Retrieval augmented generation has a well-known shape: it demos beautifully and then underwhelms the moment real users start asking real questions. Jason Liu's course exists to deal with that second phase, and it is the most useful material on the topic I have seen. The core idea is what Liu calls the RAG flywheel. Instead of tweaking prompts and chunk sizes by feel, you build synthetic evaluation sets with LLM-generated questions and answers, measure retrieval separately from generation using precision, recall and MRR, segment your incoming queries to find which kinds are failing and how often, and then fix the segments that matter most.
The later material covers fine-tuning embeddings, collecting much more user feedback than most products bother with, building specialised indexes for documents, tables and images, combining BM25 with embeddings and metadata filters, and routing queries to whichever retriever handles them best. None of this is exotic. What is valuable is the order and the discipline, which is exactly what most teams lack. Liu has the background to teach it.
He spent years building search and recommendation systems at Facebook and Stitch Fix, has consulted for many AI startups on retrieval, and wrote Instructor, the structured output library that a lot of Python developers use. Students on the page include engineers from OpenAI, Anthropic, Google and Meta, and an Anthropic solutions engineer is quoted saying the frameworks turned what they already knew into practice. That is the right audience and the right claim. This course will not teach you what RAG is.
The prerequisites say outright that you should already have a RAG system deployed. The format is six prerecorded lectures, at least six live office hours where people bring their own problems, weekly guest speakers, a set of Python notebooks with walkthroughs, and a Slack community you keep for life. Free re-enrolment in later cohorts is a genuinely generous policy and means one payment can cover you as the material evolves, which it clearly does, given that the page mentions 12 notebooks in one place and more than 20 in another. My main frustration is the price, or rather its absence.
Between cohorts the page shows nothing, so you cannot build a budget case until enrolment opens. The other thing to know is that Liu publishes a lot of this thinking for free in a long written guide on his own site. If you are disciplined, that guide plus your own logs will take you a long way. The course earns its fee through the office hours and the guest speakers, and you get the most from it if you arrive with a struggling production system and specific questions.