GraphRAG is a technology allowing AI to navigate the context of your business processes and solve complex problems, such as compliance or working with large knowledge bases.
The essence of the GraphRAG approach is to use a knowledge graph to form the context of LLM when answering user questions. By GraphRAG we mean any methods of using graphs to form context in question answering tasks, not a specific solution from Microsoft. The knowledge graph can act as a direct source for forming context, and/or be used to annotate and search for relevant fragments of text. This allows achieving greater completeness and factual accuracy of LLM answers, in comparison with standard RAG techniques based on searching for relevant fragments of text using the vector embeddings similarity. Advanced techniques also include supplementing the context with logical inferences obtained on the knowledge graph, supplementing missing facts and excluding irrelevant or contradictory facts, step-by-step reasoning based on long chains of facts.
The GraphRAG implementatons can be classified by the way they use graphs. The simplest approach is to represent text fragments (chunks) as graph nodes and establish connections between the fragments in the graph that are closest in meaning. More complex approaches build a knowledge graph, the nodes of which are entities and concepts mentioned in the indexed texts.
A knowledge graph containing a description of objects and concepts, an expression of specific facts, can complement the graph representation of indexed texts, or be used as an independent information storage and a source of context formation for LLM.
A knowledge graph can be built in accordance with a certain scheme (ontology) or without it. If an ontology is used, it can be strictly defined by the developer (for the particular applied tasks), or in a semi-automatic mode, or completely constructed ad-hoc based on text analysis using LLM. An ontology may include only definitions of classes and properties of the subject area, or also contain a level of description of lexical concepts corresponding to the elements of the domain description, or also include descriptions of constraints and rules of logical inference.
The de facto standard for storing knowledge graphs in GraphRAG solutions is Neo4j, since this DBMS combines the functions of a graph and vector database: it supports storage and search of similar vectors by cosine distance. Neo4j is not a triple store and does not support the SPARQL language (although there are plugins that allow you to import RDF ontologies there and translate SPARQL queries into Cypher). This technical feature is one of the reasons why OWL ontologies are not often used in GraphRAG solutions. Other reasons include the complexity of automated construction of OWL ontologies and a fairly high threshold for IT specialists to enter the semantic technology stack. Meanwhile, the use of OWL ontologies opens the way to the integration of knowledge from various subject areas when forming a context, provides tools for formal verification of statements, expands the possibilities of using chains of reasoning when answering a question, and, in general, for the closest possible integration of symbolic and generative AI.
An overview of GraphRAG methods and techniques as of early 2025 is given in the following papers:
1. Q Zhang, S Chen, Y Bei, Z Yuan, H Zhou, Z Hong, J Dong, H Chen, Y Chang & X Huang (Jan 2025): A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models. URL: https://arxiv.org/pdf/2501.13958
2. H Han, Y Wang, H Shomer, K Guo, J Ding & Y Lei (Jan 2025): Retrieval-augmented Generation with Graphs (GraphRAG). URL: https://arxiv.org/pdf/2501.00309
3. P Jiang, S Ouyang, Y Jiao, M Zhong, R Tian & J Han (Aug 2025): Retrieval And Structuring Augmented Generation with Large Language Models. URL: https://dl.acm.org/doi/pdf/10.1145/3711896.3736557. This review, published in Summer 2025, briefly discusses some of the core elements of the GraphRAG approach, but uses the term Retrieval-And-Structuring (RAS) to refer to them.
Here is an overview of some important new works on the topic of GraphRAG for 2025, close to the topic of our research.
1. L Yao, F Ren, K Du & Q Du (04 Jul 2025): From knowledge graph construction to retrieval-augmented generation: a framework for comprehensive earthquake emergency support, Geo-spatial Information Science, DOI:10.1080/10095020.2025.2514813 URL: https://www.tandfonline.com/doi/pdf/10.1080/10095020.2025.2514813
This paper describes the Improved HybridRAG framework, which uses knowledge graph construction as one of the methods for searching relevant content in a knowledge base to form context, along with conventional search similarities of vector embeddings. The application area of the described solution is information retrieval to support decision-making in response to earthquakes. The authors used a manually created ontology containing classes and properties describing:
Then, using LLM, information is extracted from an array of documents and transformed into a structured form in accordance with the ontology. Tasks typical for this type of process are solved, including resolving co-references and identifying entities. Ontology fragments are included in LLM prompts to make the results of its work as homogeneous as possible.
In the third step, relations between the extracted entities are built: for example, relations between seismic events and measurement methods, or response measures and specific action scenarios.
Next, all the obtained information is merged in order to connect it into a single graph, as consistent as possible. Embedding is calculated for each type of relations, then relations are grouped using the spectral clustering method, after which each cluster is summarized and generalized using LLM. The result is a unified definition of each type of relations within a cluster - this is necessary to combine relations that are close in meaning, such as "is a part of" and "included in". Individual objects are then merged based on both logical rules and vector embeddings. In more complex cases, embeddings obtained from the property values and relations of each individual are used to merge individuals. The HDBSCAN (Hierarchical Density-Based Spatial Clustering) method is used to form clusters of individuals, after which LLM decides for each pair of individuals in the cluster whether they describe the same or different real-world entities.
