Reproduce “REaR: Retrieve, Expand and Refine for Effective Multitable Retrieval” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rishita Agarwal, Himanshu Singhal, Peter Baile Chen, Manan Roy Choudhury, Dan Roth, Vivek Gupta

The REAR framework introduces an LLM-free method for multi-table retrieval in Text-to-SQL that jointly optimizes query-table relevance and table-table joinability. It achieves competitive SQL execution accuracy compared to state-of-the-art LLM-based systems while reducing token consumption by 92.24% on the BIRD dataset.

Key insights

  • Jointly optimizing for both query-table relevance and table-table compatibility is critical for achieving high-fidelity multi-table retrieval in Text-to-SQL systems.
  • Efficient, LLM-free methods, utilizing precomputed embeddings and specialized rerankers, can effectively model complex relational reasoning without the prohibitive cost and latency of online LLM calls.
  • A modular, multi-stage approach allows for targeted improvements in both recall (by expanding joinable candidates) and precision (by refining based on joint relevance and coherence) in multi-table contexts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.