Recent publications on KG-RAG

In the first half of 2026, the KG-RAG approach finally moved from the realm of experimentation to a set of generally accepted methodologies, architectures, and software solutions. The need to rely on structured knowledge across a wide range of applied problems is no longer in doubt. Researchers are focusing on solving specific engineering problems, such as accelerating the search for relevant knowledge in graphs, constructing and populating knowledge graphs for various industries, addressing the challenges of working with incomplete knowledge, solving the classical graph traversal problem as applied to query context generation in LLM, knowledge element ranking algorithms, using KG-RAG in an agent-based environment, and so on.

Recent publications, as before, primarily consider schema-less graphs, which contain facts or text fragments but do not describe the structure of knowledge (the structure is either rigidly defined externally or is absent: graph nodes and connections are not typified). Typical approaches include the use of the neo4j graph DBMS together with vector database, and the LangGraph framework for implementing flexible logic at both the graph construction and usage stages.

In the literature, the term KG-RAG encompasses a variety of approaches, which can be classified by the method of graph structure formation:

1. Graph elements are text chunks or document structural elements (sections, tables, etc.)

2. Graph elements are domain objects, i.e., specific clients, products, substances, etc. (depending on the domain), but the data schema as such is absent or rigidly defined in the application. This approach is often combined with the first.

3. The conceptual level of the knowledge graph is a full-fledged OWL ontology, reflecting the types of properties and relationships of domain objects. The ontology structure is based on known high-level or industry-specific models, or is constructed ad hoc during the indexing of knowledge sources. Moreover, facts corresponding to the ontological model are not necessarily stored in an RDF triple store due to performance and scalability requirements.

Our research interest is in KG-RAG implementations based on ontological knowledge graphs, so the studies discussed below use the third or second approach. This specific area is also seeing an increase in the quantity and quality of publications. Many of them compare, among other things, the performance metrics of ontological KG-RAG versus knowledge extraction from schema-less graphs. The SCAIR and HEAR frameworks demonstrate accuracy values ​​of up to 94.7 in complex enterprise tasks, which is an extremely high indicator of effectiveness. Comparison with schema-less approaches shows that ontological KG-RAG outperforms them in response accuracy and indexing efficiency. Positive "side effects" include a higher density of useful information in the graph, fewer errors and inconsistencies due to the use of constraints, and lower LLM token costs at all processing stages. Schema-less solutions scale more efficiently due to better storage scalability, but practical industry needs often force a choice in favor of working with structured knowledge, guided by the goal of building trusted and explainable AI for professional applications.

In previous reviews, we already discussed the process of constructing an ontological knowledge graph and the research and engineering challenges that arise.

Research issues related to automated ontology construction, semantic alignment of concepts extracted from documents with existing models, and the integration of knowledge from various subject areas in a single graph remain relevant.

We have highlighted several recent studies that are most closely related to the topics of our research and practical developments. This review does not include several technologically interesting articles describing completed projects in the fields of knowledge management in cultural heritage, medicine, and biotechnology, where the use of ontological data models has long been commonplace. We will also not consider tasks related to codebase analysis. We will focus only on KG-RAG applications in industry and the corporate sector.

1. L. Wang et al., "Hypergraph Enterprise Agentic Reasoner over Heterogeneous Business Systems", May 14, 2026. [Online]. Available: https://arxiv.org/abs/2605.14259v2

The authors examine the solution of RAG problems in a corporate context, where the source of knowledge is systems such as ERP, BPM, SRM, and others. The authors rightly note that interpreting data from these systems requires expert knowledge describing corporate procedures, rules, and concepts. Solving practical problems, such as auditing order fulfillment processes or analyzing the causes of delays, requires integrating multiple, heterogeneous and incomplete data and reducing it to a single conceptual framework. This leads to problems such as differences in the understanding of the meaning of the same indicators or business object properties between different groups of business users, as well as a drift between the meaning implied in database columns by IT system developers and their actual use at a specific enterprise. All these semantic gaps can only be closed by relying on a precise description of the meaning of each type of data object. The authors encountered particular difficulties when working with n-ary relations, since an unstructured graph is a priori designed to reflect binary relationships.

The authors developed HEAR (Hypergraph Enterprise Agentic Reasoner), an enterprise agent system based on the Stratified Hypergraph Ontology. In this paradigm, n-ary relations are reflected as hyperedges, which can be declarative or procedural and describe, respectively, the states or relationships of graph nodes or the processes of executing routine tasks.

