Employing Approaches of Raising Quality of Generated Answers of AI LLM-Systems when Accumulated Enterprise Information is Used

Abstract

The development of large language models (LLM) has extended opportunities of processing non-structured data in corporate environment. However, LLM fundamental restriction lies in their dependence on data of preliminary learning, which can decrease their ability to work safely with enterprise information non-published in other sources. The article provides analysis of current approach to using corporate data to upgrade quality of answers generated by LLM. Systems of Retrieval-Augmented Generation (RAG) are studied both in classical vector realization and in GraphRAG. Agent systems and integration of RAGand GraphRAG approaches in their architecture were discussed, as it can make it possible to build complicated systems of AI and solutions requiring multi-stage analysis of corporate non-structured data. Architecture of classical vector RAG-approach is described, which consists of three key stages: getting vector presentation of indexed fragments of papers; searching for fragments relevant to user requirement; forming the answer on extracted context combined with the initial request. As a key restriction the author showed the absence of mechanism necessary for integral analysis of loaded proprietary data. In its turn GraphRAG uses mechanism of plotting knowledge graph on context, which can help work not with semantically similar data fragments but with hierarchically clustered information by accumulated analysis of text summaries. Shortcomings and benefits of using approaches were formulated, criteria of expediency of RAG and GraphRAG practical application were identified and examples of systems for their use were provided