Tell lrnit what you want
to master.
01 · how lrnit works
One sentence in.
A whole course out.
02 · the mechanism
It builds
while you
watch.
Retrieval, end to end
Chunking and recall
Tool-calling loops
Evals before you ship
Context budgets and latency
One feature behind a guardrail
03 · inside a lesson
One lesson,
four ways in.
Watch
Read
Listen
Build
For the part you need to see happen.
For the part you need to check a source on.
For the commute, offline.
For the part you only believe once you have made it.
Your place is kept across all four.
Drawn explanation · chapters follow the syllabus
Embeddings put the answer near the question, not inside it.
A chunk that splits a definition from its example scores below a paragraph that mentions neither. That one fact explains most of what makes a good corpus retrieve badly[3] and why the boundaries, not the model, are usually what you are debugging[7].
SOURCE 3
Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) · §2.1
“So the fix isn’t a bigger context window. It’s where you cut…”
retriever.py · yours, forty lines in
chunks = split(doc, on=headings)index = embed(chunks)hits = index.search(q, k=4)answer = llm(prompt(q, hits))
The last lesson of chapter one, and the only one with no quiz after it. You either have a retriever that answers, or you don’t.
04 · built this week
Nobody asked for the same thing twice.
1,414
courses built since launch, updated daily
CRASH COURSE
Ship a RAG endpoint before Friday
FULL COURSE
Agents that call tools without looping forever
SEARCH
Why my reranker made recall worse
FULL COURSE
Evals for a chatbot nobody trusts yet
CRASH COURSE
Cut token spend without losing answers
FULL COURSE
Take a prototype notebook to production
SEARCH
Is a vector DB worth it under a million chunks
CRASH COURSE
Prompt injection, and what stops it
FULL COURSE
Earn the fine-tune before you pay for one
SEARCH
Why my agent costs four dollars a conversation