Offline NLP Assistants for Internal Knowledge Access in a Central Bank: Proof-of-Concept Systems for Semantic Search and HR Chatbot

This paper presents two offline, on-premise NLP proof-of-concept assistants built on a shared architecture for internal knowledge access in the Central Bank of Bosnia and Herzegovina: (i) a semantic document search tool for internal Word/PDF repositories and (ii) an HR chatbot that applies retrieval-augmented generation (RAG) over indexed HR policies and procedures. Rather than proposing a novel NLP method, the paper contributes by documenting a reusable offline architecture for institutional AI assistants in a security-constrained central banking environment and by providing pilot evidence on how established semantic retrieval and RAG techniques can be adapted to strict requirements of confidentiality, data sovereignty, and governance. The semantic search assistant combines exact phrase matching with embedding-based retrieval and hybrid re-ranking, while the HR chatbot generates source-grounded answers using locally hosted language models under explicit governance constraints, including B/H/S-only output, strict fallback behaviour, and transparent display of retrieved passages. Pilot results indicate that hybrid retrieval offers the most reliable performance across representative internal queries, while the HR chatbot demonstrates the feasibility of document-grounded employee support under offline institutional constraints. The findings provide preliminary evidence that offline NLP assistants can improve access to internal institutional knowledge while remaining compatible with the security and operational risk requirements typical of central banking environments.