Biswas highlighted another issue with graph databases: “the ‘supernode’ problem, where a single node in the graph is connected to many other nodes, such as a user with millions of followers in a social network or a popular product in a catalog. Querying a supernode or its neighbors can be resource-intensive and lead to scalability issues, as the database must process a large number of relationships.”
In some cases, Biswas continued, “Relational databases may be more appropriate, particularly when the data is less connected or when transactional consistency takes priority over exploring relationships between datapoints.” Implementing graph databases “can be challenging, especially in organizations where data silos are a problem,” Ward cautioned. “Gathering data from different departments requires breaking down these silos, which can slow down the process significantly. Additionally, most users are accustomed to working with tables and key-value entries in relational databases. This means there’s a steep learning curve when transitioning to graph databases, often limiting their use to data science teams familiar with the technology.”
Graph databases “are highly optimized for use cases involving complex, dynamic relationships, but they can introduce unnecessary complexity or performance issues in scenarios where relationships are minimal, data is static, or high-volume transactional and analytical processing is required,” said Perlov. “For these types of use cases, traditional relational databases, NoSQL solutions, or specialized systems provide better performance, scalability, and simplicity.”
For example, Perlov continued, “document-based data or semi-structured content is a better fit for vector databases, while financial and accounting systems require a high level of transactional integrity, support for complex SQL queries, and efficient handling of numerical data, which relational databases are designed to handle.”
In addition, “Transactional systems with simple, structured data—like inventory management or payroll systems—typically involve structured data with well-defined schemas and a focus on transactional integrity,” Perlov illustrated. “Relational databases are a better fit here; applications requiring low-latency responses for many small, individual transactions—for example, a customer chatbot—perform better with NoSQL or in-memory databases, which are optimized for quick reads and writes with minimal overhead.”
Another common challenge in developing and working with graph databases is ensuring data quality and relevance in graph implementations, said Rao. “Without well-defined relationships, performance can degrade.” For instance, “A poorly structured knowledge graph might slow down AI systems rather than enhance them,” Rao noted. “Organizations must carefully design their data models to ensure scalability. Additionally, there’s a skills gap—teams often lack the specialized expertise needed to manage graph databases. In cases where the focus is purely on transactional data, such as inventory management or simple customer records, traditional relational databases are often more suitable due to their straightforward implementation and reliability. The key is to choose the right tool for the right task—graph databases shine in scenarios where relationships between datapoints are central.”
Graph databases also place greater demands on data quality. “Any inconsistencies or errors can lead to incorrect results, as these systems rely on accurate, real-time information to provide intelligent insights,” said Ward. “The complexity of knowledge graphs can also be a barrier, as the tools and platforms available may not be as user-friendly or widely adopted, making it difficult for general users to work with them effectively.”
What’s Ahead
How do experts see the adoption and evolution of these technologies in the coming years? They point to an emerging construct called GraphRAG (graphs plus retrieval-augmented generation), a data pipeline and transformation suite that is designed to extract meaningful, structured data from unstructured text using the power of LLMs. According to Microsoft, a leading proponent, “GraphRAG is a structured, hierarchical approach to RAG, as opposed to naïve semantic-search approaches using plain text snippets. The GraphRAG process involves extracting a knowledge graph out of raw text, building a community hierarchy, generating summaries for these communities, and then leveraging these structures when [performing] RAG-based tasks.”
LLMs alone are not enough to make the most of GenAI, said Philip Rathle, chief technology officer for Neo4j. Vector-based RAG and fine-tuning can help, but have their limits: “Neither proves the certainty of a correct answer.”
Knowledge graphs open more accurate knowledge, Rathle continued. GraphRAG AI applications are starting to leverage enterprise taxonomies and ontologies in new and evolving ways, Selz said. “By combining knowledge graphs with generative AI, users can use taxonomy and ontology to perform business process automation.” For example, with customer ticketing, “RAG provides a natural language chat interface where end users can register support requests. Simple ontologies stored in the graph guide the flow of conversation, ensuring that the RAG engine asks the right questions to gather all the required input to determine the right action.”
The graph “can even store function calls so that following a natural language support session, the RAG generates one or more Jira tickets, routing all the required information to the right people,” said Selz. “Generative AI plus semantic knowledge graphs quickly delivered new value propositions; the evolving fusion of probabilistic and deterministic technologies will yield more use cases that deliver synergistic value to the enterprise.”
Despite the relative complexity of establishing graph databases that integrate with LLMs and RAG environments, “The benefits are worth it,” said Ward. “AI and ML platforms leverage knowledge graph-structured data to ground their models in objective facts, enhancing the accuracy and reliability of AI insights. Investing in these technologies can unlock the full potential of your data, providing significant advantages.”