In July 2025, I began a new collaboration with Graphwise. I’ll be an independent curator of sorts focused on the topic of graph-oriented retrieval-augmented generation, or graph RAG, in addition to related topics.

For those of you who don’t know me, I’m a freelance researcher and analyst on emerging tech topics, retired from PwC after 20 years in that firm’s R&D units and Thought Leadership groups. I’ve covered semantics and graph databases since 2009, which is when I first came across the folks at Semantic Web Company and Ontotext, the predecessor companies that merged in 2024 to become Graphwise.

Graphwise is now giving me the opportunity to curate the graph RAG space. I’ll be able to create, co-create or contribute to all sorts of content, from blogs, to white papers, to podcasts, to collaborative events.

Along the way, I’ll be able to unpack what graph RAG is in detail, and make recommendations about what to pay attention to, as well as what to ignore. Because as we all know, Noise is burying the signal when it comes to the reality of AI and the challenges ahead. We don’t want to be buried in AI noise. A topic as important as graph RAG deserves careful curation.  

What’s Graph RAG?

Graph RAG gives language models (such as the ChatGPT large language model, other LLMs, or smaller language models, for that matter) a means of harnessing the power of knowledge graphs to augment what language models have traditionally been trained on. 

From  “Modernizing Your Data Strategy with a Graph Center of Excellence,” Graphwise white paper, https://graphwise.ai/resources/white-paper/graph-center-excellence

A good graph RAG system, in other words, makes a trusted, interoperable semantic layer approach possible. Knowledge graphs (the graph part of graph RAG) can bring together all the structured, semi-structured and unstructured data that enterprises depend on. 

Good knowledge graphs, in other words, logically connect and desilo instance data with the help of a business knowledge model, relationship-rich tiers of semantic metadata known as ontologies, taxonomies and controlled vocabularies that together help thousands of heterogeneous data sources describe themselves so that AIs can tell one entity (a customer, for example) and one context (a manufacturing company) apart from another entity and context. 

Relational databases, despite the name, aren’t all that good at relations, as anyone who’s worked with foreign keys and cryptic column headers can tell you. But it’s the relations that make disambiguation and contextual computing possible. N-dimensional graphs can connect tables to documents to images appropriate for each business context. This way, organizations can leave their relational databases in place doing the important transactional work they do, while graphs can do the precision scaling and connecting.

The RAG part of the system then makes that capability available to LLMs, agents and the rest of the native AI application layer that’s currently emerging. Good graph RAG systems make the entire process reliable, precise, trustworthy and performant.

It’s a great time to unpack and curate the nascent graph RAG space, because the point isn’t just to keep doing what we did last year. Graph RAG will open up all sorts of new territory for those who haven’t been able to take full advantage of knowledge graphs yet.. 

Why should you care about semantic graph RAG?

Semantic, standards-based knowledge graphs–Graphwise’s specialty as an end-to-end provider of semantic graph database, knowledge graph platform and metadata management tooling–can provide very large scale, heterogeneous integration and interoperation. 

Scaling data management is incredibly important. It’s the difference between spending 40 to 60 percent of your IT budget on integration because of a Bad Data tax and oodles of unnecessary complexity and duplication, or using the money you’re saving on taxes so you can pivot to take serious advantage of a new agentic AI application paradigm.  

Semantic web standards and the node-and-edge graphs that they specify–which I researched in depth and wrote about first in 2009 and still believe in, before Google began to call them knowledge graphs in 2012–are how data management scales. When one department, business unit or organization builds a proper, standards-based graph, they can share selected subgraphs with another department, business unit or organization. That’s how supply chains aren’t just digitized, but become interoperable. 

An infrastructure gateway to a representation of the real world

And interoperability is how we get to what author Kevin Kelly, the co-founder of Wired, calls a “mirrorworld.” It isn’t just a virtual or augmented reality front end, or a virtual assistant. To get to a functioning, fully interactive mirrorworld, we’ll need to represent our physical world accurately, at the data layer.

In that sense, semantic graph RAG is a gateway to the reliably digitized versions of people, places, things, ideas and facts that businesses need to automate at scale, augmenting and amplifying the efforts of humans in the loop with those machines who need our direction and curation ability.