An education company with a working LMS and a decent content library is in an awkward position: the AI capability everyone is asking about does not require replacing the platform, but nobody selling to you says so.
Retrieval-augmented generation sits on top of content. It needs to read your material and put an index somewhere. It does not need to own your courses, your enrolments or your gradebook. That makes "add AI" a genuinely separable decision from "change platform" — which is worth knowing before someone quotes you for a migration.
What RAG needs from you
Three things, and only three.
Readable content. Your material has to come out of wherever it lives in a form that parses. This is where most projects lose time. Clean PDFs and structured HTML are fine. Scanned handouts need OCR. Video needs transcription. A course whose real content is embedded in proprietary interactive objects may not be extractable at all without effort nobody scoped.
Somewhere to put the index. A vector database, hosted or self-run. Small compared to the rest.
A place to ask questions. A widget in your LMS, a separate app, or an integration into whatever students already use.
Notice what is not on that list: your LMS's internals. RAG does not care how your enrolments work.
Three routes, honestly compared
Build it yourself. Full control, and the right answer if the retrieval behaviour itself is your differentiator. The cost is not the first version — a competent engineer gets a working prototype quickly. The cost is permanent ownership: embedding models improve and require reindexing, content gets revised, model providers deprecate versions. Budget for the maintenance or do not start.
Bolt on a RAG service. Several vendors sell the retrieval layer as a product. Fastest route to something real. The questions to ask are where the index lives, whether your content is isolated from other customers', and what you can export. You are handing your content library to a third party; the terms matter more than the demo.
Move the content to a platform where it is native. Heaviest lift, and the only route that avoids running two systems that both think they own the course material. Worth considering seriously if your LMS is already a source of friction — but not worth doing for the AI alone, which is exactly the mistake vendors will encourage.
The right answer depends on a question that has nothing to do with AI: is your current LMS something you are otherwise happy with? If yes, add a layer. If it is already frustrating you, adding an AI layer on top of a system you want to leave is a way of making the eventual migration harder.
RAG or fine-tuning?
Fine-tuning adjusts how a model behaves — its tone, its format, its handling of a specialised task. RAG changes what a model has access to.
For teaching from a body of material, RAG wins on three practical grounds. Updating means re-indexing rather than retraining. Answers can cite the passage they came from, which fine-tuning structurally cannot do. And content stays separable from the model, so one institution's material never leaks into another's answers.
Fine-tuning has real uses. Answering questions from a syllabus that changes every year is not one of them.
What actually makes it good
The retrieval quality, not the model.
Most disappointing RAG deployments fail at chunking and retrieval, not generation. If the system retrieves the wrong three passages, the best model available will produce a confident, well-written, wrong answer. Time spent on how documents are split and how retrieval is evaluated pays back more than time spent comparing models.
The second thing is what happens when retrieval finds nothing. A system that says "your material does not cover this" is more useful — and far more trusted by teachers — than one that always produces something.
Sikgen AI is built around this pipeline rather than bolted onto it: see how RAG works in education, the AI tutor it drives, and the AI LMS it sits inside. If you are weighing a layer against a platform, talk to us — we will tell you if adding a layer to what you have is the better call.
Frequently asked questions
Can we add RAG to Moodle or Canvas without replacing them?
Yes. RAG needs read access to your content and somewhere to put the index — it does not need to own the LMS. The practical constraint is usually content extraction: getting material out of an LMS in a form that parses cleanly is often more work than the retrieval layer itself.
How much content do we need for RAG to be worth it?
Less than most people assume. The threshold is not volume, it is whether students currently struggle to find answers that exist somewhere in your material. A tightly-written 200-page course where nobody can locate anything benefits more than a sprawling library nobody consults.
Is RAG better than fine-tuning a model on our content?
For teaching from a body of material, almost always yes. RAG lets you update content by re-indexing rather than retraining, and it can cite the source passage — which fine-tuning cannot. Fine-tuning changes how a model behaves; RAG changes what it knows about. Most education use cases want the second.