AV-SQL: Decomposing Complex Text-to-SQL Queries with Agentic Views
The rapid evolution of artificial intelligence has led to significant advancements in the field of natural language processing, particularly in the area of Text-to-SQL translation. This task, which involves converting natural language queries into executable SQL statements for structured databases, aims to empower non-expert users to access and manipulate data without needing to master SQL syntax. However, as highlighted in the preprint article “AV-SQL: Decomposing Complex Text-to-SQL Queries with Agentic Views” (arXiv:2604.07041v1), challenges remain, particularly when dealing with complex queries in real-world scenarios.
Traditional approaches utilizing large language models (LLMs) have demonstrated promising results; however, they often falter when confronted with intricate queries that require multi-step reasoning and navigation through extensive database schemas. In many instances, the complete database schema exceeds the context window of the model, leading to difficulties in generating correct SQL queries. Additionally, one-shot generation frequently results in non-executable SQL due to syntax errors and incorrect schema associations.
Introducing AV-SQL
To tackle these challenges, the authors propose AV-SQL, an innovative framework designed to decompose complex Text-to-SQL tasks into a structured pipeline involving specialized LLM agents. A core component of AV-SQL is the introduction of “agentic views,” which are agent-generated Common Table Expressions (CTEs). These CTEs encapsulate intermediate query logic and filter relevant schema elements, thus simplifying the process of generating accurate SQL from complex natural language queries.
How AV-SQL Works
The AV-SQL framework operates in three distinct stages:
- Rewriter Agent: This agent is responsible for compressing and clarifying the input query, enhancing its clarity and focus.
- View Generator Agent: This agent processes chunks of the schema to produce agentic views, which encapsulate relevant elements and logic necessary for the final SQL query.
- Collaborative Agents: A planner, generator, and revisor agent work together to compose these agentic views into a coherent and executable SQL query.
Performance and Results
Extensive experiments conducted with AV-SQL demonstrate its effectiveness in generating executable SQL queries. The framework achieved an impressive execution accuracy of 70.38% on the challenging Spider 2.0 benchmark, surpassing existing state-of-the-art methods. Furthermore, AV-SQL remains competitive on standard datasets, achieving accuracy rates of 85.59% on Spider, 72.16% on BIRD, and 63.78% on KaggleDBQA.
The source code for AV-SQL is publicly available, inviting further exploration and development within the academic and professional communities. For those interested in delving deeper into the capabilities of AV-SQL, the code can be found at GitHub – AV-SQL.
Conclusion
AV-SQL represents a significant advancement in the field of Text-to-SQL translation, particularly for complex queries. By employing a structured approach that leverages agentic views, the framework not only improves execution accuracy but also enhances the accessibility of structured data for non-expert users. As the demand for data-driven insights continues to grow, innovations like AV-SQL are poised to play a crucial role in bridging the gap between natural language and database interaction.