📄 中文摘要
该项目展示了将大型语言模型（LLM）直接应用于数据库模式以生成 SQL 查询在企业用例中是根本性失败的原因。核心论点强调，缺乏语义元数据层（如本体、知识图谱或模式注释），LLM 只能在商业语义上进行猜测，且往往会出错。研究表明，随着模式复杂性的增加，文本到 SQL 的准确性急剧下降，这一现象在 Spider 基准测试和 Bird 基准测试中得到了验证。
文本到 SQL 的失败示例
发布: 2026年3月4日
📄 English Summary
Text-to-SQL Failure Demo
This project illustrates the fundamental failure of naively applying a large language model (LLM) to generate SQL queries from database schemas in enterprise use cases. The central argument emphasizes that without a semantic metadata layer, such as an ontology, knowledge graph, or schema annotations, an LLM is merely guessing business semantics, which often leads to errors. Research indicates that the accuracy of Text-to-SQL drops sharply as schema complexity increases, a trend validated by the Spider and Bird benchmarks.