In retrieval-augmented generation (RAG), tables serve as crucial knowledge sources, but a retrieved table may lack sufficient evidence to answer a query, a property termed answerability. While answerability broadly concerns whether a source or collection contains sufficient evidence, retrieval models optimized for semantic relevance do not guarantee this even in single-source cases, leading to a fundamental mismatch.
To address this, we introduce TCR-Bench, a diagnostic benchmark for Table Content-level Answerability in RAG, built around sibling tables—tables with highly similar schemas but subtle content differences. On TCR-Bench, the dense retrievers we evaluate consistently exhibit a Semantic-Answerability Gap: they often retrieve the correct sibling group yet struggle to identify the uniquely answerable table within it, dropping QA performance from 0.755 (oracle) to 0.330 (top-5 retrieved).
Our analysis suggests this gap is associated with semantic accumulation, schema-level cue dependence, and weak row-column binding. As a diagnostic probe into the source of this gap, we test a lightweight two-stage pipeline, Answerability-Aware Reranking (AAR), which applies direct query-table answerability judgment to recover performance: it raises top-1 target retrieval from 18.2% to 57.4%, and this significant gain itself indicates that much of the observed failure reflects a missing answerability verification step, rather than an inherent limitation of model capacity alone.
Blogger's Review: This article introduces the TCR-Bench benchmark to analyze the semantic answerability issues in table RAG, revealing deficiencies in retrieval models when handling similar data. The proposed answerability-aware reranking method offers a new perspective for enhancing model performance, warranting further exploration and practical application.