What is RAG, plainly?
Fetching text relevant to a question and putting it in front of the model before it answers, so the answer comes from your documents rather than from what the model absorbed in training. It is the most reliable single improvement available for factual work, and its failures are mostly search failures.
Why do you say the model is rarely the problem?
Because when we take systems apart with students, the relevant passage usually was not in what the model received. It was never retrieved, or it ranked far below the cutoff, or it was truncated. A model cannot answer from text it was not given, and swapping in a better one changes nothing about that.
How much does it cost?
A group batch of five to eight students is USD 100 a month and one to one is USD 150 a month, billed in US dollars, the same flat figures everywhere we teach outside India. No enrolment fee and no minimum term, and the first class is free without a card.
Do I need machine learning knowledge first?
You need to write Python comfortably and to have handled messy real data, which matters more here than any modelling background. Most of this course is engineering: splitting documents sensibly, combining two kinds of search, measuring whether the right passage came back. The mechanism page is a better prerequisite than a maths course.
Is RAG obsolete now that context windows are enormous?
No, and the claim is worth addressing directly. Putting an entire corpus in the window is expensive, slow, and measurably worse at locating one specific fact than good retrieval. More room in the window is not the same as attention landing on the part that mattered, and cost scales with every single query.
What corpus should a student use?
One they know well enough to grade. Personal notes, a rulebook for a game they play, a public dataset in a field they follow. Familiarity is the requirement, because evaluating retrieval means knowing what the right answer was, and a student who has not read the documents is guessing at every measurement.
Can this be taught to teenagers?
Yes, to teenagers who write Python confidently, and it lands well because the failures are so concrete. A fifteen year old who has watched their own system quote the wrong section of a rulebook understands the retrieval problem better than most adults who have read about it.
How do you evaluate a RAG system?
In two separate halves, and separating them is the whole method. First, retrieval alone against a labelled set of questions with known correct sources. Only once that is acceptable do you judge generated answers. Teams that judge only the final answer cannot attribute a failure and end up tuning prompts against a search problem.
What about private or sensitive documents?
In the course, nothing sensitive goes anywhere: students use their own material or public data, and we do not have them wire real confidential corpora into third-party services as a teaching exercise. What we do teach is how to reason about where documents travel, which is a decision people frequently make without noticing.
What happens after I send the form?
A mentor calls to arrange the free class. Bring a corpus you know well and a few questions you would want answered from it, and we will look at how it should be chunked. That conversation is usually more informative about the difficulty of your particular problem than anything we could tell you in advance.
