If you've heard of AI tutors that "answer from your course material," the technology making that possible is called RAG — Retrieval-Augmented Generation.
Understanding what it is (and isn't) helps you make better decisions about which AI learning tools to trust with your students.
The problem RAG solves
Standard large language models (LLMs) like the technology behind ChatGPT have a fundamental limitation for institutional education: they answer from their training data — which is effectively "everything on the internet up to a cutoff date."
This creates two problems for a coaching institute or college:
- Off-syllabus answers. A student asks about a TNPSC topic; the LLM might give a correct general answer that diverges from your specific notes or interpretation.
- Hallucinated "facts." LLMs can state incorrect information confidently. In an exam-prep context, this is actively harmful.
RAG solves both problems by grounding every AI response in documents you provide.
How RAG works (non-technical version)
Think of it as a two-step process:
Step 1: Retrieval. When a student asks a question, the system searches your uploaded documents — PDFs, notes, past papers — and finds the most relevant passages. This is done semantically, not just by keyword: a question about "the role of oxygen in cellular respiration" retrieves passages about metabolism even if they don't contain those exact words.
Step 2: Generation. The AI reads those retrieved passages and generates an answer based on only that context. It doesn't blend in general internet knowledge.
The result: every answer is traceable back to a specific page in a specific document you uploaded.
Why this matters for institutions
Trustworthiness
Because the AI cites its source — "This answer is from Chapter 3, Page 47 of your TNPSC General Studies notes (confidence: 91%)" — students can verify the answer. They develop a habit of checking sources rather than passively accepting AI output.
Syllabus alignment
Your AI tutor is, by construction, limited to what you uploaded. It cannot give students information that contradicts your curriculum or deviates from the exam board's interpretation.
Safe deployment
You control the knowledge base. To update the AI tutor's knowledge, you upload a new document. To remove a topic, you remove the document. No retraining, no API changes.
RAG vs. fine-tuning: what's the difference?
A common alternative to RAG is fine-tuning — training a model on your specific content until it "memorises" it.
| | RAG | Fine-tuning | |-|-----|-------------| | Update cost | Upload a new document | Retrain the model ($$$) | | Traceability | Source-cited | No citations | | Hallucination risk | Low (grounded in docs) | Higher (model interpolates) | | Setup time | Hours | Weeks | | Suitable for | Course-specific Q&A | Domain-specific tone/style |
For educational Q&A, RAG is almost always the right choice. Fine-tuning is better suited for applications like customer service bots that need a specific tone.
What to look for in an LMS with RAG
Not all "AI LMS" tools use RAG. Some simply wrap a general-purpose LLM with a prompt that says "pretend to be a tutor." This produces plausible but uncitable answers.
Look for these signals that a platform's AI is genuinely RAG-based:
- Source citations with page numbers — if the answer doesn't show where it came from, it's not using your documents
- Confidence scores — genuine RAG systems can quantify how well the retrieved context matches the question
- Behaviour on out-of-scope questions — a proper RAG system says "I couldn't find information on this in your course material" rather than making something up
- Per-institution knowledge base — your documents are indexed separately from every other institution's
Sikgen AI uses RAG for its AI tutor, with per-institution document indexing, confidence scoring, and source-page citations. See it live →