Translation. Region: Russian Federation –
Source: Novosibirsk State University –
An important disclaimer is at the bottom of this article.
The RAGU project, presented by the development team of the Applied Digital Technologies Laboratory International Scientific and Educational Mathematical Center of NSU, won the "Innovations in Artificial Intelligence" category of the "Code Without Borders" competition, held as part of the "Code Without Borders" grant program by GitVerse, Cloud.ru, and Habr. Over 200 applications from across the country were submitted, but the project by NSU researchers was recognized as the best. Its concept forms the basis of the "Menon" chatbot, which they are developing for NSU applicants. The RAGU software library was presented at the Datafest conference. A research paper on the library and its applications is currently planned. Ivan Bondarenko, a research fellow at the Laboratory of Applied Digital Technologies at the International Scientific and Educational Mathematical Center at NSU, spoke about the grant program and how his RAGU project became a winner.
RAGU (Retrieval-Augmented Graph Utility) is an open-source software library designed to integrate knowledge graphs with large-scale language models (LLMs), improving the accuracy and reliability of responses and reducing the risk of hallucinations. Its architecture is similar to the GraphRAG approach, but is based on the principle of "stepwise" knowledge graph construction: a multi-step process is used, with a pre-trained smaller model used for the first step, reducing resource requirements.
Ivan Bondarenko explained that the key to success lies in combining knowledge graphs and modern language models, which improves accuracy and reduces the risk of hallucinations in responses.
"The initial idea behind RAGU was to open access to tools for the efficient, synthesized operation of knowledge graphs and LLMs. Our open-source software library enables the integration of large language models with knowledge graphs to improve the accuracy, reliability, and reduce the hallucination of responses from large language models to user questions. We used a multi-step approach—we specifically retrained a generative neural network to be an effective tool for constructing a knowledge graph, and to do so in multiple steps rather than in a single step. This approach reduces hardware requirements and accelerates the process. With the original approach, efficient knowledge graph construction often required enormous language models (up to ~32 billion parameters). Our approach reduced the size to approximately 600 million parameters through retraining and a multi-step architecture, while maintaining or even improving the quality compared to traditional solutions within the GraphRAG methodology," the researcher explained.
The project attracted participants from various cities and universities across Russia, highlighting its nationwide reach. In addition to NSU students and staff, it included representatives from Lomonosov Moscow State University, Immanuel Kant Baltic Federal University, MISIS University of Science and Technology, Far Eastern Federal University, and ITMO University: Ivan Bondarenko (NSU), Mikhail Komarov (NSU), Yana Dementyeva (NSU), Roman Shuvalov (NSU), Nikita Kukuzei (MSU), Ilya Myznikov (IKBFU), Alexander Kuleshevsky (MISIS), Stas Shtuka (FEFU), Matvey Soloviev (ITMO), and Fyodor Tikunov (NSU).
"We didn't come up with the concept itself. We borrowed the idea for the GraphRAG architecture from a Microsoft paper published a year ago. It turned out to be a good one, but we noticed a number of shortcomings: a very lengthy knowledge graph construction procedure and non-deterministic results. We were able to speed up the process and improve reliability using our approach. The architecture includes multi-step tuning and retraining of a smaller model, which allows us to reduce the model size and hardware requirements. The knowledge graph is built on nodes—named entities—and arcs—the relationships between them. This allows us to create a human-readable and reliable world graph, separated from the "black box" of a neural network," explained Ivan Bondarenko.
RAGU is already the basis for accelerating processes and demonstrates speed advantages over larger models. Ivan Bondarenko plans to write a scientific paper about the library and its application. This paper will be further developed and ported to the Menona engine within NSU.
Material prepared by: Elena Panfilo, NSU press service
Please note: This information is raw content obtained directly from the source. It represents an accurate account of the source's assertions and does not necessarily reflect the position of MIL-OSI or its clients.