Recent Publications on KG-RAG

In our previous review, we used the term GraphRAG to describe a technology for extracting knowledge from text into a graph and using this graph to generate LLM context for answering user questions. From this review onward, we will refer to it as KG-RAG (Knowledge Graph RAG). This term avoids confusion with Microsoft's GraphRAG implementation.

1. Serratore, G., & Clunis, J. (September 2025). The Integration of Artificial Intelligence and Ontologies: Transformations in Knowledge Representation and Application. Proceedings from the North American Symposium on Knowledge Organization, 1–11. https://doi.org/10.7152/nasko.v7i1.95643

The "AI for the construction and populating of ontologies" section of this article is devoted to an overview of frameworks and methodologies for the automatic creation of ontologies using LLM and NLP. Both approaches using fine-tuning LLM and those oriented toward prompt engineering are considered. A number of problems inherent in some of these implementations are noted, such as insufficient semantic expressiveness of the generated classes and properties (in comparison with manually created ontologies), incorrect choice of relationship direction, incorrect assignment of the scope and range of property values, problems generating syntactically correct Turtle files, etc. The positive effect of using ontology fragments in LLM prompts is noted, allowing the language model to internalize conceptual structure of the domain model and improve the accuracy of responses.

2. da Cruz, T., Tavares, B., & Belo, F. (November 2025). Ontology Learning and Knowledge Graph Construction: A Comparison of Approaches and Their Impact on RAG Performance. https://arxiv.org/abs/2511.05991v1

The authors of this review compare vector RAG, Microsoft's GraphRAG implementation, and information extraction from graphs populated with information from relational databases and text to investigate the impact of using ontologies on the accuracy and completeness of dialog system responses. The conclusion of the study is that adding text fragments as properties of graph nodes improves the accuracy of information extraction from the graph. Another clear observation is that knowledge graphs populated with data from relational DBMSs are cheaper and easier to create and maintain compared to graphs using text as an information source. Of interest to us is that this work also considers the automation of the formation of the conceptual level of the ontological model, that is, the extraction of a set of object types and properties from a data source. For relational DBMS, the ontology is formed on the basis of DDL tables, and for text, the authors follow the approach described in the article Bakker, R.M., Di Scala D.L., & de Boer M. H. T. (2024) Ontology Learning from Text: an Analysis on LLM Performance. In Proceedings of the 3rd International Workshop on Natural Language for Knowledge Graph Creation co-located with the 20th International Conference on Semantic Systems (SEMANTiCS 2024). CEUR Workshop Proceedings, vol. 3874, pp. 70-87. In this work, the authors investigated several approaches to generating the conceptual level of ontology using LLM: a direct prompt to create an ontology (an obviously unsuccessful option); Dividing the ontology into three subtasks—extracting classes, individual objects, and relations; and a variant of the second approach, in which classes are extracted from the entire text at once, and individuals and relations from each sentence separately. The result was compared with a manually constructed ontology. The latter variant, as expected, proved to be the most effective.

3. The work by the same authors: Bakker, R.M., Di Scala, D.L., & de Boer, M.H.T. (December 2025). Ontology Learning with LLMs: A Benchmark Study on Axiom Identification. https://arxiv.org/abs/2512.05594 is devoted to testing various LLMs on the task of extracting axioms from several existing ontologies using different prompt construction strategies (zero-shot, one-shot, etc.). The authors focus on axioms describing class and property relationships, such as superclass-subclass, property-subproperty, domains of application and property value ranges, and others. LLM is fed an axiom-free ontology as input, and the model's task is to propose correct axioms that can complement it. A wide range of models were tested, including GPT-4o, DeepSeek R1, Mistral 3.1, Llama 3.3, Qwen3, and others. The philosophical capabilities of LLM are currently insufficient to achieve high precision and recall metrics in this task. LLMs are most successful at identifying superclass-subclass relationships, and GPT-4o demonstrated the best performance among all the models considered.