In our opinion, the term "hypergraph" is used excessively here: when constructing ontologies, classes are often created that describe complex relationships between objects (reification), which is called "hyperedge" in the context of this work. However, we understand that this term helps distinguish the generally accepted approach of constructing a simple graph in neo4j from methods for formally expressing much more complex knowledge. Unfortunately, developers and researchers remain insufficiently aware of the power of ontological tools and the modeling tasks they can solve, and the use of such terminology can be considered justified to popularize the idea of ​​applying such modeling approaches to solving "ordinary" business problems (not related to scientific research).

From a technological perspective, the authors do not use OWL ontologies or ready-made OWL models, but their approach to knowledge formalization is to some extent similar to the principles of ontological modeling. Their approach is also close to our own developments in their desire to formally describe and make elements of structured knowledge accessible to LLM. The authors describe the structure of business processes, the objects involved, and their states in the form of a knowledge graph, and create a mapping between the definitions of these knowledge elements and the actual information in business applications, which allows LLM to effectively extract and analyze information.

2. P. Chaturvedi et al., "SCAIR: Schema-Conditioned Agentic Iterative Reasoning for Enterprise Knowledge Graphs," June 2, 2026. [Online]. Available: https://arxiv.org/abs/2607.22571v1

This paper also considers an important applied problem: working with a configuration database (CMDB) for IT infrastructure or industrial systems. Such databases are fertile ground for using graph descriptions, as they include many nodes of different types, connected by various structural and functional relationships. A number of ready-made ontologies and data models exist that provide a ready-made conceptual model for creating a CMDB. Problems solved using such DBMSs include dependency analysis, fault finding, and component replacement planning. Using LLM in such problems requires deep and precise integration of the language model with a structured data source. SCAIR embeds a structured description of the knowledge schema, provides schema-aware graph traversal during multistep reasoning, and controls the propagation of entities across topics to balance exploration and exploitation during search.

In the SCAIR approach, the description of the knowledge schema is used to constrain and guide the agent's efforts to find solutions. During the planning stage, the knowledge schema is used to decompose the user's question into a set of subqueries to be answered. Constraints are also formulated to avoid obviously unproductive agent actions. The result is not a rigid query execution plan, but a set of recommendations for the agent.

During the data extraction stage, the agent finds candidate objects, filters them using a set of constraints, and evaluates their relevance to the question. The process terminates when sufficient initial data is found to answer the question, or the maximum graph search depth is reached. Practical evaluation of the results shows significant improvements in all metrics for complex questions, compared to implementations of the ReAct approach (a cycle of reasoning and action repeated until an acceptable result is achieved). SCAIR uses OWL ontologies and the SPARQL language for data access.

3. M. Campbell and H. Michaels, "GraphRAG and Knowledge Grounding for Agentic Systems in SAP Environments," [Online]. Available: https://www.scienceopen.com/hosted-document?doi=10.14293/PR2199.004145.v1

This paper proposes a framework for solving knowledge extraction problems from SAP data. It consists of four components: a graph construction pipeline that extracts facts from structured data; a graph data mining engine; a multi-agent framework for orchestrating reasoning, verifying results, and coordinating agent actions; and a continuous learning module that enriches the knowledge graph with collected feedback. Neo4j is used in combination with a vector database to store the knowledge graph. The graph structure (the list of entity types and relationships), as far as can be judged, is primarily constructed by experts and data architects in accordance with SAP metadata. The resulting model is presented in YAML or JSON format; a formal OWL ontology is not created. The review format does not allow us to describe the authors' proposed architecture for the rather complex process of searching for data in the graph and formulating answers to questions. A multi-stage approach is used, including entity extraction, the formation of relevant subgraphs, the work of agents in formulating and executing queries, and the collection and validation of results. Importantly, the built-in learning agent analyzes the meaningfulness and effectiveness of queries executed by subagents and formulates recommendations for improving agent behavior or expanding the knowledge graph. Benchmarks achieved an accuracy of 94.8 in responses to a relatively small knowledge graph of 45,000 facts, significantly exceeding the results of simpler RAG approaches.

