Abstract
Questions that span relational tables and knowledge graphs (KGs) depend on a valid join between sources. When identifiers differ in format, hierarchy level, or vocabulary, exact matching can miss a valid link, while accepting a plausible but incorrect link can produce a wrong answer. We propose RelationalRAG-Tables , which uses a small learned verifier to guide this decision. Our planner checks each candidate join, tries another candidate after rejection, and abstains when uncertainty is high. We study how this use of existing matching methods affects question answering (QA), supported by two conditional theorems on recall gains and the utility of abstention. Our evaluation spans 20 synthetic seeds, constructed medical records linked to Wikidata, BIRD value linking and query generation, and HybridQA evidence retrieval. On BIRD with corrected query-alias resolution, our verifier recovers drifted values at five-seed recall 0.91 versus 0.00 for exact matching. On the constructed table+KG benchmark, gating reduces silent wrong answers, incorrect answers returned without abstention, from 14.0% to 0.7%. The benefit depends on how the match is used. Direct join gating reduces wrong answers, while SQL hints remain limited by selecting the correct database column. Hierarchy rules, lexical overlap, and tuned rule cascades tie or beat learned verification on the data they suit. We release a dependency-light implementation, benchmark generator, evaluation harness, and label-audit package. Keywords: Data integration, Entity matching, Record linkage, Knowledge graphs, Retrieval-augmented generation, Value linking, Uncertainty calibration, Selective prediction, Text-to-SQL