Development of a Virtual Assistant for Shift Operators of the BM@N Experiment

This paper presents the development of a virtual assistant designed to address knowledge fragmentation challenges faced by shift operators at the BM@N (Baryonic Matter at Nuclotron) experiment. The system employs a hybrid approach that integrates full-text and semantic retrieval methods using the BGE-M3 model embeddings to solve challenges related to domain-specific terminology. The article details the conversion of documents into the Markdown format and the chunking algorithm that preserves contextual relationships. The implemented system uses PostgreSQL, Valkey, and Qdrant databases with NVIDIA Triton inference server. It provides user access through both a Telegram bot and a web interface with OpenID Connect authentication. The system was evaluated using a domain-specific dataset with the mean reciprocal rank metric. The evaluation yielded MRR@5 equal to 0.68, indicating its capability to retrieve relevant information from specialized documentation. This semantic search engine is the foundation for a future complete RAG system that can generate coherent responses to operator queries.