ANALYSIS OF RETRIEVAL-AUGMENTED GENERATION METHOD IN THE AREA OF LEGAL CONTRACTS GENERATION

Abstract

The subject of the study is the method of Retrieval-Augmented Generation of machine learning for generating contracts under limited resources and methods of comparing and assessing their effectiveness. The work aims to analyze the method of search-augmented text generation for the development of independent specialized systems and to assess their effectiveness for generating contracts in different languages ​​in different legal systems. The following tasks are solved in the article: determining the method of Retrieval-Augmented Generation for adapting models to narrowly focused industries; analyzing the methods of evaluating and comparing such systems; identifying the limitations of existing solutions and approaches; finding the optimal approach under limited resources. The following results were obtained: the method of RAG was investigated in combination with large language models; the architecture of systems based on this approach, its main structural components, and possible variations were considered; the advantages and disadvantages of use in specialized industries were determined; two methods were compared: large general-purpose language models without additional tuning and modifications and systems using RAG for adaptation to the selected subject area. As a result of a practical experiment, it was determined that the RAG method significantly improves the accuracy and completeness of answers in the legal field compared to standard models but requires more time and computational resources. The article provides an overview of Retrieval-Augmented Generation method for text information generation using large language models, considers their advantages and disadvantages, and discusses the possibility of application in limited financial and human resources conditions. The paper considers a specialized legal field and the problem of contract generation and determines the effectiveness of the selected method for its solution.