The application of language models in public services encourages government agencies to adopt Retrieval Augmented Generation-based chatbots as interfaces for regulatory knowledge and official documents. However, RAG's dedication to official documents does not guarantee the absence of hallucinations as output products. RAG also does not reduce public trust and legal confidence. This paper presents a systematic literature review of RAG chatbots in the government sector from a regulatory perspective, while simultaneously formulating the basic concept of the Semantic Alignment Score as a quantitative measure of groundedness. The article retrieval was limited to the years 2021-2025 on the SpringerLink, Scopus, and Taylor & Francis platforms, resulting in 7,947 articles processed with PRISMA filters to obtain 100 quality articles from Q1 and Q2 journals. Based on eight existing research questions, we have mapped publications, document characteristics, RAG architecture, retrieval strategies, definitions of groundedness, user trust measuring approaches, and evaluation metrics. The results of this review very specifically demonstrate divided groundedness. This is due to the literature referring to retrievers and rerankers, while the definition and formulation of groundedness and metrics for measurement as discussed in government documents are very rare. Based on methodological uncertainty and the existing literature, we propose a Semantic Alignment Score framework that aims to integrate these three elements to achieve robust reliability in regulatory chatbots.
Rumambi et al. (Sun,) studied this question.