[Paper Review] CORE-T: COherent REtrieval of Tables for Text-to-SQL

Hassan Soliman, Vivek Gupta+2 more|TUbilio (Technical University of Darmstadt)|Jan 19, 2026

Data Quality and Management0 citations

TL;DR

CORE-T is a training-free framework that retrieves a coherent, joinable subset of tables from pooled open-book corpora for text-to-SQL by offline enrichment and a single-shot LLM-based selection with a compatibility-based restoration step.

ABSTRACT

Realistic text-to-SQL workflows often require joining multiple tables. As a result, accurately retrieving the relevant set of tables becomes a key bottleneck for end-to-end performance. We study an open-book setting where queries must be answered over large, heterogeneous table collections pooled from many sources, without clean scoping signals such as database identifiers. Here, dense retrieval (DR) achieves high recall but returns many distractors, while join-aware alternatives often rely on extra assumptions and/or incur high inference overhead. We propose CORE-T, a scalable, training-free framework that enriches tables with LLM-generated purpose metadata and pre-computes a lightweight table-compatibility cache. At inference time, DR returns top-K candidates; a single LLM call selects a coherent, joinable subset, and a simple additive adjustment step restores strongly compatible tables. Across Bird, Spider, and MMQA, CORE-T improves table-selection F1 by up to 22.7 points while retrieving up to 42% fewer tables, improving multi-table execution accuracy by up to 5.0 points on Bird and 6.9 points on MMQA, and using 4-5x fewer tokens than LLM-intensive baselines.

Motivation & Objective

• Address the bottleneck of retrieving relevant and joinable tables in open-book text-to-SQL over large heterogeneous table collections without db_id signals.
• Enrich tables offline with purpose metadata and build a lightweight table–table compatibility cache to approximate joinability.
• Provide a scalable online pipeline that uses dense retrieval for high recall, a single LLM pass for coherent subset selection, and a restoration step to recover strongly compatible tables.

Proposed method

• Offline, table enrichment with LLM-generated purpose descriptions and 5-row Markdown snapshots to create a table embedding for dense retrieval (embedding f_tbl).
• Column-level signals (header embeddings, value overlap, and relational constraints) are used to build a table–table compatibility cache CS(t_i, t_j) approximating joinability.
• Online, retrieve top-K tables via dense retrieval on enriched embeddings (RS(q,t) = cosine(e_q, e_t)).
• A single LLM pass acts as a schema analyst to select a coherent, joinable subset from the top-K candidates using provided compatibility evidence and sampled table metadata.
• An additive adjustment step restores strongly compatible tables from the original top-K using the compatibility cache, yielding the final set S(q).
• All steps are designed to be training-free and to minimize LLM usage while maximizing join coherence.

Experimental results

Research questions

• RQ1 How can open-book multi-table retrieval be made join-aware without explicit db_id or gold foreign keys?
• RQ2 Does offline enrichment plus a lightweight online LLM selection improve table coherence and end-to-end SQL execution over pooled corpora?
• RQ3 What is the impact of a compatibility-based restoration step on recall and downstream SQL accuracy?
• RQ4 How does CORE-T’s efficiency compare to iterative or MIP-based join-aware baselines in open-book settings?

Key findings

• CORE-T improves table-selection precision and F1 while retrieving fewer tables compared to strong baselines across Bird, Spider, and MMQA.
• A single LLM selection plus a lightweight restoration step yields more coherent schema slices and better multi-table execution (EM ≥2T) than dense retrieval alone.
• CORE-T achieves higher end-to-end SQL execution accuracy on multi-table questions across different SQL generators, reducing headroom relative to oracle gold-table settings.
• The approach reduces LLM usage by about 4–5x in selection and shows up to ~5x token savings compared to heavier multi-draft baselines.
• Compared to join-aware methods that require db_id or MIP optimization, CORE-T offers competitive or superior table-set quality with substantially lower inference overhead.