Yes, it uses RAG
Somebody looked at Ghost and called it RAG. They were describing a real mechanism inside it and mistaking that for the product. Here is the line, and here is the part of it I cannot prove yet.
My first instinct was to argue. That would have been a mistake. It retrieves, obviously it retrieves, and denying that would only tell a technical reader how much to trust the rest of what I say.
So here is the honest version. RAG is how Ghost reads. It is not what Ghost is.
What retrieval actually does
The pattern is simple and it earned its popularity. A question arrives, the system finds the passages most likely to be relevant, hands them to a model, and the model answers with that material in front of it instead of from memory. It is a very good answer to the problem of a model that does not know your stuff.
The graph variants change the retrieval structure, not the shape of the idea. They pull entities and the relationships between them instead of whatever chunks look similar. That gives better recall, and it lets you reason across a whole body of material rather than a handful of passages. Better retrieval. Still retrieval.
Which is exactly why the graph cannot be my argument. If I lead with a connected graph and models reasoning over it, a technical reader has already filed the product before I get to the part that is different.
Where the line actually is
Retrieval changes the answer. What I am building has to change the operating state of the company.
A retrieval system is finished when the answer is good. The unit it produces is a response, and its scorecard is about that response: was the right passage found, is the answer faithful to it, did the person get what they asked for. All real problems, all solved at the moment of the answer.
The unit I care about is not an answer. It is a decision, with everything attached to it: the evidence that led there, what was believed and how confidently, who agreed or overruled it and why, what was actually done, and what reality said afterward. That object has a life after the answer, and it is dated, so you can ask what was believed in March rather than only what is believed now.
That is the whole distinction, and it is not a claim about being smarter. It is a claim about what gets kept.
Why somebody can still look at it and say RAG
Three reasons, and I think all three are fair.
The first one is my fault. I kept leading with the graph, and graph is the same word every retrieval product uses. Say it first and a technical reader has already put this in that box before hearing what is different about it. The graph is how the record holds together. It is not why the record matters.
The second is that the visible behaviour is identical. You ask, it gathers, it reasons, it answers. From the outside that is indistinguishable from good retrieval until you can see what got written back and how it changed the next run.
The third is the honest one. The part that makes it different is the part I have proven least.
The five things that would settle it
If somebody wants to know whether this is more than retrieval, these are the questions I would want them to ask, and I would rather publish them than be asked them.
The state transition. Can it show how a specific new piece of evidence changed a belief it already held, rather than quietly overwriting the old one?
Correction inheritance. Can it point at an operator correction that changed a later ranking or recommendation, without anybody re-explaining it?
Outcome attribution. Can it connect a recommendation to the action taken and the result that followed?
Cross-canvas continuity. Is the organization on one canvas literally the same object on the others, or a copy that will drift?
The model swap. Can the reasoning engine change while the evidence, the corrections, the outcomes and the standards stay exactly where they are?
Which of them pass today
One, four and five pass. Beliefs carry their evidence, their source and their dates, and a change is a transition rather than an overwrite. The objects are shared across the canvases by construction rather than by sync. And the model is swappable by design, which is the easiest of the five to demonstrate because it is architectural rather than accumulated.
Two and three do not pass yet, and I am not going to write around that. Corrections are captured with their reason and their author. They are not yet feeding back into scoring, which means today a correction is preserved rather than applied. Outcomes are attaching as cases close, and there are not enough closed cases for the number to mean anything.
The architecture is past retrieval. The learning is not proven at volume. Both halves are true, and I would rather say both.
The standard I am holding myself to
The temptation is to define learning as storing the correction, because that part is done and it sounds like the same thing. It is not the same thing.
A system has learned when a correction changes a later behaviour without the person who made it having to remember it and say it again. Until that is true, what exists is a very well kept record and an honest intention. I have the first. The second is close, and close is not the same as done.
What would end the argument
Not a better diagram. One trace, start to finish: a signal that arrived, the belief it formed, the human who overruled it and why, the action that followed, the outcome that landed, and the next recommendation visibly different because of all of it.
That is one artefact, and it settles the question in a way no architecture argument can. It is the next thing I owe anybody who asks.
RAG is how it reads. The record is how the company learns. The second half is the one I still have to show you.
The companion pieces are The model is not the memory on what the model cannot supply on its own, and Evidence compounds on why the accumulated half is the only half that is yours.