RAG

Production RAG Systems: A Hands-On Course

Course Overview

Large language models are fluent, but they do not know your organization's documents, cannot show where an answer came from, and will confidently invent facts. Retrieval-Augmented Generation (RAG) solves this by finding the right passages in your own content first, then having the model answer only from them, with citations.

A RAG demo takes an afternoon. A RAG system that is accurate, secure, fast, and affordable in production takes much more, and that is what this course teaches. Across fourteen modules you build one project, ShopSphere, a support assistant for a fictional e-commerce company. It starts as a few dozen lines of code and grows into a production service that ingests messy real documents, respects who may see what, answers with citations or honestly refuses, and blocks quality regressions before they ship.

Every module teaches through runnable code, with a plain-language explanation after each example, a hands-on lab, and interview questions. All tools have free options: Python, open-source embedding and reranking models, Qdrant, FastAPI, and your choice of Groq, Gemini, or a fully local model with Ollama.

The course is organized into four stages: Foundations, covering Modules 1–3: RAG Foundations, Document Ingestion, and Chunking; Retrieval, covering Modules 4–8: Embeddings, Vector Stores, Search and Hybrid Retrieval, Query Understanding, and Reranking and Context Assembly; Answers and Quality, covering Modules 9–11: Generation and Grounding, RAG Patterns, and Evaluation; and Production, covering Modules 12–14: Production Engineering, Advanced and Emerging Topics, and Capstone: A Production RAG Service.

What You Will Learn

• Build a complete RAG pipeline and explain the trade-off at every stage
• Turn PDFs, web pages, spreadsheets, and tables into clean, permission-aware data
• Choose chunking, embedding, and indexing strategies, and justify them with measurements
• Combine keyword and vector search, rerank results, and rewrite conversational questions
• Generate cited, validated answers that refuse when the evidence is weak
• Defend against prompt injection, data leaks, and cross-customer exposure
• Evaluate quality with real metrics and block regressions automatically in CI
• Run RAG in production with caching, tracing, cost control, and graceful failure handling
• Judge when advanced techniques such as GraphRAG, agents, and fine-tuning are worth their cost.

Prerequisites

• Intermediate Python: comfortable with functions, classes and dataclasses, type hints, comprehensions, modules and imports, exceptions, JSON and files, and virtual environments with pip.
• Understanding of LLMs:what prompts, tokens, and context windows are, why models hallucinate, and how to call an LLM API from code. No knowledge of model training or internals is needed.

-Developer basics: the command line, Git, and what an HTTP API request and response are.

• Math: high-school level. Vectors, similarity, and the statistics used for evaluation are taught from scratch.
• Setup: a computer with 8 GB of RAM (no GPU needed), Python 3.11 or newer, a GitHub account, and a free Groq or Google AI Studio key, or Ollama installed locally.