Domain 2 of 6

Engineering the Context

The work that begins once one prompt stops being enough. What changed when the field renamed itself, the context window as a budget with competing claims on it, the four things a team can do with that budget, the ways a long context degrades rather than fails, and retrieval as the main way relevant material gets in.

Concepts

~18%

Of the exam

Practice questions

Concepts in this domain

01Context engineeringThe rename that happened in June 2025 and what it changed, the seven kinds of material that reach a model on one call, one request shown as a bare prompt and then as an assembled context, and where prompt engineering sits inside the larger job.02The context windowThe window as a budget with seven claims on it, one support assistant's budget worked twice six months apart, the room an answer needs reserved before anything else is assembled, and why a longer window moves the problem instead of removing it.03Write, select, compress and isolateThe four things a team can do with a context budget, what each one costs and what each one risks, a compression prompt written out in full, and the ordering rule that says selection comes before compression.04Context rot and long context failureWhat the measurements show about accuracy falling as input length grows, the four named ways a context goes wrong, the move from the previous page that answers each one, and why a degraded context returns a plausible answer instead of an error.05Grounding and retrievalWhat grounding means, the three kinds of data a system can ground against, the three arrangements that reach them, a grounded prompt written out in full, why retrieval reduces invention without removing it, and the sampling settings that shape what comes back.06Retrieval Augmented GenerationThe pipeline stage by stage, why chunk boundaries decide retrieval quality, what an embedding similarity score measures, hybrid search and reranking, and why retrieval has to be measured on its own before any prompt is tuned.

Try a question from this domain

A pet supplies retailer measures one order question end to end. The assembled context comes to 1,262 tokens, of which the system instruction is 180. The rest is tool definitions, two policy clauses kept from eight retrieved, a customer memory record, three turns of history with the earlier ones summarised, one tool result and the customer's question. The team's improvement plan is a week of rewriting the system instruction. What does the breakdown say about that plan?

AThe instruction owns 180 tokens and the other 1,082 are the output of six decisions taken at run time, so the failures that reach customers sit in retrieval, eviction and tool filtering rather than in the wording.
BThe instruction is the only part the team writes in full, so it is the only part the team can improve and a week spent on it is the right call.
CThe plan is sound because the instruction sits at the start of the window, which is one of the two positions research says gets used best.
DThe breakdown says nothing until the same call is measured on a longer window, since the share each part takes changes with the size of the window.

16 questions on this domain.

One per page, with a worked explanation.

Start the set