CORE-T introduces a training-free framework for improving table retrieval in text-to-SQL systems by combining dense retrieval with LLM-generated metadata and compatibility caching. The approach achieves significant performance gains—up to 22.7 points in table-selection F1 and 24.4 points in multi-table execution accuracy—while reducing inference tokens by 64-76% compared to LLM-intensive alternatives.
CORE-T addresses a critical challenge in database query systems: accurately identifying which tables from large, heterogeneous collections should be joined to answer natural language questions. Traditional dense retrieval methods achieve high recall but introduce noise through irrelevant results, while specialized join-aware approaches require additional infrastructure or computational overhead. The framework's innovation lies in its hybrid approach, leveraging LLM-generated purpose metadata to enrich table descriptions and pre-computing compatibility relationships without requiring model retraining.
The research emerges amid growing interest in making large language models more practical for enterprise applications. As organizations integrate LLMs into data pipelines, the ability to efficiently retrieve relevant database tables directly impacts system reliability and cost. Current solutions either sacrifice precision for recall or demand significant engineering overhead, creating friction in production deployments.
The practical implications are substantial. CORE-T's ability to reduce selection tokens by 1.64-4.20x translates directly to lower API costs for organizations using LLM-based database interfaces, while simultaneously improving accuracy. The training-free nature makes adoption frictionless—teams can implement the framework without model fine-tuning or retraining cycles. The consistent improvements across multiple benchmarks (Bird, Spider, MMQA, Beaver) demonstrate generalization across different database schemas and query patterns.
Looking forward, this work suggests a broader trend toward intelligent middle layers that optimize LLM-database interactions. As enterprises scale LLM applications, similar efficiency-focused retrieval improvements may become standard components of production systems, creating competitive advantages for teams who optimize token usage and accuracy simultaneously.
- →CORE-T achieves 22.7-point F1 improvement in table selection while reducing returned tables by 40% through LLM-generated metadata and compatibility caching
- →The framework uses 64-76% fewer tokens than LLM-intensive baselines, directly reducing inference costs for production deployments
- →Training-free design enables immediate adoption without model retraining or fine-tuning cycles
- →Consistent performance gains across four major benchmarks demonstrate broad applicability to diverse database schemas
- →Hybrid approach combining dense retrieval with LLM filtering balances recall and precision more effectively than single-method alternatives