COURSE PREVIEW / FORMAT & CURRICULUM
From documents
to grounded answers.
Build a small retrieval system, evaluate its evidence, and develop a capstone you can explain. This is a written, self-paced course with practical Python labs.
Inside the course
- Design the context contract
Define what an assistant may know, what it must prove, and how to fit useful evidence into a finite context window.
- Build a retrieval baseline
Turn source material into traceable retrieval units, compare lexical and semantic approaches, and combine ranked lists without confusing scores.
- Evaluate before you optimize
Measure retrieval coverage and answer support separately, create honest development and holdout sets, and use controlled experiments to diagnose improvement.
- Connect evidence across sources
Plan multi-hop questions, inspect explicit graph relationships, retain provenance, and diagnose the earliest point where useful evidence is lost.
- From evidence to a useful answer
Build answers with traceable claims, define bounded tool workflows, and preserve useful memory without carrying every past token.
- Reliable context pipelines
Design predictable failure behavior, version your knowledge pipeline, and collect the signals needed to explain a bad answer.
- Security, economics, and release control
Keep access boundaries outside the model, test hostile retrieved content, and plan a measured deployment with a clear rollback path.
- Capstone: build, evaluate, and decide
Assemble an internal knowledge or customer-support assistant, review the evidence, and make a justified release recommendation.
How you will learn
Each module includes three written lessons, two labs, and a six-question knowledge check. Complete the capstone using the provided rubric. Plan roughly 40–60 hours with practice; basic Python and API familiarity are recommended.
Course access follows your account. Notes and progress are stored in your browser and can be exported. Optional model providers and compute have separate costs.