Kirti Kumar Sahu · August 2, 2026 · 10:00 · 2 min read
I've been taking notes on RAG lately, and it's worth writing up in my own words instead of leaving it buried in a notes app. Splitting it into a few parts since there's more here than one post's worth.
RAG stands for Retrieval-Augmented Generation. It shows up everywhere now — customer support, medical, legal, finance — because it's one of the most practical ways people are actually using AI in production, not just in demos.
A good example of why: get a blood test done, and the app doesn't just email you a PDF. It summarizes the report, answers questions specific to your numbers, and gives suggestions based on your actual data. ChatGPT or Gemini alone can't do this — they've never seen your report. That's the gap RAG fills.
Student 1 studies from a stack of books, memorizes everything, then sits a closed-book exam tomorrow and answers purely from memory. That's a normal LLM — GPT, Gemini, Claude. Trained once, then it generates answers from whatever it learned during training. That's the "Generation" half of RAG.
Student 2 studies the same books, but the exam is open-book. They can flip back to the source, verify a fact, check something in real time. Student 2 does better, because they're not relying on memory alone — they're looking things up as they go. That's RAG.
RAG isn't a separate model or a product you buy. It's a technique: give the model access to real, current data at the moment it's answering, instead of trusting it to remember everything from training.
Say your flight gets delayed and you ask a generic chatbot "why is my flight delayed?" It has no idea which flight is yours — you get a generic non-answer. But if the airline's support chatbot is RAG-based and you're logged in, it already has your PNR, your flight, your delay reason. Same question, a genuinely useful answer, because the model was handed the right context before it answered.
Next part: why this actually matters — hallucination, cost, and the knowledge a model was never trained on.