You already know: hash maps and sorting (D1 Data structures); TF-IDF (D3 Feature engineering); context budgets and delimiting sources (D7 Context engineering, Prompt fundamentals).
Why this matters for Atlas. Retrieval-augmented generation (RAG) is the pattern behind almost every assistant that answers from a company's own knowledge: search first, then let the model answer from what was found, citing it. The model's answer can only be as good as what retrieval hands it. If the right passage isn't retrieved, the best model in the world will either guess or say it doesn't know. This lesson builds retrieval from the ground up — the inverted index and BM25, which still power most search engines — and measures exactly where word matching breaks, which motivates everything else in this domain. Lesson II (Retrieval II) takes it to production: analyzers, knowing when not to answer, speed, sharding, permissions and debugging.
Where you'll use this. Every RAG system has a retrieval step, and most production ones still include BM25: in Elasticsearch, OpenSearch, Postgres full-text search or a library like Tantivy. Knowing how it scores tells you why a query found the wrong passage, when keyword search will beat embeddings (codes, names, error messages), and what the dense retrievers in the rest of this domain are trying to improve on.
1. The RAG pipeline
The question. What actually happens between a user's question and a grounded answer?
Start from nothing. Two phases: