Here is a question no LLM on earth can answer: "How many casual leaves do I get this year?" Not because it's hard — because ChatGPT has never read your company's HR policy. The technique that fixes this is called RAG, and it is the single most-asked applied topic in AI interviews today — at TCS, Infosys and Accenture as much as at Amazon and Google. This free course teaches it properly: chunking, embeddings, vector databases, hybrid search, reranking, evaluation and system design, with diagrams and real company questions after every lesson.
What is RAG?
RAG stands for Retrieval-Augmented Generation. Read it backwards and it explains itself: you want generation (an LLM writing an answer), you augment its prompt with text you retrieved from your own documents. The model still does what it always does — predict the next token — but now the most plausible continuation is your actual policy, because your actual policy is sitting right there in the prompt.
Nothing about the model changes. No training, no GPUs, no waiting. That's why RAG became the default architecture for enterprise AI in roughly eighteen months, and why every AI job description now mentions it.
The open-book idea
Think of two exams. In a closed-book exam you answer from memory — and when memory fails under pressure, you write something confident and plausible. (Every student knows this move. LLMs call it hallucination.) In an open-book exam you first find the right page, then answer from it. Same brain, completely different reliability.
RAG turns an LLM's closed-book exam into an open-book one. Everything in this course is engineering in service of one goal: making sure the right page reliably shows up in the prompt, fast and cheap, for whatever the user asks.
Your first retrieval — right now
Retrieval sounds heavy. Here it is in nine lines: three chunks of a handbook, a question, and one similarity call.
import chromadb
col = chromadb.Client().create_collection("handbook")
col.add(
ids=["c1", "c2", "c3"],
documents=[
"Employees get 12 casual leaves per calendar year.",
"The cafeteria serves lunch from 12:00 to 14:00.",
"Laptops must be returned on the last working day.",
],
)
hits = col.query(query_texts=["how many CLs do I get?"], n_results=1)
print(hits["documents"][0][0])Result
Employees get 12 casual leaves per calendar year.
Notice what just happened: the question and the answer share no words at all — "CLs" never appears in the document. Meaning matched anyway. Paste that chunk into a prompt and you have a working RAG system. Everything else in this course is about making this work over ten million chunks, in six languages, with permissions, evaluation and a latency budget.
Why learn RAG (the honest version)
One: it's the most-asked applied GenAI topic in interviews — "explain RAG" and "design a chatbot over our documents" are the reverse-a-linked-list of AI rounds. Two: it's what AI projects in Indian IT services actually build — document assistants, support bots, policy Q&A, contract search; if you join a GenAI project, this is likely the codebase. Three: it teaches genuinely transferable engineering — search, ranking, caching, evaluation — that outlives any particular model.
How this course works
Every lesson opens with a real situation, builds the intuition, then names the technical concept, with diagrams wherever a picture beats a paragraph and runnable Python wherever code beats prose. Each lesson ends with a Practice Zone — click-to-answer MCQs and hands-on tasks, with the companies where each pattern has been asked shown under the question.
This course assumes you know what an LLM is. If tokens, prompts and hallucinations are new, read the Generative AI course first — especially lesson 2. Its RAG lesson is the ten-minute version of what you're about to learn properly.
Course roadmap — 12 lessons
| # | Lesson | What you'll be able to answer |
|---|---|---|
| 1 | What Is RAG? | Explain RAG in 45 seconds — and when NOT to use it |
| 2 | Embeddings | Meaning as coordinates; cosine similarity; the same-model rule |
| 3 | Semantic Search | Dense vs sparse, BM25, ANN, metadata filters |
| 4 | Vector Databases | HNSW, Pinecone vs Chroma vs FAISS vs pgvector, index maintenance |
| 5 | Chunking | Size, overlap, structure-aware and semantic chunking |
| 6 | Retrieval Techniques | Hybrid search, RRF, reranking, query rewriting, HyDE |
| 7 | Build a Pipeline | Write the whole thing in ~30 lines of Python |
| 8 | Advanced RAG | Self-RAG, Corrective RAG, multi-hop, agentic RAG, GraphRAG |
| 9 | Evaluating RAG | recall@k, faithfulness, golden sets, LLM-as-judge |
| 10 | RAG in Production | Latency budgets, caching, staleness, multi-tenancy, cost |
| 11 | RAG vs Fine-Tuning | The interview question everyone gets — settled properly |
| 12 | RAG System Design | The full design-round answer, start to finish |
Company-wise RAG PYQs
After the lessons comes a section of company-wise RAG interview questions — the patterns reported from TCS, Infosys, Accenture, Amazon, Google and more — in attempt-then-reveal format, each linked back to the lesson that teaches it.
How to study this course
Lessons 1–7 build the pipeline in order and are worth reading sequentially — chunking makes no sense before embeddings do. Lessons 8–12 are the depth that separates candidates: if you're short on time, do 1, 2, 5, 7 and 12 — the pipeline plus the design answer. And build something small, even a bot over your own class notes; interviewers can tell within one question who has run a retrieval query and who has only read about one.
FAQ
Do I need to know machine learning to learn RAG?
No. RAG is engineering — search, ranking, prompting, evaluation. You need basic Python and an understanding of what an LLM does. No training, no math, no GPUs.
Is RAG becoming obsolete now that context windows are huge?
No — pasting a whole corpus into every request is expensive, degrades accuracy when the answer is buried, and doesn't scale past what fits. Retrieval and long context work together: lesson 10 covers the economics with real numbers.
Which framework does this course use — LangChain or LlamaIndex?
Neither, deliberately. You'll write the pipeline directly so you understand every step; frameworks then become convenience rather than magic. That's also what interviews test — nobody asks you to recall a framework's API, they ask what happens between the question and the answer.
What comes after this course?
AI Agents — where retrieval becomes one tool an agent can choose to use, and the system starts making its own decisions.
Ready? Start here — Lesson 1: What Is RAG? →