During the Q&A following the talk I gave recently on the characteristics of a true semantic layer at the Graphwise AI Summit, an attendee asked how to explain what I’d presented on, including terms like graph RAG and knowledge graphs, to leadership. I gave an answer that emphasized not delving into explanations of graphs, for example.

With this post, I will now try to craft an answer that assumes leadership genuinely wants to learn a bit about semantic graphs and AI.

Background: A formula for explaining important tech to execs

Back during my days at a Big 4 consulting firm writing for a C-suite audience, my boss’s boss, the CTO of the US firm, recommended our emerging tech research publication team answer these five questions in each blogpost:

1. Why should I care?
2. What is this?
3. How’s this different from X?
4. Why would I do this?
5. What’s the downside?

I like this list of questions a lot because it assumes your target audience won’t read about emerging tech unless they understand you’re demonstrating consistently that you’re respectful of their time.

Imagine now we’re speaking about AI strategy to an audience with members of our company’s C-suite in attendance. During the Q&A, they express curiosity about graph RAG and related concepts.  We have ten minutes to explain the main concepts. What do we say? Let’s see if we can answer the former CTO’s Questions 1-5 sufficiently within ten minutes.

Why should the C-suite care about graphRAG?

Currently, AI’s missing at least half of the information it needs to deliver trustworthy responses or take actions correctly and appropriately on our behalf. 

Most gen AI hallucinates, particularly when it’s casting about for a straight answer to a question and what it’s been trained on is not explicit or detailed enough. Findability requires rich metadata that can help agents navigate to the right answer. Unfortunately, data science teams are in the habit of stripping explicit relationship-rich information down, rather than enriching it. 

Graph RAG uses a knowledge graph approach to deliver desiloed, contextualized AI-ready, trustworthy inputs to the AI that can fill in the gaps.

Gen AI on its own is a black box–if the input to the black box is garbage, the output will be garbage too.

What is graph RAG?

Good databases for decades now have stored useful, relevant, up to data facts the business can use. RAG (=retrieval augmented generation) is a way gen AI or multiple agents can retrieve responsive, trustworthy facts from databases so they can give solid answers to questions or take the right actions.

Graph RAG, by contrast with plain RAG, uses a graph or network database so that the information can be desiloed (not trapped in blind file directories or application code) and contextualized (describing how people, places, things and ideas are related to one another). It’s like a network described in a machine and human-readable way in metadata, with more than just people in it.

Ordinary databases silo the information in sparsely described tables and files. Vector databases merely approximate how closely related things are to one another in a space–they’re not precise or reliably reusable.

With solid, factual, richly connected and described context in graph databases, machines can see the whole picture and find the pieces of the picture they need to share with the user to respond to a question or direct an agent to take an action. This same context can be reused and refreshed repeatedly.

How’s this method different from what the ordinary RAG vendor is pitching?

Ordinary RAG vendors only know how to deliver statistical guesswork because they’re only using statistical machine learning methods to predict the next word in a string of words.

Graph RAG providers, by contrast, add deterministic, highly articulated and logically connected methods that businesses have relied on for decades. Information systems need more than just statistics to support business decisions and operations–they need hard facts and rules too, and they need to tap a network of networks – interconnected business contexts – to retrieve the right facts and rules for the right purpose, to the right people, at the right time and place.

Why would we do this?

In most companies, employees are uploading documents to something like SharePoint. But SharePoint without graphRAG doesn’t share well. There’s not enough context for people to find what’s relevant to them. As a result, SharePoint sites are often neglected or even abandoned.

It’s hard to find what you’re looking for in SharePoint. Gen AI or agents will have the same trouble humans do finding what they’re looking for.

In most companies, employees copy tabular datasets and share them in a data lake, or integrate them into a lakehouse. Even if the integration process is reasonable, it won’t scale the way a graph with a context model does.