ASAPO: A Framework for Efficient Text-to-SQL Generation through Adaptive Schema Filtering

Large Language Models (LLMs) have demonstrated remarkable capabilities in Text-to- SQL tasks, yet their deployment faces significant challenges due to token consumption and associated costs. We present ASAPO, a novel framework for efficient Text-to-SQL generation that reduces token usage by up to ۶۸% while maintaining competitive accuracy on the Spider benchmark. ASAPO employs dynamic schema filtering, relevance-based table selection, and adaptive few-shot example retrieval to minimize prompt size without sacrificing semantic understanding. Our experiments demonstrate that ASAPO achieves ۸۲.۳% execution accuracy with an average of ۸۴۷ tokens per query, compared to over than ۲,۵۰۰ tokens in recent approaches. The framework is model- agnostic and integrates seamlessly with existing LLM-based Text-to-SQL systems.

کلیدواژه ها:

Text-to-SQL ، Large Language Models ، Prompt Optimization ، Few-Shot Retrieval ، Schema Filtering ، Iterative Refinement ، Token Efficiency ، ASAPO Framework