By this point a plain chat window is no longer the whole story, and the syllabus turns to the architectures built on top of a language model. Retrieval-augmented generation, which grounds answers in a document collection the organisation controls. Agents, which chain steps and call tools instead of returning a single block of text. Fine-tuning, which adjusts the model itself. And the operational layer around all of it — versioning, evaluation and monitoring of prompts and models over time.
It is the smallest chapter on the exam: five questions of 40, five points of 46, and every one of them K2. Nothing here asks you to design or implement any of these solutions. The questions ask you to understand what each approach buys and what it costs, which is a much shorter list to learn than the chapter's subject matter might suggest.
Prepare by being able to choose between the options rather than define them. Stale or missing domain knowledge points at retrieval; a consistently wrong style or format points at fine-tuning or at a better prompt; a task that needs several steps and a real system call points at an agent. Keep the costs attached to each answer too — a retrieval index that is never refreshed serves confident, outdated results, and that trade-off is exactly the kind of thing a K2 question is built around.
This chapter is covered most heavily (5 questions) in CT-GenAI — Mock Exam 1 (Prompting for Test Analysis, Test Data and Pilot Governance), CT-GenAI — Mock Exam 2 (Model Behaviour, Technique-Driven Prompting and Operational Risk), CT-GenAI — Mock Exam 3 (Checkable Output, Retrieval Operations and Governance)