Multi-pattern retrieval-augmented framework for Text-to-SQL with Poincaré-Skeleton retrieval and meta-instruction reasoning

Chunxi Guo, Zhiliang Tian, Jintao Tang, Shasha Li, Ting Wang

Abstract: Highlights•We propose a multi-pattern framework to enhance SQL generation, leveraging dynamic demonstration selection based on query and reasoning patterns.•We use the Poincaré ball model to project sentence embeddings into hyperbolic space, capturing the semantic hierarchy of language and database schemas more effectively.•To handle complex SQL queries, we break the process into meta-instructions, creating a multi-category repository of Chain-of-Thought fragments to aid SQL generation.•We tackle potential biases and inaccuracies in SQL generation by implementing a revision mechanism. It allows the LLM to interact with databases, interpret its own output, and make necessary corrections, thus improving the overall reliability.•Ours demonstrates superior generalization capabilities, outperforming existing models in SQL generation tasks.