Building LLM Applications: APIs, RAG, Embeddings, and Tool Use

Open the source of any LLM application that actually holds together in production, and the impressive prompt is the smallest thing in it. What's around the prompt is the engineering: the schema that constrains the output, the index it consults for facts it doesn't already know, the tool that takes the action, the evaluation that catches the regression before it ships. That surrounding code — not a clever incantation — is what turns one good answer into a system you can stand behind. This course teaches that surrounding code as a small set of durable primitives. The API surface as a stable contract under model names that rotate every few months: messages, structured outputs, streaming, prompt caching, and the token-and-latency math that comes with them. Embeddings, ANN indexes, hybrid retrieval, and cross-encoder reranking as the standard way to ground a model in your own data, with chunking and ingest treated as first-class engineering rather than a notebook hack. Function calling, the agent loop, idempotent tool effects, and the Model Context Protocol as the standard way to let a model act on real systems, with capability scope and human-in-the-loop where reversibility demands it. And the engineering loop — golden datasets, LLM-as-a-judge with its known biases, the RAG-eval triad, tracing, and lethal-trifecta defenses — that turns "it worked once" into evidence you can trust. You will finish with five worked builds: doc Q&A with citations, structured extraction at scale, an internal assistant with read-only tools, a customer-action agent with mutating tools, and a multi-step research agent. Across thirty-nine lessons you build the habit underneath all of them: given a use case, reach for the right primitives, compose them, evaluate them, ship.