📄 中文摘要
在处理 LLM 生成的 SQL 查询时，常常会花费大量时间追踪错误原因。将自然语言请求交给模型后，返回的 SQL 字符串是否符合预期结果至关重要。通过对话，作者获得了一系列灵感，包括确定性检查（行数、列覆盖）、AST 比较的深层语义分析，甚至是一个 AI “评审”来指出缺失或多余的部分。该框架整合了这些元素，提供了可复制的最小代码、批量处理数百个查询的方法，以及 LLM 层提供的可操作反馈，无需仪表板。
构建生产就绪的 SQL 评估引擎与 LLM
出处: Build a Production‑Ready SQL Eval Engine with LLMs
发布: 2026年3月30日
📄 English Summary
Build a Production‑Ready SQL Eval Engine with LLMs
The challenges of debugging LLM-generated SQL queries often lead to significant time investment in identifying errors. The key concern is whether the SQL string returned from a natural language request aligns with expected results. Insights gained from discussions include deterministic checks such as row counts and column coverage, deeper semantic analysis through AST comparisons, and the introduction of an AI 'judge' to highlight missing or extraneous elements. This framework integrates these aspects, offering minimal code for easy replication, methods for bulk processing of hundreds of queries, and actionable feedback from the LLM layer without the need for dashboards.