For Instructors

For instructors

What you get

• A 51-slide deck (separate from this repository)
• This repository as the worked implementation
• Twelve weeks of structure: three tiers, three parallel advanced tracks
• Seven labs, a capstone brief, and a 20/20/20/40 rubric
• An evaluation harness students can point at their own corpus

All MIT licensed. No permission is required to adopt, edit, refactor, translate or teach from any of it. You do not need to ask or notify anyone.

Adapting it

Replace data/ with documents from your own institution and rewrite eval/cases.json against them. That single change makes the whole course local, and it takes an afternoon. Students engage far more with their own handbook than with a sample corpus.

Nigerian examples are used throughout — naira amounts, bursary circulars, course codes like CSC 508 and DSA 708. Swap them freely.

The lab machine problem

This is why the offline providers exist. Thirty students on a lab network with no budget for API keys can all run the full pipeline, including evaluation, with zero cost and zero network. Introduce real providers in week 6, once the concepts are settled and only one key is needed for demonstration.

Sequencing advice

1. Run it before explaining it. Week 1, session 1: clone, index, ask a question. Understanding follows.
2. Return to the seven-step map constantly. Students who hold the whole shape debug far better than students who hold one step.
3. Teach --hits in week 1 and insist on it in every bug report. It breaks the habit of blaming the prompt.
4. Make them break grounding on purpose (lab 7). Seeing a confident wrong answer once inoculates against trusting one later.
5. Reward negative results. A correctly measured change that made things worse teaches more than an unmeasured one that helped.

Assessment notes

The failure analysis in the capstone carries more weight than the demonstration. A system presented as flawless is marked as one that was not examined.

Be explicit that AI assistance is permitted and expected, must be declared, and that students will be asked to explain any line they submit.

Common student misconceptions

They think	Actually
Bigger context window means retrieval matters less	Retrieval decides what goes in the window
The model hallucinated	Retrieval failed and the prompt allowed it
More retrieved chunks is better	Six good chunks beat twenty mediocre ones
Lowercasing and stop-word removal help	Those harm embeddings; they are keyword-search habits
Fine-tuning would fix this	Fine-tuning changes behaviour, not knowledge

Getting help

Open an issue. Questions from instructors adapting the material are welcome and usually improve the documentation for everyone.