When forming a context, key concepts are extracted from the user's query, for each of which elements of the knowledge graph with similar embeddings are determined. Then, neighborhoods of these elements are extracted from the graph, sufficient to form a meaningful context taking into account the specifics of the domain area. The constructed subgraphs are converted into a text prompt and fed to LLM as a context used to answer the user's question.
Evaluation of the results shows a higher quality of LLM answers compared to other context formation strategies, especially noticeable in terms of factual accuracy of answers.
2. A Kaisera, C Leoveanu-Condreia, R Golda, M Dinua & M Hofmarcher (23 Jul 2025): HyDRA: A Hybrid-Driven Reasoning Architecture for Verifiable Knowledge Graphs. URL: https://arxiv.org/abs/2507.15917
The HyDRA framework proposed by the authors is designed for automated construction of the conceptual level of an ontology and filling the knowledge graph with factual information. The authors reconstruct some steps of the conventional methodology for manual ontology creation using interacting AI agents. By "reaching consensus" between the AI agents, stakeholders and other persons interested in using the ontology are identified. Competency questions are formulated, to help clarify the boundaries of the ontology and the functional requirements for it, etc.
Methods for verifying the constructed ontology are also proposed, based on checking the structural constrints and functional requirements for ontological models.
The novelty of the approach is that ontology construction is considered not just as a generative task, but as an iterative process built on the principles of Design-by-Contract, only the participants in this process are not people, but AI agents. For each step of the process, clear requirements are defined, the fulfillment of which is automatically monitored during the process execution. If the requirements are violated, the feedback mechanism forces the AI to adjust the result until it meets the requirements.
It is important that the authors of this work strive to generate an OWL-DL ontology, and not just a conceptual scheme formalized arbitrarily by specifying a set of entity types and relationships, as is often done in other approaches. The content of an OWL-DL ontology is much richer than a conventional conceptual schema, as it includes statements about characteristics and relationships of properties and classes, constraints, etc.
3. M Banf, J Kuhn (28 Apr 2025): A Tripartite Perspective on GraphRAG. URL: https://arxiv.org/pdf/2504.19667v1
The paper proposes a method for constructing a knowledge graph that links three types of elements: domain objects mentioned in the text, concepts used to formulate statements about objects, and the parts of the text from which information is extracted. From each text fragment, the authors extract statements related to a certain object and/or concept and save them together with the vector embedding as properties of the relationship between the object or concept and the text in which it is mentioned. This allows to reduce the volume of the context generated when querying LLM, extracting from each chunk only facts directly related to the objects and concepts relevant to the query.
This approach does not include the step of generating a formal ontology - instead, a set of lexical concepts found in the processed texts is used.
When creating a context, it includes text fragments saved in all relations of the object(s) identified in the user's query, as well as information from other mentions of the same concepts in texts, filtered by the semantic similarity metric.
4. Y Tiwari, O Lone, M Pal (31 May 2025): OntoRAG: Enhancing Question-Answering through Automated Ontology Derivation from Unstructured Knowledge Bases. URL: https://arxiv.org/pdf/2506.00664v1
The paper proposes another approach to the automated formation of the conceptual level of the ontological model for use in GraphRAG tasks. Clustering algorithms are used to extract classes, in particular K-means and the Leiden algorithm. To determine the semantic proximity of model elements, in particular, to assign individuals to classes, the semantic similarity between embeddings reflecting the name and description of each element is used. The approach is "bottom-up": first, it identifies entities mentioned in texts, then divides them into classes, generalizes the relationships between specific entities into types of possible relationships between class objects. The ontology is verified by the analyst.
Let us mention several more works that are interesting to us in certain aspects.
5. C Ma, S Chakrabarti, A Khan, B Molnar (Jan 2025): Knowledge Graph-based Retrieval-Augmented Generation for Schema Matching. URL: https://arxiv.org/pdf/2501.08686
The paper proposes an algorithm for matching "schemas" (conceptual level ontology models) for integrating data from different knowledge graphs in GraphRAG tasks. This algorithm is useful for enriching context with information from external data sources, including publicly available large knowledge graphs such as Wikidata.
6. K Sharma, P Kumar & Y Li (Dec 2024): OG-RAG: Ontology-Grounded Retrieval-Augmented Generation For Large Language Models. URL: https://arxiv.org/pdf/2412.15235
The authors use LLM to extract facts about specific entities from text using a pre-defined domain ontology.
7. S Toro, A Anagnostopoulos, S Bello et al. Dynamic Retrieval Augmented Generation of Ontologies using Artificial Intelligence (DRAGON-AI). J Biomed Semant 15, 19 (2024). https://doi.org/10.1186/s13326-024-00320-3
The paper proposes DRAGON-AI, a method for using LLM for semi-automatic ontology generation. The method adds to the prompt as a sample structured description of ontology elements similar to the one that needs to be described, as well as contextual information describing the element of interest to the user. LLM generates a description of the required element according to the template, which is then validated.
8. M. Ahmad (May 2025): Towards a Unified Framework for Information Retrieval in Large Language Model Applications: Balancing Textual and Graph-Based Knowledge Sources. URL: https://aaltodoc.aalto.fi/server/api/core/bitstreams/8ba53084-d1eb-4839-b874-f5aa23f5ee99/content.
The master's thesis provides a practical example of GraphRAG implementation.