RAG: Retrieval-Augmented Generation from First Pipeline to Production

Learn to make a language model answer from your own documents. Build one RAG system over a small company handbook, step by step: chunking, embeddings, vector search, prompts, hybrid search, metadata filters, re-ranking, query rewriting, parent documents, compression, conversations, fresh data and caching, agents, evaluation, hallucination checks, guardrails, serving and a finished web service. Every example runs on your laptop with a free local model and shows the output it actually produced.

Who it is for

Developers who know a little Python and want to build question-answering over their own documents with a language model, and know how to measure and trust the result.

Before you start

• Basic Python (functions, lists, dictionaries, running a script)
• A laptop with about 8 GB of RAM and 6 GB of free disk (for the embedding model and a small local language model)
• No machine-learning background is needed; embeddings and language models are explained from scratch
• No paid account or API key is needed; the course uses the free Ollama runtime on your own computer

What you will be able to do

• Explain why a language model needs retrieval, and build the retrieve, prompt and generate pipeline from scratch
• Chunk documents, create embeddings and search them with exact vector search or a vector database
• Improve retrieval with hybrid search, metadata filters, re-ranking, query rewriting, parent documents and context compression
• Handle conversations, changing documents, caching and multi-step questions with a tool-using agent
• Measure a RAG system with retrieval metrics, fact checks and a judge model, and catch unsupported claims
• Defend against prompt injection and personal-data leaks, serve the pipeline over HTTP and reason about latency and cost