Building RAG systems

About this course

A course on building a search-plus-generation (RAG) system that actually answers from documents instead of making things up. It covers the full pipeline: extracting text from sources, chunking for answers, retrieval, generation with citations, and quality evaluation — with a focus on where errors actually happen.

Six modules move from "do you even need RAG" through knowledge preparation, retrieval, and generation to evaluation and operations. In "Retrieval for answers" you build working search over 100 documents with query normalization. In "Answer generation" you build a contract for empty-context behavior and claim-to-source citation. In "Evaluation" you assemble at least 80 questions with reference answers and take your first quality measurement.

For people who write code and work with their own product's real documents — you'll need a collection of at least 100 files. Requires Python and basic familiarity with LLM API calls. This isn't an "answer from the document" prompt course — you'll deal with text extraction, metadata, and metrics.

What you'll learn

• Judge when RAG is needed versus long context or plain search
• Extract text from varied source formats, including tables and images
• Chunk documents for answers rather than for the index
• Rewrite and expand a user's query before retrieval
• Rerank retrieved chunks and filter out irrelevant ones
• Build an answering prompt that ties claims to sources with citations
• Assemble a question set with reference answers and compute hit rate, coverage, and hallucination metrics
• Separate retrieval errors from generation errors when triaging failures

What you'll end up with

• A working RAG pipeline over your own collection of 100+ documents
• A set of 80+ questions with reference answers and a first quality measurement
• A report breaking down latency and cost by pipeline stage
• A finished operations project for the RAG system (paid tiers include reviewer feedback and a certificate)

Not for you if you don't have documents to work with, or if plain full-text search without generation already solves the problem.