From the Semantic Web to Agentic AI: Ora Lassila on Knowledge Graphs, LLMs, and the Architecture of Machine Understanding

Q1. In 2001, you co-authored the original Semantic Web paper with Tim Berners-Lee and Jim Hendler — a vision for web content that would be meaningful to computers and would enable AI agents to act on it autonomously. Twenty-five years later, at the Knowledge Graph Conference 2026 you received the KGC Lifetime Achievement Award for that work, and the audience could reasonably observe that large language models and agentic AI have now made a version of that vision real — but through a completely different technical path than the one you described. What does it feel like to see the goal arrive by a route you did not anticipate, and what does the Semantic Web vision still offer that the LLM-driven approach does not?

I don’t necessarily agree with the “different path” notion. The “agentic systems” we see today are not agents in the way the agent research community defined them in the 1990s and what we discussed in the original Semantic Web vision. But we are now in a fantastic position to realize the vision, given the robust natural language processing capabilities of LLMs and the mature state of ontology and knowledge graph technology. Agents — real intelligent and autonomous agents — will require technologies that have been around for quite some time: apart from ontologies and knowledge graphs, they will need planning techniques to realize true autonomy. To build trustworthy, accountable and explainable systems it is imperative we know which knowledge these systems are basing their answers on, and we must be able to audit and curate that knowledge.

Q2. You hold a January 2026 patent on “Knowledge Graph Assisted Large Language Models” — a direct combination of the two most significant threads of your career. Can you explain the core technical architecture behind that approach — what specifically the knowledge graph contributes to the LLM that the LLM cannot provide for itself, where the boundary between the two systems sits, and what classes of enterprise problems this combination solves that neither approach handles adequately on its own?

This patent started out as an experiment to make the knowledge an LLM embodies more explicit. The approach is to take a knowledge graph query (a SPARQL query) from a user, and have an LLM construct a knowledge graph that this query could run against. The LLM would take its (implicit) knowledge and structure it as a knowledge graph where this knowledge is now explicit and can be audited and curated.

Generally, I would like to use LLMs for what they are really good at: interpreting human language. And since we are not really in control of how the LLMs were trained, what training data was used, I would much rather see the source of answers be something I can control. Also, I am not talking about RAG here, since that technique is still subject to the known shortcomings of LLMs.

Q3. You have articulated what is being called “Ora’s Rule of Knowledge Graph Implementation” in the context of the ongoing RDF versus LPG debate. Can you state that rule precisely and explain the technical reasoning behind it — and more broadly, for an enterprise architect who is currently deciding between an RDF-based knowledge graph and a labeled property graph for a production AI application, what are the most important criteria that should drive that decision in 2026, and where does the conventional wisdom in that debate get it wrong?

The conventional wisdom wrongly thinks all graphs are somehow created equal. This originates from the naive idea that a knowledge graph is a mere data structure. RDF graphs come with the whole “Semantic Web stack” of standards (OWL, SPARQL, SHACL, R2RML, etc.). To replicate that functionality with LPGs requires those technologies to be re-built. LPGs have no formal semantics, RDF has very carefully designed formalization. This makes it possible to, for example, predictably merge RDF graphs without any additional engineering (merging happens, for example, when you load an external ontology). SPARQL supports federated queries. R2RML makes it possible to take relational databases and either migrate their data to an RDF graph or treat them as virtual graphs. LPGs, on the other hand, typically support graph analytics: this makes them more suited for specific, narrow use cases that require lots of computation.

I think the “shorthand” decision criteria is this: if you are building a knowledge graph, use RDF; if you are building something else, think carefully about what you need.

Q4. GraphRAG — using knowledge graphs to augment retrieval-augmented generation — has moved from research curiosity to active enterprise deployment over the past two years. Previously as Principal Graph Technologist at Amazon Neptune and someone who has spent decades thinking about knowledge representation, what do you see as the most significant unsolved technical problems in GraphRAG implementations today — the failure modes that practitioners encounter in production that the published benchmarks and research papers consistently miss?

GraphRAG is essentially a technique to use a graph to make LLMs perform better. It does not necessarily mean you are using a knowledge graph; often the graph is used to connect related “chunks” of text for better focusing the LLM to those parts of the source text where the answer might be. I am more interested in the “inverted” scenario: how can we use LLMs to make knowledge graphs work better? LLMs offer us a fantastic means of interacting with the user, something that the original Semantic Web vision (and autonomous, intelligent agents in more general) needed. Simplest scenario is to have an LLM write queries against a knowledge graph, but that is just the beginning.

Q5. You have worked on knowledge graphs, ontologies, and semantic technologies across Nokia Research, State Street, AWS, and now Accenture — spanning mobile computing, financial services, and cloud infrastructure. That longitudinal, cross-domain perspective is exceptionally rare. Looking at the current moment, where LLMs are producing fluent outputs while often lacking verifiable grounding in structured knowledge, what do you believe is the most important architectural investment enterprises should be making today in their knowledge infrastructure — and what will separate the organizations that successfully integrate symbolic and neural approaches from those that discover, too late, that their AI systems are confidently wrong at scale?

To me, the key investment is enabling the use of ontologies and knowledge graphs to be used as a means of “grounding” the LLMs. Many people talk about “context” and “context graphs”. My take on this is the following: everything you know should be considered a context, and the context graph is your knowledge graph. It should live in the “semantic layer” of your architecture, a layer that insulates applications and agents from your legacy data sources (and performs data integration). Making all your agents “see” a unified, uniform view of your enterprise data will, in part, prevent different agents from basing their answer on different source data. And mind you, not everything is about agents, there is still a role for more conventional applications, made better by virtue of the semantic layer being their way of accessing data – and accessing services like ontological reasoning.

Qx. Anything else you wish to add?

Despite what many people believe, AI did not get its start just a few years ago. AI has been studied since the 1940s, and from the vast body of research (and practical application) we have many techniques and technologies that are important, crucial even, to make the “new” AI work and succeed. In summary, the good news for those working with new AI is that we have solved some key issues around semantics and context needed to get it to work. And conversely, the new AI makes these old techniques easier to use.

We should also remember some of the interesting philosophical criticisms of AI people have voiced over many years. Weizenbaum’s ELIZA (the first chatbot, in the 1960s) and Searle’s “Chinese room experiment” (from 1980) are highly relevant today, and we should be asking the same questions these researchers were asking back then.

…………………………………………………………………..

Ora Lassila  

Ora Lassila is a Finnish-born American computer scientist and a pioneering figure in knowledge graphs, ontologies, and the Semantic Web. He earned his Ph.D. in Computer Science from Helsinki University of Technology (now Aalto University) and has built a career spanning academia, research, and industry.

Lassila has been conducting research into the Semantic Web since 1996, and was co-author — alongside Tim Berners-Lee and James Hendler — of the landmark article “The Semantic Web,” published in Scientific American in 2001, now the most cited paper in the Semantic Web field. His foundational contributions include co-editing the original RDF Model and Syntax specification, launched during a stint as a Visiting Scientist at MIT’s Laboratory for Computer Science in 1996–1997.

He served as an elected member of the World Wide Web Consortium (W3C) Advisory Board from 1998 to 2013. Over his career, he has held senior roles at Nokia Research Center, Pegasystems, State Street, and Amazon Web Services. 

In April 2026, he joined Accenture to define what Knowledge Engineering means for this new generation of AI; he focuses on ontologies and knowledge graphs. He also serves as co-chair of the W3C RDF & SPARQL Working Group, which is defining the next version of the RDF standard. He is the author of more than 100 conference papers and journal articles.