Text-to-SQL

Text-to-SQL助手的核心工具链由四个Tool组成：ListTables、GetSchema、VerifyQuery和ExecuteQuery

Early Text-to-SQL systems relied on rule matching and template filling, with limited accuracy and difficulty handling complex queries

大模型在Text-to-SQL场景中面临幻觉问题，可能生成语法正确但表名/列名错误的SQL

GPT-4等大模型在零样本（zero-shot）条件下即可生成高质量SQL

Text-to-SQL is a research direction in NLP with the core goal of automatically converting natural language questions into SQL query statements executable on relational databases

The paper proposes the first schema-documentation-free Text-to-SQL framework, which instead learns database structure from historical query logs.

The success of existing Text-to-SQL methods on academic datasets is built on the premise that the database has complete schema documentation.

Intent Classification is a fundamental task in natural language processing, referring to determining the semantic category of user input, and often serves as the first gate before query generation in Text-to-SQL scenarios.

Schema selection (schema selection/schema linking) is a front-end step of the Text-to-SQL pipeline, is essentially a retrieval and ranking task, and is more sensitive to the model's instruction following, output format stability, and attention allocation over long lists.

In Text-to-SQL systems, combining static syntax checks, execution plan review, and LLM semantic checks into a multi-layer defense is more reliable than a single checker.

The SIGMOD'27 paper 'Benchmarking Text-to-SQL under Role-Based Access Control' (arXiv 2607.22115) found that under RBAC, showing the model only the columns it is authorized to view does not stop a Text-to-SQL agent from still piecing together an answer when the answer requires hidden columns, rather than refusing to answer.

A SIGMOD'27 paper released a public benchmark for Text-to-SQL in RBAC scenarios.

Schema grounding is a key step in Text-to-SQL, referring to the system mapping business terms in natural language onto the database's actual tables, fields, primary and foreign key relationships, and constraints, usually accomplished with the help of metadata catalogs, field comments, and business glossaries, and it is a prerequisite for the correctness of the subsequent SQL.

Databricks Genie falls within the category of Text-to-SQL or the broader natural language interface (NL2Data) technology; its underlying logic is to use a large language model to translate natural language questions into structured query statements and execute them on a data warehouse or data lake.

The core challenge of natural language data query tools lies in accurately understanding ambiguous terms in business contexts, where the same term may have different definitions across different departments.

Genie combines "Text-to-SQL" and "Text-to-Insight" capabilities in its technical architecture, converting natural language questions into SQL and retrieving results from the underlying data lake or data warehouse.

The core mechanism by which AI agents process data is converting natural language questions into SQL queries or API calls, a process known as Text-to-SQL or NL2SQL.

Research shows that Text-to-SQL accuracy on real enterprise databases is far lower than on standardized benchmarks, with non-standard field naming and missing business semantics being one of the primary causes.

The implementation path for applying LLMs to structured data queries typically involves first translating natural language into structured query statements (such as SQL or API calls), and then having the LLM semantically interpret the results—a path also known as Text-to-SQL or NL2Query.

Natural language query (Text-to-SQL) functionality introduces new attack vectors such as prompt injection and indirect data leakage.