4. Cappelli, M. A., & Di Marzo Serugendo, G. (2025). Methodological Exploration of Ontology Generation with a Dedicated Large Language Model. Electronics, 14(14), 2863. https://doi.org/10.3390/electronics14142863

The authors focus on the interaction between humans and generative models in the process of ontology creation. They describe the ontology creation process, which includes six phases: theoretical definition (a human defines the domain and defines its key concepts), domain exploration and analysis, ontology synthesis, formalization, self-assessment, and integration.

LLM acts as an assistant, suggesting new concepts and properties, as well as ways to incorporate them into the ontology, and performing some verification tasks for the resulting model. The human retains the management role and final decision-making authority over the model's composition and structure.

It should be noted that the authors of this article, like most researchers, use large, general-purpose models (e.g., ChatGPT-4o), whereas we aimed to implement a method that works with simpler models.

5. In Abolhasani, M. S., & Pan, R. (December 2025). OntoKGen: A Genuine Ontology and Knowledge Graph Generator Using a Large Language Model. https://doi.org/10.1109/rams48127.2025.10935139 proposes another approach to constructing an ontology generation pipeline using LLM. This approach includes the following steps: identifying concepts (and defining individual objects corresponding to these concepts), defining relationship types and linking specific objects, defining property types and assigning specific property values, and verifying and checking the graph's completeness.

As far as the publication can tell, this approach uses expanded LLM prompts aimed at producing ready-made classes and concepts. The processing result is stored in a schema-less graph in Neo4j, meaning a full-fledged ontology is not necessary: ​​it is sufficient to typify objects and relationships without constructing hierarchies of classes and properties, introducing constraints, and other ontological axioms. In our opinion, knowledge graphs stored in Neo4j cannot be fully considered ontological. They are suitable for solving problems in subject areas where the conceptual structure of knowledge is not very complex, but the number of individual objects and relationships, on the contrary, is very large.

6. The work of Park, J., Jeon, H., Lee, Y., Hong, J., & Kim, M. (December 2025). Ontology-Based Knowledge Graph Framework for Industrial Standard Documents via Hierarchical and Propositional Structuring. https://arxiv.org/pdf/2512.08398 considers the problem of extracting facts and requirements from industrial standards texts, necessary for creating the KG-RAG system. The authors propose a pipeline consisting of three stages. The first stage, Hierarchical Ontology Modeling, consists of reflecting the hierarchical structure of standards in the model. For well-structured regulatory documents, the division into sections, subsections, and chapters actually carries a semantic load and can be used to form an ontological framework. The second stage, Atomic Proposition Ontology Modeling, reflects the meaning of propositions from the standard text into triplets. At this stage, special attention is paid to transforming tables into fact sets. Often, each cell contains the value of some quantity that must be applied under the conditions defined by the intersection of a row and a column. It should be noted that the authors interpret a triplet in a sense not typical for ontologists—as a construction of two nodes representing elementary entities and a typified relationship between them. In the authors' methodology, triplets are elements of a case description, or a requirement application scenario, associated with a condition (the value of the if clause) and the applicable requirements (the then clause). A few-shot learning technique is used, in which the LLM prompt includes several samples of the desired result of processing other similar expressions. During requirements extraction, a dictionary of entities and relationships is constructed. To eliminate duplicate definitions, clustering is used using HDBSCAN vector embeddings obtained from terms defining entities and relationships. Three embeddings are generated and concatenated based on the text definition of the term, the context of its use, and the term's position in the graph structure. For relationship definitions, a fourth embedding is added, describing the types of nodes linked by the relationship. Redundant links and duplicate statements are removed from the graph. In the third stage, Ontology-based Knowledge Graph Construction, the assertion triplets are linked to the ontological schema of the document structure.

To find assertions in the graph during the knowledge extraction stage, a search for document sections is used based on the proximity of the semantic embeddings of the query and each section of the standard. The context of the LLM query is formed from assertions extracted from the relevant sections. The relationship between the document structure and the specific assertions extracted from them allows for the formation of a proof. (evidence) of each statement from the LLM response.

