Text-to-SQL is a definition problem: build the canonical model first

Text-to-SQL challenges stem not from the inability to write SQL but from the lack of understanding of data semantics, necessitating a canonical knowledge layer that defines data meanings before building models. This approach suggests prioritizing the creation of a canonical model that acts as a knowledge graph, providing a structured understanding of data relationships and definitions to guide both model building and query answering. By integrating the canonical knowledge layer at the outset, data engineering becomes more efficient, reducing the drift between the data model and its semantic understanding, and enhancing the reliability of automated query responses. This method aligns with the practices of organizations like Anthropic, which have demonstrated significant improvements in query accuracy by embedding data semantics into their analytics processes. The strategy is to write the definitions first, allowing for accurate, consistent answers that stakeholders can verify, ultimately streamlining the path from question to answer and enabling self-service analytics with built-in checks for accuracy and relevance.