Grouping-Aware Numerical Verification for Offline Retrieval-Augmented Question Answering over Institutional Documents

Abstract: Institutions hold large volumes of policy documents, reports and spreadsheets that are still searched manually. Retrieval-augmented generation (RAG) offers a natural interface to such collections, but mainstream deployments route confidential material through hosted language-model APIs and provide no guarantee that figures quoted in an answer actually occur in the source text. This paper presents an offline RAG system for institutional document question answering in which ingestion, embedding, retrieval and generation all execute on commodity hardware with no external network dependency, and in which every numeral in a generated answer is verified against the retrieved context before the answer is shown. The central observation is that naive numeric extraction — the digit-matching regular expression used in comparable pipelines — fragments digit-grouped quantities such as 45,23,000 into unrelated components, causing a verifier built on it to reject faithful answers. We describe a grouping-aware extraction and verification procedure that handles Indian and Western grouping conventions, decimal and percentage forms, and enumeration markers. On a 15case extraction benchmark the proposed extractor is exact on 15/15 cases against 5/15 for the baseline, and substituting the baseline extractor into the verifier introduces false rejections of faithful answers. Measured retrieval latency over an exhaustive 384dimensional index remains below one millisecond to roughly 5,000 chunks and grows linearly to about 10.5 ms at 50,000 chunks, which bounds the corpus size at which exact search remains appropriate.

Keywords: retrieval-augmented generation, hallucination detection, offline inference, vector search, document question answering, privacy-preserving systems.