An evaluation of the results shows that the performance gain of the method (F1 metric) relative to a simple text-based RAG is small. The advantage of this approach is that it maintains performance as the size of indexed documents increases.

In our opinion, this approach does not fully utilize the advantages of graph-based knowledge organization. It is tailored to processing documents of a specific structure and cannot be applied to less formalized documents. The structure of statements, which become elementary triplets of the knowledge graph, is fixed and does not allow for the reflection of complex combinations of conditions under which certain requirements must be applied. The conceptual model extraction step is not isolated from the overall pipeline and does not result in the formation of a machine-readable representation of the model's conceptual level. The conceptual model is not used in the knowledge extraction step. The extraction itself occurs in the traditional way, using vector embeddings – no attempt is made to transform the user's question into a query to the knowledge graph. All this limits the applicability of the approach and the results that can be obtained with it.

Several recent studies are devoted to the problems of processing various codes: from disease codes in medical ontologies to industrial product codes or emission factors of industrial enterprises. LLMs perform poorly in finding precise encodings and matching codes for the same object from different classifiers; they can confuse them or invent non-existent codes. In such problems, the need to combine LLM with the use of structured knowledge is particularly evident.

7. Zhao, Y., Dai, C. H., Niyato, D., Tan, C. F., Xiang, K., Wang, Y., Yeo, Z., Tan, D. Z. L., Low, J. Z., & Ho, E. H. Z. (June 2025). A Graph-Retrieval-Augmented Generation Framework Enhances Decision-Making in the Circular Economy. https://doi.org/10.21203/rs.3.rs-6795706/v1

The authors of the work construct a knowledge graph that combines information on the types of industrial emissions, materials, andThe authors define waste classifications, materials, and environmental factors (GWP100) according to several standards and use it to answer user questions using LLM. The authors call their solution CircuGraphRAG. The work focuses on extracting data from the graph; the process of its construction is not discussed in detail. It can be assumed that the authors use structured or semi-structured data sources and a predefined conceptual schema (ontology). A significant portion of the graph content consists of entities described in industrial standards and assigned specific code values: this eliminates the problem of identifying and classifying entities when processing regulatory texts. To answer questions, LLM generates SPARQL queries against the graph, which return precise results. The importance of the solution's correct operation with numerical values ​​of various indicators in tasks related to monitoring compliance with regulatory requirements is emphasized: CircuGraphRAG demonstrates 100% accuracy in numerical responses, which compares favorably with the naive RAG, which, in the authors' tests, either fails to provide answers to such questions or provides incorrect answers. The authors experiment with several approaches to constructing such queries: using predefined query templates (the model simply selects an appropriate template and inserts data corresponding to the user's question), generating a SPARQL query by the model, or a fuzzy approach: generating a SPARQL query with contextually provided hints regarding entity type identifiers and properties. Generating a SPARQL query by the model from scratch does not produce practically usable results, and the effectiveness of the fuzzy approach depends on the type of LLM used. The approach with a pre-prepared template library is reliable, but it limits the range of tasks that can be solved. In our opinion, the optimal balance between versatility and answer accuracy is provided by an approach not considered by the authors of this article: constructing a SPARQL query with an imperative algorithm using a combination of techniques, such as access to LLM, identifying entity and relationship types using vector embeddings, and validating and adjusting the query structure based on the conceptual schema of the ontology.

8. Wang, Z. Y. (July 2025). Retrieval-Augmented Question Answering System Based on Large Language Models and Knowledge Graphs. Applied and Computational Engineering, 178(1), 236–245. https://doi.org/10.54254/2755-2721/2025.po25800

This work is devoted to the creation of a RAG solution for a knowledge base in the field of dermatology. The article describes the full solution development cycle – populating a knowledge graph with information from text documents and implementing a dialog interface for generating answers to user questions. A predefined knowledge graph structure is used, based on 19 key disease properties, such as risk factors, diagnostic criteria, symptoms, etc. The LLM is tasked with representing the contents of text documents as triplets in terms of a predefined ontology. The result is stored in the Neo4j database. When processing a user question, the entities mentioned in it and their types are first identified, after which the most relevant graph nodes are found. Using Cypher queries, the environment of each relevant node is extracted. The context for LLM is formed from the descriptions of graph node groups.

