The demo works in an afternoon. Making it right often enough that people stop double-checking it is the actual job.
You ask your AI assistant a question about your own company's handbook and it answers with total confidence, in fluent prose, about a policy that does not exist. Bolting a search step onto a language model, a technique called retrieval-augmented generation or RAG, fixes that in about thirty lines of code, and then breaks in a dozen new ways. This course walks the whole ladder. You'll build the naive version first, watch it fail, and then earn each fix: splitting documents so chunks keep their meaning, adding old-fashioned keyword scoring (BM25) alongside vector similarity and fusing the two, reranking the top fifty down to the top five, rewriting vague user questions before they hit the index, and chaining hops when one passage can't hold the answer. Then the parts that decide whether it survives contact with users: measuring retrieval and answer quality with numbers you trust, forcing every claim to carry a citation, letting an agent decide when to search at all, holding latency and cost steady at a hundred queries per second, and gating releases on evaluations that run in continuous integration. You finish by shipping one over a real corpus, with the metrics to prove it works.
Built by Lakshya Kumar
We grant free access case-by-case — students, career-switchers, builders on a tight budget. Sign in to send us a note.
Sign in to applyFinished the tasks? Take the prompt to your AI and get tested on it. We copy the prompt and open the app — just paste it in.
Semantic search alone misses exact terms and returns near-duplicates; a few cheap extra passes lift answer accuracy 30-50%.
Users say 'cancel my thing'; the doc says 'terminate a subscription' — without a bridge, even good retrieval returns nothing.
Some questions need two documents combined — search in steps, cap the cost, and prove the system reasoned instead of guessing.
Retrieval scores can rise while answers get worse — learn to catch that before your users do, not after.
A confident answer nobody can check is a liability — trace claims to real sources, flag outdated docs, and refuse when unsure.
One search isn't always enough — give the model tools, and it can retry, reformulate, and stop when it actually has the answer.
A demo that feels instant for ten users can crawl — and bill $20K a month — once real traffic arrives. Fix it before it ships.
One prompt tweak can quietly wreck answer quality — gate every change in CI, watch live traffic, and roll back in minutes.
Complete all modules, then submit the required number of capstone projects. Each must earn a passing rating from an admin reviewer.
Pick a real corpus (your company docs, a public dataset like Wikipedia or arXiv, or your own knowledge base). Ship a production-quality RAG with: chunking dispatcher, hybrid + rerank, query rewriting, agentic mode for complex queries, citations, eval gate in CI, live monitoring, and the full production checklist. Submit the live URL + the metrics + the checklist.
Build an eval framework that another team could drop in: golden-set tooling, retrieval + generation metrics, LLM-as-judge with calibration, CI gate, online signal integration, dashboard. Submit the framework as a small npm/pypi package or repo.
Paste this into any AI chat. Fill in the bracketed parts with your context — you'll get back a straight answer on whether this belongs on your plate.
I'm taking a "RAG Systems" course that runs from naive RAG through hybrid + rerank, query rewriting, multi-hop, evals, citations, agentic RAG, scale, and production DevOps. It uses Python (with Node alternatives) and lots of real 2026 production tricks (Anthropic Citations API, prompt caching, Contextual Retrieval, MCP). Here's my context: 1. My current product/project is: [describe] 2. My current RAG state: [haven't built one / naive prototype / shipping / scaling] 3. My corpus: [size, doc types, growth rate] 4. Where I think RAG is failing me: [my guess] Given that, answer: - Which module should I prioritize, and why? - Name 3 concrete wins this course would unlock for my situation. - Name 1 thing the course won't help me with so I don't have wrong expectations. - If I only had 2 hours this week, which single technique gives me the biggest lift? How would I measure that it worked?
Build an agentic RAG that uses ≥3 tools (docs, structured DB, web). Include streaming, observability, retry/reformulate, citations. Submit the live demo + traces.
Scale your RAG to 100 QPS sustained: prompt caching, multi-layer caches, batched ingest, vector topology, per-tenant isolation, cost dashboard. Submit load-test results + cost numbers.
Pick a domain you know nothing about, build a RAG on it in 1 week. Document every wrong turn, every fix, every metric. Submit the writeup — it's the most honest learning artifact in this course.
Module 2 leans on this directly. Pair with the Citations API doc.