Building Sophisticated A.I. Chat Models Using GraphRAG

The latest thing in A.I. Chats is GraphRAG, something we've been expanding models in early adopter cases to extend far beyond what we've been able to do with A.I. chats in the past.

How Does GraphRAG Improve on A.I. Chats?

In the past when you've chatted with A.I., it was mostly a 2-way conversation, you ask a question, that goes to a data repository, that provides a response. Question in, Answer back.

The concept of "graph" is a cascading thread of questions and responses, where the answer to one question can be the start of a conversation on a whole new topic or broad new extension of a topic.

Graph in the Microsoft world is not new, it's the foundation of how all Microsoft 365 content is webbed together (from your logon, to your user account info, to your Outlook mail content, to your Teams folders, to your Intune endpoint management content, etc)

How We've Used GraphRAG in A.I.

When we're working with GraphRAG in chatting with content, each chat response can lead us to chatting and uncovering other information down a completely different path of data.

In one use case where we're working in a medical research environment:

• The first chat query is simply to summarize stored knowledge in the system about symptoms of a patient
• The first response from A.I. is a series of potential conditions that could be investigated
• Before going down the rabbit hole on any one of the eight potential areas initially uncovered, we chat with A.I. across 2-3 additional sequences to add in more information that A.I. narrows down the list to 3 potential conditions
• We put a "pin" on that level of 3 conditions and start diving down the Graph on one condition, to read up and better understand if that condition seems plausible.
• We can then quickly go back up the tree to the other 2 and dig deeper on these conditions
• Diving down one of the 3 paths, we find a completely different set of 4 potential conditions that expands out the thinking that helps us broaden our thinking, so we "pin" an additional branch and start expanding on questions down the 4 new paths.
• Down one of these paths we find a completely different set of input areas that we hadn't addressed in all previous questions and chats, so we "pin" this juncture and start to ask questions and uncover a completely different angle to the knowledgebase, gathering more information to see if these additional paths might lead us to more insight to information.

Using GraphRAG in Retail and Supply Chain Models

In another use case, we expanded beyond internal sales, manufacturing, and inventory data housed in corporate an ERP system and expanded the Graph to include supplier data as well as weather predictions, economic data, and current events news.

By injecting multiple datasets into the A.I. analysis, we've been able to "see" far beyond just internal data produced from a traditional ERP system, and look at impacts on sales caused by economic events, or hindered (or enhanced) by upcoming weather patterns (hurricanes or snowstorms that can slow down sales in some industries, or could increase sales in other industries)

GraphRAG in these scenarios injects new data sources that causes different outcomes on datasets that produces different results. Instead of statically thinking a question only has 1 answer, when additional data is added to the overall information set, the response changes, forming different paths based on the weight of incoming information.

Early Experiments and Experiences

GraphRAG is still very new in how it operates, and the learning curve for GraphRAG practitioners takes a bit of time to navigate the web of information to remember where you're at, what your train of thinking was, why you were heading down a particular path and what other paths you were thinking you wanted to explore further.

So it's still new in what can be done with the tool ("a science project"), however those that have worked with my company in working with Azure A.I. already have the basis of tech in place and some initial grounding on what can be done with the two-dimensional A.I. we've had the past year and a half, so this is the next level...

Wrap-up

A.I. has always been about digging into data to get output and answers, however GraphRAG provides us with more than just a simple question with a resultant answer. GraphRAG provides us a much broader canvas to work from, where answers to questions can lead to a whole new series of questions that drives further down different paths.

And rather than digging deep down a single rabbit hole, digging deeper and deeper, that hole can expand to many branches, with different outcomes based on the variation of your questions.

Where we thought chatting with A.I. was a bit of an art before, now it's gotten even more sophisticated, with so many more angles to pursue.

Bottomline as I've shared in the past, you don't want to "wait" until A.I. is perfect and "ready" to jump in and start using it, the whole A.I. world evolves monthly, and it is a lot easier to learn and experience a little at a time than try to jump in months later with SO much to try to catch up one down the line.