4. Dong Guo et al. "ComplianceNLP: Knowledge-Graph-Augmented RAG for Multi-Framework Regulatory Gap Detection," April 26, 2026. [Online]. Available: https://doi.org/10.48550/arXiv.2604.23585

The authors present the ComplianceNLP framework, which monitors changes in regulatory documents in the banking sector, extracts requirements, and analyzes discrepancies between regulations and internal bank documents. The system consists of three components: a RAG pipeline based on the regulatory framework knowledge graph, a multi-task requirements extraction module combining NER, deontic classification, and cross-reference resolution using the LEGAL-BERT encoder, and requirements analysis that compares requirements with the bank's internal policies and assesses the severity of their violation. The achievement of the F1 metric = 87.7 is declared in the task of identifying discrepancies between the internal regulatory framework and external regulatory requirements. The authors distinguish four deontic modalities (requirement, permission, prohibition, recommendation), and represent each requirement as a tuple <entity, action, modality, application conditions, source>. A metric of semantic similarity between external regulatory requirements and the content of the bank's internal policies is calculated, as a result of which each pair of elementary requirements is assessed as "compliant", "partially non-compliant", or "completely non-compliant". The result of the work is a structured discrepancy report. The authors examine in detail ways to improve the solution's performance, as document analysis is a resource-intensive task. The cost-effectiveness of the proposed solution is compared with manual document processing: in the second phase of the project, costs were $48.2K per month versus a monthly budget of $45K for manual processing; in the third phase, costs are planned as $33.2K; and in the fourth, costs shall be $20.2K. Thus, the third and fourth phases shall demonstrate savings of 26% and 55% of the budget.

Other works offering interesting ideas include:

5. Q. Shi et al., "Trust-Aware KG-RAG: a framework for trustworthy question answering for engineering equipment," [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1474034626004933

The proposed framework integrates trust metrics into all stages of knowledge extraction, graph construction, context generation, and answer search. Entities and relationships between them are assigned weights characterizing the degree of trust in these data elements. At the response generation stage, these weights are taken into account when resolving contradictions between different sources and assessing the validity of conclusions.

Review Works on KG-RAG

Recent reviews of KG-RAG technologies are also appearing.

6. D. Kumar and V. Rana, "Ontology Engineering for Trustworthy AI: A Production-Grounded Survey of Knowledge Graphs in the LLM Era," [Online]. Available: https: //www.researchgate.net/profile/Dheeraj-Kumar-151/publication/405935402_Ontology_Engineering_for_Trustworthy_AI_A_Production-Grounded_Survey_of_Knowledge_Graphs_in_the_LLM_Era/links/6a2133f583a9957d732b661a/Ontology-Engineering-for-Trustworthy-AI-A-Production-Grounded-Survey-of-Knowledge-Graphs-in-the-LLM-Era.pdf

This work provides an overview of all types of technologies, standards, and common models used in the implementation of ontology-based KG-RAG, including methods for automated construction and alignment of ontology models.

7. Zhou, Dongzhuoran et al. (2026). What Breaks Knowledge Graph-based RAG? Benchmarking and Empirical Insights into Reasoning under Incomplete Knowledge. 2522-2538. 10.18653/v1/2026.eacl-long.114.

This paper analyzes the reasoning performance of KG-RAG solutions. The authors consider questions that require more than just extracting a ready-made fact from a knowledge graph, but instead requiring constructing missing facts through logical reasoning. Can KG-RAG solutions correctly find the necessary premises in the knowledge base, and can LLM draw the correct logical conclusion from them? The authors generate a test dataset from which facts that are logical consequences of other facts contained in the graph are artificially removed. An algorithm for calculating the performance metrics of various KG-RAG algorithms on such a dataset is introduced. An analysis of how the authors of various papers describing KG-RAG frameworks calculated the metrics for their solutions is conducted, and a number of shortcomings in their algorithms and discrepancies between the paper texts and the code implementation are identified. Using the unified metrics proposed by the authors of this paper, a comparative analysis of the effectiveness of various KG-RAGs on different datasets is performed. The best results are demonstrated by the RoG (Luo et al., 2024; a planning approach, where the LLM creates a plan for data extraction from the graph) and GNN-RAG (Mavromatis and Karypis, 2024; graph neural networks are used to select possible answers in the graph and find the shortest paths to them). Results on complete and incomplete datasets differ significantly (on average, by about 10% for the Hits@Any and F1 metrics), demonstrating the limited effectiveness of KG-RAG solutions in drawing logical conclusions from knowledge graph contents. Overall, the metrics of most KG-RAG solutions are not very high (40-60 for Hits@Any), and only leading solutions show Hits@Any = 60-90 on individual datasets.

8. It is worth noting the publication of the first (to our knowledge) book devoted to an overview of the entire technology stack used in combining LLM and knowledge graphs: W. Wei and H. Wu, "Advanced Retrieval-Augmented Generation: Bridging Large Language Models and Knowledge Graphs," ISBN: 978-1-394-37471-7, July 2026, Wiley-IEEE Press.