UNIT 016 h allocated
Model use and limits
- Pretraining versus inference
- prompting
- sampling
- model selection and cost
← Course catalogue
LANGUAGE & GENERATIVE AI / CORE
AI320Generative AI, RAG & bounded agentic workflows
24 h productive allocation9 h 36 m learning + 14 h 24 m training4 units · 1 assessed practical brief
Detailed course plan and training brief. A complete taught course pack is in preparation. The linked foundation module, where shown, is a smaller available resource.
BEFORE YOU BEGIN
A clear starting point.
Prerequisites: NLP310
Core studio / CPU-friendly local or notebook environment
WHAT YOU SHOULD BE ABLE TO DEMONSTRATE
- Explain the assumptions and failure modes in model use and limits.
- Complete the practical: build a small source-grounded assistant for an approved public corpus and a bounded tool workflow with explicit refusal and approval cases.
- Defend the resulting corpus and licence record with reproducible evidence.
THE SUBJECTS / FOUR CONNECTED UNITS
What you will study.
UNIT 026 h allocated
Retrieval-augmented generation
- Chunking
- indexes
- retrieval ranking
- evidence and abstention
UNIT 036 h allocated
Agentic systems
- Tool contracts
- state and memory
- approval boundaries
- bounded execution
UNIT 046 h allocated
Evaluation before demonstration
- Task datasets
- source support
- prompt injection
- tracing and failure handling
THE TRAINING STUDIO
Make the subject
visible in your work.
Build a small source-grounded assistant for an approved public corpus and a bounded tool workflow with explicit refusal and approval cases.
Evidence to produce
- Corpus and licence record
- retrieval benchmark
- evaluation report
- trace of a failed and a successful task
Training brief for supervised delivery. Code, datasets, answer keys and environment acceptance must be completed in the full course pack before delivery is confirmed.
ASSESS THE REASONING
A question you
should be able to answer.
Why does connecting a retriever or tool not automatically make an answer trustworthy?
Review dimensions
Technical correctness · reproducibility · interpretation of results · failure analysis · individual explanation. Course-level rubric and faculty moderation form part of the final adopted teaching pack.
External reference access, hardware, cloud accounts and paid services are separate. No third-party certification is included by implication.