We did not review many works describing the application of similar approaches to solving various industry problems in biotechnology and medicine, IoT, engineering, telecommunications, manufacturing, and construction.

Analysis of Current Approaches to KG-RAG Implementation

Researchers working with tasks where obtaining correct answers in a dialog system requires precisely finding the desired fragment of source information, often among vast amounts of text with identical vocabulary, are faced with the fact that adequate navigation through a text corpus is impossible using vector embeddings alone. This leads to the idea of ​​​​representing the source information in a more structured and formalized form to ensure accurate selection of information for context. In cases where a precise and verifiable answer to a question is required, especially in situations of requirements compliance monitoring or dialogic interaction with regulatory documents, forming the context of an LLM query, even from relevant text fragments, carries a serious risk of hallucinations. In such cases, the response itself must be generated based on a query executed against structured data, and LLM can only transform it into a more human-readable form. These premises lead researchers to formulate and solve problems of knowledge graph formation for RAG tasks.

When creating KG-RAG solutions, many questions must be addressed, the main ones of which can be formulated as follows:

  • Conceptual schema of the knowledge graph (set of object types and properties): should it be formed manually in advance or built based on text processing with varying degrees of automation?
  • Transformation of knowledge according to the conceptual schema: should this task be entirely delegated to the LLM, providing it with examples of formally recording information in the form of a graph, or should some combination of imperative algorithms and NLU tools be used with calls to the LLM?
  • Finding relevant fragments of the knowledge graph: compare vector embeddings of the question and graph nodes, or transform the user's question into a graph query in SPARQL or Cypher? Various combinations of these approaches are possible, including thanks to the ability to store embeddings in some graph storage systems and their support for vector operations.
  • If a graph query is chosen, the next choice concerns the method of generating it: using a template library, creating the entire query using LLM, or constructing the query piecemeal. The latter option is preferred when complex relationships between graph nodes need to be found: in this case, the query construction task must be based on a conceptual framework to include only valid relationship types in the template and exclude the search for formally valid but essentially incorrect relationships, as well as to use various variants of the shortest-path algorithm.

Depending on the answers researchers choose to these questions, the resulting solution has greater or lesser universality and accuracy. We strive to create the most universal solution without sacrificing accuracy, which requires the use of complex, combined techniques for creating a conceptual schema, extracting facts from text, and generating an LLM context for answering questions.

In our view, the main conclusion from the above works is that it is necessary to rationally divide tasks into those implemented by imperative algorithms and those implemented by LLM. For example, the task of generating a Turtle file is unambiguously implemented by a standard software procedure that guarantees the syntactic correctness of the result. Those parts of the task that are primarily performed using LLM, such as determining the scope and value range for properties, should be verified both by control questions (in a feedback loop that terminates only after a comprehensive check of the correctness of the resulting assertions) and by algorithmic procedures of formal refinement (for example, replacing the list of all subclasses of a class in the scope of a property with the class itself).

Many approaches strive to involve humans in all stages of automated ontology construction. Depending on the complexity of the domain and the requirements for the accuracy and expressiveness of the ontology, this may make more or less sense. From the perspective of our practical tasks, human verification of the ontology should occur only once, at the very end of its formation.

Our review shows that the KG-RAG approach remains an area of ​​active research, where a generally accepted approach has not yet been developed. We note the internal competition between approaches generating rigorous conceptual OWL/RDF models and constructing weakly structured graphs in neo4j. The advantages of both approaches over naive RAG are obvious and have been quantitatively proven in many studies, but the advantages of various variants of the KG-RAG approach remain a subject of debate.

It is also interesting to note that the use of graph neural networks (GNNs) in KG-RAG tasks remainse is the focus of attention of researchers.