A RAG serving system built from scratch.
Ask a question, find the passages that answer it, and let the model reply from those sources instead of guessing. This is the system that does it, end to end, built from first principles and held together under a real latency budget.
The four hops, every question, under one budget
text → vector nearest passages build the prompt grounded answer Turn the question into a vector, find the passages closest in meaning, paste them into the prompt, and let the model answer from them. Rerank and a cache make it better; the budget keeps it fast.
Learn, the course
23 lessons across 7 units, with interactive demos. Embeddings, retrieval, grounding, reranking, and the latency budget, the ideas you need to build the serving system yourself.
Start learningDocs, run & serve
Installation, the API, and how to run the service against your own documents. Arriving with the first release.
Read the docsWhat makes it different
Answers from sources
The model stops guessing from memory and answers from passages you retrieved, with citations. Grounding is the whole point.
Taught, not just shipped
A full interactive course ships alongside the code. It teaches the concepts; you write the serving system, which is where understanding sticks.
Measured, not just wired
v1 is the honest pipeline. v2 is the headline: where do the milliseconds and dollars go, and does reranking earn its cost? A real latency and cost budget, not a demo.