Retrieval-augmented generation is easy to demonstrate and hard to make good. A weekend prototype answers five questions impressively and then falls apart on the sixth, because the answer lived in a paragraph the retriever never returned. This course is about the retrieval half, which is where nearly all the quality lives.
You build a pipeline over a genuinely awkward corpus: internal documentation with tables, code blocks, deep heading hierarchies, near-duplicate pages and a changelog. Naive fixed-size chunking is implemented first so you can watch it fail specifically, splitting a table from its header and stranding a code sample from the sentence explaining it. Then you fix it, with structure-aware splitting, heading context carried into every chunk, and metadata that makes filtering possible.
Search comes next. Dense retrieval, keyword search with BM25, and the hybrid combination that beats both, followed by a reranking pass over the top candidates and the query rewriting that turns a lazy user question into something searchable. Each change is measured on a labelled question set, so improvements are demonstrated rather than claimed.
Only then does generation appear: grounded answers with citations, a refusal when retrieval came back weak, and separate evaluation of retrieval failure and generation failure. Postgres with pgvector throughout.