A new study from the National Innovation Centre for Data has shown combining large language models with graph technology can significantly improve AI accuracy, completeness and trustworthiness.
Research at a glance
- 510 complex questions tested.
- 80% improvement in truthfulness versus standard vector RAG.
- More than 2x increase in questions answered.
- Uses automatically generated graph structures, making enterprise adoption practical.
Introduction
Newcastle, UK – New research from the UK’s National Innovation Centre for Data (NICD), which is powered by Newcastle University, has found that combining large language models (LLMs) with graph database technology can dramatically reduce AI hallucinations and omissions. The research highlights a potential solution to two of the biggest barriers preventing organisations from confidently deploying AI for real-world use cases.
The study tested AI systems against 510 complex questions from a recognised benchmark dataset designed to challenge reasoning across multiple sources of information. Researchers found that an AI system combining a GPT model with graph database technology achieved the highest level of truthfulness, significantly outperforming both conventional Retrieval-Augmented Generation (RAG) approaches and AI models relying solely on their training data.
Why AI hallucinations and omissions matter
While LLMs have rapidly transformed the way organisations access information, they remain prone to two critical weaknesses: hallucinations and omissions.
Hallucinations occur when an AI system confidently generates information that is incorrect, unsupported or entirely fabricated. Omissions happen when an AI response leaves out important facts or context needed to answer a question accurately. In both cases, the answer can appear credible and convincing, making it difficult for users to identify mistakes.
For businesses, these issues can create significant risks in areas such as compliance, legal research, customer service, due diligence and knowledge management. As organisations increasingly rely on AI to analyse complex information and support decision-making, improving the accuracy, completeness and reliability of AI-generated answers has become a major focus for researchers and industry alike as otherwise, the consequences can be significant, particularly in regulated or high-stakes environments.
The research suggests that adding graph-based context helps AI systems better connect information across multiple documents and entities, leading to more complete and reliable answers.
Key findings
Compared with a standard vector-based RAG system, the graph-enhanced approach:
- Increased AI truthfulness scores by approximately 80%.
- More than doubled the number of complex questions answered, increasing answer rates from 29% to nearly 66%.
The findings indicate that organisations can improve AI reliability without fundamentally changing how employees interact with AI tools, instead strengthening the information retrieval process behind the scenes.
"The work of AI experts at the UK National Innovation for Data has shown that integration with a graph database can significantly reduce the two most significant problems holding back the exploitation of LLMs for real-world applications: hallucinations and omissions. This is especially important for organisations deploying LLMs in applications where regulatory compliance and the avoidance of reputational or financial damage is key."
Professor Paul Watson – Director, National Innovation Centre for Data
How the research was conducted
Researchers evaluated three different approaches to answering 510 complex questions drawn from a recent peer-reviewed research benchmark designed to test reasoning across multiple documents and sources.
The study compared:
- A traditional vector-based agentic RAG system.
- A graph-enhanced agentic RAG system combining semantic search with document structure-based graph tools.
- A standalone LLM with no retrieval tools.
Unlike many previous approaches, the graph technology used in the research was automatically generated from document structures such as article titles, sections, paragraphs and links. This demonstrates that organisations can benefit from graph-enhanced AI without needing to invest in highly complex or manually curated knowledge graphs, making it easier for organisations to apply to internal knowledge bases, policies, reports, technical documentation and operational data. This approach is practical and scalable for organisations looking to improve enterprise AI deployments.
What this could mean for business
As businesses continue investing in AI-powered customer service, employee assistants, research tools and decision-support systems, reducing hallucinations and missing information could help:
- Improve operational efficiency.
- Reduce time spent verifying AI-generated responses.
- Lower the risk of costly errors and misinformation.
- Increase employee confidence in AI-powered tools.
- Improve customer experiences through more accurate answers.
As organisations across the UK race to deploy AI assistants, internal knowledge bots and customer-facing chatbots, the research suggests graph-enhanced AI could help bridge the gap between experimental AI and enterprise-grade systems that can be trusted in high-stakes environments. By reducing misinformation and helping AI systems retrieve more complete answers, the approach has the potential to improve productivity, reduce business risk and accelerate the adoption of AI across regulated industries.
“LLMs are amazing, but their tendency to hallucinate introduces significant risks to businesses, even as the token costs continue to mount. In an agentic world, where autonomous systems can make significant decisions over time, this cannot hold.
NICD’s landmark study demonstrates how organisations can make agentic systems much more dependable, while also reducing end-user costs. It shows that better models alone cannot solve the problem, but that the right data at the right time can provide significant benefits in terms of accuracy, responsiveness, and cost.
Eschewing traditional vector-only approaches, the NICD team equipped their agents with graph data that provided high-value context to the model. In benchmarking, this solution easily outpaced other solutions, being both better and cheaper. They have shown that enterprises no longer need to compromise on the quality of their AI, while they can compromise on cost. This is an unusual and highly welcome finding.”
Dr. Jim Webber – Chief Scientist, Neo4j
Supporting responsible AI innovation in the UK
Researchers conclude that AI systems should not rely solely on semantic search when answering complex questions. Instead, supplementing AI with graph-driven document relationships can help retrieve relevant information more effectively, reduce unsupported responses and improve overall answer quality. This is particularly relevant for applications such as due diligence, regulatory compliance and technical support.
While the researchers note that further testing across additional datasets, AI models and real-world enterprise environments will be important, the results demonstrate the potential for graph-enhanced retrieval to play a significant role in the next generation of enterprise AI systems. The study was conducted using a single benchmark dataset and one LLM model (GPT 5.4), highlighting opportunities for future research and broader validation.
Find out more
The study, Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation, was carried out by NICD and funded by Neo4j.
You can read the pre-print on arXiv and catch NICD’s data scientists presenting the findings of this paper as part of the 3rd International Workshop on Data Management Opportunities in Bringing Agents with Graph Data at the 52nd International Conference on Very Large Data Bases in Boston, MA, USA: Aug 31st - Sep 4th, 2026.
The National Innovation Centre for Data is always open to exploring collaboration opportunities. If you would like to explore an idea you have, please don’t hesitate to reach out to us. We’d be delighted to discuss how we can assist you in your data and AI journey.
