A new paper on arXiv tackles a common problem in consumer banking search: people describe what they need in plain language, not in the terms a bank's website uses. The authors give the example of a request like "a card for travel," which may share few keywords with the relevant product page. This mismatch can cause standard retrieval systems to miss the right results.
The paper proposes an evidence-gated bilingual RAG (retrieval-augmented generation) approach. The idea is to treat the consumer's natural-language query and the bank's formal product terminology as two sides of the same intent, then use retrieval augmented with a gating mechanism to weigh evidence before generating an answer. The goal is more reliable search outcomes in a domain where precision matters.
The abstract is truncated in the source, so specific experimental details and results are not available here. Still, the core contribution is clear: a framework designed to reconcile user-friendly phrasing with institutional vocabulary, potentially improving both recall and trust in banking search tools.