Implementing Retrieval-Augmented Generation (RAG) Pipeline for a Virtual Library Assistant (ViLA): An In-house Approach

Abstract

This paper demonstrates the application of technique in AI called Retrieval-Augmented Generation (RAG) for integration into a Virtual Library Assistant (ViLA). The RAG technique helps in the retrieval of relevant information from a database, which is then used by an LLM to generate accurate responses to user queries. The paper shows the entire process of setting up a RAG pipeline to LLM inferencing. The RAG pipeline employs an open-source embedding model alongside a vector database to function as the knowledge base. Various LLMs, typically ranging from approximately 150 million to 8 billion parameters in size were tested with the RAG output, using BERTScore, ROUGE-L and AI evaluation techniques, to identify the most suitable model for ViLA. The knowledge base consists of a collection of frequently asked questions (FAQs) from a library setting. The resulting system serves as a concrete example of AI integration into the daily workflows of information professionals and managers.