Published on: 2026-06-16
Source: Novosibirsk State University –
An important disclaimer is at the bottom of this article.
The development of NGU to reduce “hallucinations” of neural networks became the winner of the Yandex Open Source competition, held with the support of the SourceCraft platform for developers in the “Artificial Intelligence” track. Among the best was the RAGU project (Retrieval-Augmented Generation & Understanding), worked on by a team led by an associate professor of Novosibirsk State University.Ivana Bondarenko.Â
RAGU is a modular engine of GraphRAG, designed for cases when simple vector search is no longer sufficient and it is necessary to build full-fledged knowledge graphs from textual documents. It allows collecting, ranking, and flexibly querying structured data based on graphs of relationships between concepts, not just on string matching of formulations.
—The raw text is automatically cut into stable fragments, from which the neural network extracts entities and relationships, after which RAGU assembles a graph, merges duplicate nodes, and breaks the network into clusters according to the Leiden algorithm,— described the essence of his workIvan Bondarenko.
The result is a universal constructor of language model integration systems with knowledge bases. When the user has a question, especially a complex one, the system allows generating an answer based precisely on knowledge about the subject area, correctly, quickly, and without hallucinations. Thus, RAGU helps artificial intelligence rely on a formal knowledge model extracted from documents, rather than on general statistical associations.
RAGU developments are now used in the “Menon” system — a digital assistant for NGU applicants, which answers questions about admission, academic programs, and university life. Unlike standard search services, “Menon” engages in dialogue with the user and formulates answers in its own words, relying on a specially structured university database.
—Such a system can be adapted for other industries as well — from construction to working with regulatory documentation. Everywhere there is an archive of documents or a knowledge base and it is necessary to use it effectively when making decisions, an important system is one that, on the one hand, understands the language well, and on the other — is capable of building a formal knowledge model based on texts, — emphasized Bondarenko.
According to the researcher, the team systematically compares its solution with foreign and Russian counterparts.
—We compared RAGU with the Chinese solution LightRAG, which is currently considered one of the most advanced in the world, and with several other libraries. Our solution demonstrates better performance quality both in terms of the MenoLite neural network itself and the RAGU library in which this model is used., — he noted.
Winning the competition gives the project both reputational and practical effects, since the winners are awarded grants to use Yandex Cloud resources for the development of open projects.
—For us, this is, first of all, confirmation of product quality, and secondly, the opportunity to conduct additional experiments using Yandex’s cloud resources. We are now actively using these capabilities to manage to complete a series of experiments by the deadline of the largest conference on natural language processing, where we will submit an article about RAGU., — said Ivan Bondarenko.
The project is being worked on by a team, headed by Bondarenko himself as a representative of the Laboratory of Applied Digital Technologies of NGU, its graduate Mikhail Komarov and other guys – both current students and graduates already working in IT companies. Students from other universities also participate in the project — from the Far Eastern Federal University to the Immanuel Kant Baltic Federal University.
In the near future, the project participants plan to teach the system to better understand “live” human speech and complex dialogues. The developers are working on a component that rewrites the user’s question into a form more independent of context for searching the knowledge graph: it should correctly handle ambiguous abbreviations like “IT” (information technology or Institute of Thermophysics), take into account the history of the dialogue, and properly link pronouns and omissions in phrases.
Please note; this information is raw content obtained directly from the information source. It represents an accurate report of what the source claims and does not necessarily reflect the position of MIL-OSI or its clients.