were observed on enforced guarantees (filters and row-level security) across four models from three providers, though judgment-dependent behaviors like refusing undefined metrics remain fallible.

Why It Matters

This work directly addresses the primary barrier to production deployment of LLM-based analytics: hallucinated or unsafe SQL generation. For AI practitioners building enterprise analytics tools, GROUND demonstrates that semantic governance and access control must be treated as first-class constraints, not afterthoughts. The findings challenge the common assumption that RAG over documentation or semantic layer definitions alone are sufficient for safe enterprise deployment.

Technical Details

  • GROUND introduces a governed semantic layer that supplies approved metric definitions, dimensions, join paths, filters, and row-level security policies, binding user intent to these governed constructs before SQL generation.
  • A multi-rule validation engine checks generated SQL against schema correctness, metric fidelity, valid join paths, correct grain, filter compliance, row-level security enforcement, and cost rules, with retry or abstention on violations.
  • Evaluation includes a 100-question synthetic enterprise-reporting benchmark, replication on real U.S. NHTSA vehicle-safety data with hand-authored gold standards, and adversarial testing across four models from three providers.
  • Comparative baselines include direct schema-only text-to-SQL, schema-RAG, and semantic-only grounding, all under a shared model to isolate the effect of governance.

Industry Insight

  • Enterprise AI teams should invest in governed semantic layers with explicit access policies rather than relying solely on schema grounding or RAG; metric accuracy without security governance is insufficient.
  • The zero-violation result on enforced guarantees across diverse models suggests that pre-execution validation frameworks can make LLM-based analytics production-ready, but organizations should still account for fallibility in judgment-dependent behaviors like metric refusal.
  • The adversarial testing results highlight the need for robust security validation in any text-to-SQL pipeline, as even semantically correct systems can leak data without explicit row-level security enforcement.

TL;DR

  • GROUND框架通过治理语义层约束LLM生成的企业分析SQL,解决幻觉指标、无效连接、错误粒度、不安全数据访问等生产级问题
  • 在100题合成基准测试中,GROUND是唯一在所有六个评估类别中零幻觉的系统,而未治理系统多次违反行级安全策略
  • 仅依赖精确指标定义无法替代完整治理,缺少访问控制策略仍会导致数据泄露
  • 在真实NHTSA车辆安全数据和对抗性测试中,GROUND的过滤器和行级安全保证在所有四个模型上均实现零违规

为什么值得看

本文针对企业级LLM分析系统的核心痛点——幻觉与数据安全——提出了可验证的治理方案,为text-to-SQL从实验走向生产提供了关键路径。研究结果明确区分了"语义准确性"与"治理合规性",对构建可信AI分析系统具有重要参考价值。

技术解析

  • GROUND(Governed Retrieval Over Unified Normalized Definitions)框架将LLM生成的分析约束在治理语义层,提供经批准的指标、维度、连接路径、过滤器和行级安全定义,并在执行前对生成的SQL进行模式、指标、连接、粒度、过滤器、安全和成本规则的全面验证
  • 基准测试采用100题合成企业报告数据集,对比直接模式text-to-SQL、模式RAG和仅语义 grounding 三种方案,评估维度涵盖幻觉指标、无效连接、错误粒度、不安全访问等六类问题
  • 实验验证在真实美国NHTSA车辆安全数据上进行,使用独立手工编写的gold标准,并在来自三个提供商的四个模型上进行对抗性测试,验证框架的泛化能力
  • 关键发现:仅语义条件(精确指标定义但无访问策略)仍会泄露数据,证明治理不能仅靠指标保真度替代;强制执行保证(过滤器和行级安全)在所有模型上保持零违规,但依赖判断的行为(如拒绝未定义指标)仍存在不确定性

行业启示

  • 企业级LLM分析系统必须将治理层(治理语义、访问控制、安全策略)与生成层分离设计,仅优化生成准确性无法解决生产环境的安全与合规问题
  • 在构建text-to-SQL系统时,应建立"验证-重试/放弃"机制,对生成的SQL进行多维度规则校验,而非依赖模型自身能力保证正确性
  • 行业需要建立企业分析场景的标准基准和评估体系,区分"语义正确性"与"治理合规性"两个独立维度,推动可信AI分析工具的发展

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Research 科学研究 Security 安全 Evaluation 评测