BuildRAG

Developers learning retrieval-augmented generation are often taught through tutorials that begin with a vector database, which makes the two concepts feel inseparable even though a vector database is only one search tool among several. Many RAG frameworks hide the underlying search and ranking decisions inside large abstractions, making it hard for learners to understand what each component actually does. This leaves developers without a clear picture of how to choose between file search, structured database lookups, and embeddings depending on the question being asked, and without a way to evaluate whether their retrieval system actually works beyond a single summary score.

BuildRAG is a course that teaches RAG from first principles, starting with the smallest working version of a system before layering on complexity. It walks through discovering sources with glob, searching contents with ripgrep, structured lookup, embeddings and brute-force vector search, PostgreSQL with pgvector, hybrid ranking, reranking, and generating grounded answers with citations. Later sections cover agentic retrieval where a model chooses read-only search tools, plus production concerns like RAG security, caching, monitoring, serving with FastAPI, and testing. The course emphasizes comparing search methods against the same set of 40 test questions and reading failed examples directly rather than relying on a single aggregate score, using focused libraries instead of large frameworks so the search and ranking logic stays visible.