Development of a chat-bot prototype based on HybridRAG to simplify access to documents of the National Archives and Records Administration of the United States

The subject of the research is the problem of effective information retrieval in archives, specifically in the catalog of the National Archives and Records Administration (NARA) of the United States. The aim of this work is to create a prototype chatbot capable of answering users' questions based on materials from the NARA catalog. Special attention is given to justifying the choice of a specific approach to generation, taking into account the specifics of the subject area, which is necessary to achieve the goal. The algorithm for creating a knowledge graph based on catalog materials and the architecture of the chatbot are discussed in detail. Various components used in the script for creating the knowledge graph and within the proposed chatbot architecture are analyzed. Examples of both correct and incorrect operation of the system using Mistral-7B-Instruct-v0.3 are examined. The methodology is based on the implementation of the HybridRAG approach, a variant of GraphRAG, which involves a combination of semantic search using vector representations and search via graph queries. The result of this work was the creation of a script that generates a knowledge graph based on catalog materials using Neo4j, a script for analyzing the generated graph, and a prototype chatbot that implements the HybridRAG approach using the generated graph. The area of application for the results is information retrieval systems for archives. The novelty lies in the adaptation of the HybridRAG approach to work with the complex hierarchical structure of the catalog. The prototype demonstrated the ability to provide correct answers based on catalog data. At the same time, key limitations of the prototype were identified: the large language model used within it must be resilient to "noise," refrain from answering questions when the necessary information is absent in the provided context, and be able to integrate information from multiple documents. The results of the work, including the created scripts, prototype code, and additional materials (tables, images, prompts, Cypher queries), are available in the GitHub repository at the following link: https://github.com/AlexanderPavlov36/NARA_ChatBot.