KADP-SQL: Knowledge-Augmented Generation and Dual-Path Validation in Text-to-SQL

The widespread application of Large Language Models (LLMs) has significantly advanced the development of the Text-to-SQL task, which aims to translate natural language questions posed by users into executable structured query statements. This allows non-expert users to efficiently access databases. However, existing models still struggle to generate correct SQL statements in the absence of sufficient knowledge, particularly under complex query scenarios, where they are prone to syntactic errors and semantic deviations. Moreover, most current approaches rely solely on execution feedback for SQL correction via a single path, making it difficult to detect semantic errors. To address these challenges, we propose KADP-SQL, which divides the Text-to-SQL task into two main modules: Knowledge-Augmented Generation and Dual-Path Validation. This framework structurally represents manually annotated evidence and incorporates web-based external knowledge to enhance SQL generation. Additionally, we introduce a dual-path SQL validation method to detect and correct errors in generated SQL queries. Through extensive experiments conducted on multiple closed-source and open-source LLMs, our proposed KADP-SQL achieves an execution accuracy of 70.53% on the BIRD development set and 88.22% on the Spider test set. These results demonstrate the effectiveness and adaptability of the proposed method.