Build a RAG serving system from scratch.
You'll build the system that takes a question, finds the passages that answer it, and lets a language model reply from those sources. This course teaches the ideas, not the implementation: by the end you'll understand every hop well enough to write the serving system yourself, and to say where each millisecond goes.
Unit 0, Why RAG
What retrieval-augmented generation is, the hallucination it fixes, and the pipeline in one picture.
What RAG Actually Is
Answer from sources, not from memory. The whole idea in one picture.
Why Models Make Things Up
Memorized facts, a knowledge cutoff, and why retrieval fixes both.
The Pipeline, End to End
Embed, search, augment, generate; where the time and money go.
Unit 1, Embeddings
How a sentence becomes a vector, how closeness is measured, and how a document is split for retrieval.
Unit 2, Retrieval
The vector store, nearest-neighbor search, the metric that grades it, and mixing in keywords. All self-contained.
The Vector Store
Where the vectors live and how you look one up.
Nearest-Neighbor Search
Brute force to approximate; the top-k passages.
Recall@k, the Metric
How many of the truly-relevant passages you caught. Interactive.
Hybrid Retrieval
Keywords plus vectors; when each one wins.
Unit 3, Augmentation
Pasting passages into the prompt, the limited space they fit in, and why their order matters.
Unit 4, Generation
The grounded answer and its citations, streaming it back, and the failures retrieval does not fix.
Unit 5, Reranking
Why the first pass is sloppy, the scout-and-judge split, and what the second pass costs you.
Unit 6, Serving
The latency budget across all four hops, the embedding cache, whether rerank earns its cost, and the whole service.
The Latency Budget
One p99 number to defend, split across four hops. Interactive.
The Embedding Cache
Don't re-embed what you've seen; hot queries, hit rate.
Does Rerank Earn Its Cost?
Quality bought vs latency spent; the measured verdict.
The Whole Service
Every piece, under one budget. Where the numbers go.