Let's state the objection exactly as it lands, because it always lands first. A CIO, an architect, watches the demo and says: “basically, it's RAG over your content, right?”
It's a good question, and it deserves a precise answer, not a dodge.
RAG, retrieval-augmented generation, does one thing and does it well: it grounds the AI's answer in your real content, instead of letting it invent from its general knowledge. You ask a question, it fetches the relevant passages from your documents, and uses them to answer. It's useful, it's solid, and the Practice Graph uses it internally. RAG isn't an adversary. It's a building block.
But RAG, alone, is blind to three things. And it's precisely those three things that turn an answer into an asset.
The first is structure. RAG sees text: passages, chunks, similarity vectors. It doesn't see that a practicePracticeA unit of know-how captured in Marylink: not a document but an executable structure (content, prompt, rules, style). is composed of a prompt, a rule, a style, a tool, assembled together. It retrieves pieces of documents. It doesn't manipulate typed objects you can recombine, update cleanly, reuseReuseThe same practice serving many times, across many spaces, the key measure of the Practice Graph's value. elsewhere. Change a rule: with RAG, you have to fix it in every document where it lingers. In a typed practice graphPractice GraphThe living architecture that links your practices, concepts, roles and spaces, executable by your teams and by AI., you change it once, and it updates everywhere it's used.
The second is governanceGovernanceThe roles, validation steps and reviews that ensure the quality of shared practices.. RAG returns the closest answer, not the most validated one. It doesn't know who endorsed this practice, which version is authoritative, whether an expert reviewed it, what validation stepValidation stepA stage in a publication's path (draft → review → approved), with rules that gate each transition. it's at. It has no notion of role, reviewReviewAn expert's assessment of a practice: score, comments and recommendations against criteria., or step. It retrieves the most similar text, even if it's an abandoned draft, even if the right version is elsewhere. Semantic relevance isn't quality. A practice graph knows how to tell what's validated from what's just lying around.
The third is executability. RAG retrieves to inform: it feeds a human's or a model's answer. It doesn't make the practice executable by an agent. In a practice graph, a practice is searched, combined with the right style and the right model, executed, then republished for the team, via MCPMCPModel Context Protocol: the open standard that connects an AI assistant (Claude, etc.) directly to your Marylink space.. RAG stops at “here's relevant text.” The practice goes all the way to “here's the validated method, executed, and deposited back.”
The shift comes down to two verbs.
RAG retrieves. The Practice Graph executes.
One improves your model's answer today. The other builds a capital that accumulates: typed, governed, executable, and dependent on no particular model.
That's why “isn't it just RAG?” is the right question, with the wrong conclusion. RAG is inside. But reducing the Practice Graph to RAG is mistaking the library for the practice of the house.
So the real question isn't: do I retrieve the right information.
It's: is the right practice typed, validated, and executable, by a human as much as by an agent?