Dynamic-SQL: an adaptive NL2SQL framework with multi-path fusion reasoning and execution feedback correction.

Area of Science:

• Artificial Intelligence
• Database Management
• Natural Language Processing

Background:

• Current Natural Language to SQL (NL2SQL) systems struggle with practical constraints like limited prompt length and inability to fine-tune large language models (LLMs).
• Performance degradation is significant in complex databases due to inaccurate schema linking, vague value conditions, and weak self-correction.

Purpose of the Study:

• To introduce Dynamic-SQL, an adaptive framework designed to enhance NL2SQL system performance under practical constraints.
• To address limitations in schema linking, value conditioning, and self-correction within complex database querying scenarios.

Main Methods:

• Developed an adaptive framework coupling multi-path chain-of-thought fusion with execution-based feedback correction.
• Constructed a dense-sparse hybrid vector space for dynamic schema element retrieval.
• Leveraged large language models (LLMs) to generate explicit schema subgraphs and enriched prompts with real-value and few-shot exemplars.
• Implemented diverse reasoning paths for candidate SQL generation, fused chains of thought, and utilized execution feedback for iterative self-correction.

Main Results:

• Reduced average prompt length by 50.83% on the BIRD benchmark.
• Improved strict schema-linking recall from 72.63% to 90.66%.
• Achieved 63.23% execution accuracy using the qwen2.5-coder-32b-instruct model.

Conclusions:

• Dynamic-SQL offers a transferable paradigm for deploying LLMs in complex database querying.
• The framework systematically addresses key challenges in schema linking, exemplar augmentation, multi-path reasoning, and self-correction.
• This approach significantly enhances the robustness and accuracy of NL2SQL systems in real-world applications.