Abstract

In the domain of official statistics, public institutions face the dual challenge of managing increasing volumes of user inquiries while maintaining rigorous standards of factual accuracy and formal communication. The Central Bank of the Republic of Türkiye (CBRT) manages the Electronic Data Delivery System (EVDS) which is a critical platform serving researchers, financial analysts, academia and media. To address the inefficiencies of manual support processes, this study investigates the automation of email responses using Large Language Models (LLMs), specifically focusing on the trade-off between semantic precision and institutional voice.

We conducted a comparative analysis of two distinct generative AI architectures implemented within the CBRT infrastructure:

1. Retrieval-Augmented Generation (RAG): This approach integrates a generative model with a dynamic knowledge base derived from historical support logs and documentation, prioritizing factual grounding and up-to-date information.

2. Low-Rank Adaptation (LoRA): This parameter-efficient fine-tuning method adapts the open-source Gemma-3 model to the specific domain by aiming to internalize the stylistic norms and phrasing of the EVDS support team without extensive retraining.

The performance of both systems was assessed using a hybrid framework combining automated metrics (BERTScore and ROUGE) and human evaluation by domain experts across dimensions of accuracy, usefulness and trustworthiness.

Results reveal significant insights in terms of accuracy and style balance. The RAG architecture demonstrated superior performance in semantic reliability deriving from high BERTScore and factual consistency, significantly reducing the risk of hallucinations by anchoring responses in retrieved context. While LoRA proved effective in generating linguistically fluent drafts that aligned to the institution's communication style, it was limited by the temporal rigidity of its training data and lacked the real-time precision of the retrieval-based system.

The study concludes that while fine-tuning captures the form of institutional tone, retrieval mechanisms are essential for the substance required in official statistics. We propose a hybrid framework that leverages LoRA for stylistic alignment and RAG for factual grounding as the optimal strategy for deploying generative AI in sensitive public sector environments.