Semantic Layer

AI governance encompasses the policies, processes, and technical controls organizations apply when using AI systems, including data classification access control, manual review of model outputs, security scanning of AI-generated code, token cost quota management, and compliance audit logs.

The semantic layer is an abstraction layer situated between raw data storage and business applications, transforming technical database structures into business-friendly conceptual models and defining unified calculation logic for business metrics.

语义层描述约4KB即可包含核心业务指标定义、维度层级关系、业务规则和字段别名映射等信息

构建结构化语义层比模型微调或复杂Prompt工程成本更低、可维护性更强，且对不同模型都能产生正向效果

A semantic layer is an abstraction layer between the data warehouse and upper-layer applications, defining the query interface in business terms rather than physical table columns, with benefits including governance, reuse, and decoupling.

In essence, the semantic layer establishes a translation dictionary between raw data tables and final queries, encapsulating SQL fragments as business terminology.

The semantic layer pre-defines metrics, dimensions, and business rules, giving AI a clear frame of reference when parsing natural language.

The semantic layer has existed within the OLAP (Online Analytical Processing) ecosystem since the early days of BI tool development.

Databricks Genie introduces the concept of a semantic layer, which reduces the probability of the model generating incorrect queries by pre-defining business metric definitions and relationships between tables.

Traditional semantic layers face structured queries (SQL) with clear rule boundaries, whereas LLMs receive natural language intent, whose boundaries are ambiguous and dynamically changing.

The concept of a semantic layer has existed in the data analytics field for many years, typically used to define business metrics, dimensions, and measures so that different tools have a unified understanding of concepts.

The Semantic Layer establishes a mapping between raw database fields and business language; without it, Agents can produce inconsistent results due to variations in how questions are phrased.

Using an LLM shifts developers from the role of 'thinking about how to implement' to the role of 'reviewing and patching AI output'.

When querying data, AI data agents interact with the semantic layer rather than directly manipulating underlying SQL tables, thereby mapping natural language questions to validated, standardized metric definitions.

dbt, Looker Semantic Layer, and Cube.js are the mainstream choices enterprises currently use to build semantic layers.

When user questions involve specific metric definitions or business terminology internal to an organization, models lacking semantic-layer metadata context tend to generate queries that are syntactically correct but semantically misaligned with business intent.

Modern data platforms such as dbt's Metrics layer and Looker's LookML are exploring the standardization of the semantic layer.

The same business term (such as active users or net revenue) can have completely different definitions and calculation methods across different teams and different systems.

The concept of the semantic layer originated in the business intelligence (BI) domain and was popularized by tools such as Looker.

Current mainstream semantic layer implementations include the dbt Metrics Layer, Cube.js, and proprietary semantic layer solutions developed by various cloud vendors.