Design and Evaluation of a Modular Multilingual RAG Architecture for Trustworthy and Auditable AI Systems: The PrivacyEthical Chat Tool

Context and motivation. The growing adoption of Large Language Models (LLMs) in institutional settings introduces new challenges related to trustworthiness, transparency, and ethical compliance. In critical domains such as ethics, privacy, and information security, it is essential that automated systems provide traceable and contextually accurate answers. Retrieval-Augmented Generation (RAG) offers a promising approach to meet these demands by grounding generated content in external, verifiable sources. However, little research has examined how RAG architectures can be effectively designed and evaluated for multilingual and governance-sensitive environments. Question/problem. This study investigates how a modular RAG architecture can be designed to produce high-precision and auditable outputs in multilingual institutional contexts. Specifically, it explores how heuristic language detection, domain-specific routing, and semantic retrieval can be combined to ensure accuracy, interpretability, and scalability in automated support systems. Principal ideas/results. We designed and implemented a three-stage RAG pipeline comprising (i) heuristic language detection, (ii) domain-specific classification, and (iii) semantic retrieval using high-precision vector embeddings. The system was evaluated on a bilingual (Portuguese/English) dataset covering ethics and privacy requirements. Experimental results indicate an accuracy of 84.3% in language detection, embeddings achieved clear semantic separation between domains, and cosine similarity scores for relevant documents consistently exceeded 0.7, demonstrating robustness and retrieval effectiveness. Additionally, the architecture successfully generated structured and traceable user stories grounded in validated ethics and privacy requirements. Contribution. This work contributes a validated multilingual RAG architecture that operationalizes principles of trustworthy and explainable AI through transparent retrieval and accountable response generation. The proposed design supports institutional governance and accountability by combining accuracy, interpretability, and scalability, advancing the development of ethical, auditable, and reliable AI systems across organizational contexts.