Knowledge Graphs and Generative AI: Stardog's COLLIDE 2023 Keynote
In a keynote at COLLIDE 2023 in Atlanta, Stardog VP of Product Navin Sharma argued that enterprise knowledge graphs give generative AI the business context it lacks. He said the pairing can improve accuracy, cut hallucinations and make answers explainable.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Data readiness comes first. Companies that made data accessible and shareable are best placed to unlock generative AI, Sharma said.
- Context is lost in data lakes. Pulling data out of systems of record strips the business meaning that models need.
- Graphs model concepts, not tables. A knowledge graph represents suppliers, customers and orders and the relationships between them, abstracted from where data is stored.
- Federate instead of copy. Connecting to sources limits data sprawl and gives access to current information.
- Graphs plus LLMs. Sharma said the pairing improves factual accuracy, reduces hallucinations and adds explainability.
At COLLIDE 2023, the AI Connect data conference held October 3-4, 2023, at the Center Stage Theater in Atlanta, Navin Sharma, VP of Product at Stardog, argued that knowledge graphs are the missing foundation for enterprise generative AI. In a keynote titled "Charting the Data Universe: Navigating with Knowledge Graphs, Assisted by AI," he said companies that have made their data accessible and shareable are the ones most likely to unlock the technology's potential.
Sharma, speaking at his first in-person AI Connect event, framed the talk around a simple point: generative AI is moving fast, but the data underneath it is not ready.
Why is generative AI adoption growing so quickly?
Sharma opened by telling the audience they are "generation AI," pointing to the way ChatGPT, GPT-4 and Llama 2 had made it possible to generate text, images, video, audio and source code. A show of hands in the room suggested most attendees were already considering or using generative AI in their own organizations.
He cited research that he attributed to McKinsey, which found that a third of the large enterprises it studied were using generative AI in at least one business function. He also cited a Gartner projection that by 2025, 80% of business applications would have some form of embedded virtual assistant. Both figures were presented as context, and the captions garble some of the source names.
What stops enterprises from trusting generative AI today?
The tools and processes needed to make generative AI work inside large organizations are still incomplete, Sharma said, and costs are high for those who can afford them. The deeper problem is data.
Companies have invested in data lakes, but "at best you've managed to collocate a lot of data," he said. When data is pulled out of a system of record, it loses the business context that explains what it means and why it was generated.
Sharma named three problems for training models on that data: it is incomplete because it is siloed, inaccurate because it lacks context or has changed since training, and potentially a black box that cannot be verified.
"Being right sometimes just does not work when we are talking about mission critical data, sensitive data, personal information," he said, pointing to regulated industries such as banking.
What is an enterprise knowledge graph?
Stardog's answer, Sharma said, is where knowledge graphs meet large language models. A knowledge graph represents institutional knowledge: the concepts an organization talks about, such as suppliers, customers, orders, parts and compounds, the relationships between them and the properties that describe them. "We don't talk about strings in a database," he said.
He noted that the technology is not new. Pharmaceutical companies use it for drug discovery and repurposing, financial services firms for operational risk and customer 360 views, and manufacturers for digital twins and supply chain traceability. Knowledge graphs also fit modern architectures such as data fabric and data mesh, where they help make data findable, accessible, interoperable and reusable, he said, citing Gartner's data fabric guidance.
What do knowledge graphs give enterprises on their own?
Sharma listed three benefits. First, they capture institutional knowledge by mapping the metadata landscape and modeling information as business concepts, or ontologies, abstracted away from where data is stored. That lets teams put multiple lenses on the same data and annotate models with domain-specific business rules.
Second, they reduce the need to move and copy data. By connecting and federating access to sources, a graph limits data sprawl and security exposure, and gives access to current information at the moment a question is asked.
Third, they support self-service. Users without special skills can ask questions in terms of concepts they understand and explore beyond the initial question without learning a new technology, he said.
How do knowledge graphs improve generative AI answers?
When the two technologies come together, Sharma said, the graph supplies connected, up-to-date data that lets a model answer a question completely. He argued this can eliminate hallucinations, make retrieval augmented generation "intelligent" through added context, improve factual accuracy, help fine-tune models for a domain, and add verifiability and explainability.
The payoff, in his view, is data democratization. "English is the new programming language," he said, and pairing natural-language models with a graph that carries the enterprise vocabulary delivers accuracy, full context and explainability, which he tied to higher productivity, faster innovation and lower costs.
He closed with a line that summed up the pitch: "knowledge is power."
Where to hear more enterprise AI sessions
COLLIDE returns on Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta. Billed as the largest AI conference in the Southern U.S., it expects more than 1,000 enterprise data and AI leaders. Details are at /events/collide/, and registration for 2027 events is coming soon.
Frequently asked questions
What is an enterprise knowledge graph?
Sharma described it as a model of institutional knowledge: the business concepts an organization uses, the relationships between them and the properties that describe them, abstracted from where data is stored.
How do knowledge graphs help generative AI?
He said they supply connected, current data and enterprise vocabulary, which makes retrieval augmented generation more intelligent, improves factual accuracy and adds explainability.
Why are data lakes not enough for generative AI?
According to Sharma, data lakes at best co-locate data, and removing data from systems of record loses the context about its meaning and origin.
Do knowledge graphs require moving all data to one place?
No. Sharma said graphs can connect to and federate access over sources, limiting physical copies and data sprawl.
This article is based on remarks by Navin Sharma, VP of Product, Stardog at COLLIDE Data + AI Conference 2023 on October 3, 2023. More coverage: COLLIDE AI Conference.
Filed under: Context, Semantics & RAG, Data & AI Infrastructure, AI Governance & Trust
