CCDV-F - Model Selection and Optimization (16.8% of the exam) - Section 5.2
Apply the foundational technical concepts that support AI application development.
Published skill weight 6.1 percent. Covers the basic engineering practices underneath a Claude integration, including working with SDKs that wrap REST APIs and with websockets, so that a candidate can reason about the transport rather than only about the model.
SDKs wrapping REST APIswebsocketsfoundational engineering practice
Practice question for this objective
lock_openFree sampleModel Selection and Optimizationeasy
A two-person team must ship a Claude-backed summarisation service in three weeks. They are deciding between calling the Messages API over plain HTTP with a hand-written client and using the official SDK for their language. Their stated constraints are the three-week deadline and a requirement that request and response handling be covered by the same maintenance as the rest of their dependencies. Select TWO sound reasons to take the SDK.
AThe SDK wraps the same REST endpoints and already implements request construction, response parsing and retry handling that the team would otherwise write and test itself.check_circle Correct
BThe SDK removes the need to design the prompt or decide which model to call, because those choices are made inside the client library on the team's behalf.
CThe SDK is versioned and installed through the team's existing package manager, so upgrades and security fixes follow the dependency process they already run.check_circle Correct
DThe SDK opens a persistent connection to the service for the lifetime of the process, which a client written against plain HTTP cannot be made to do.
EThe SDK guarantees that responses never need validating in the application, because everything it returns has already been checked against the team's own schema.
An official SDK is a maintained wrapper over the published REST API, supplying client plumbing rather than application decisions. The SDK and a hand-written client speak to the same REST endpoints. What the SDK adds is maintained plumbing, namely request building, response parsing, retry and error handling, delivered through the normal dependency channel. It does not take over prompt design, model selection or validation of generated content, which stay with the application.
Why A is correct: Correct: an SDK is a maintained client over the published REST surface, so the boilerplate the deadline does not allow for is already written and exercised by other users.
Why B is wrong: Tempting because the SDK does supply sensible defaults, but prompt design and model choice stay application decisions the team must make and own.
Why C is correct: Correct: this satisfies the stated maintenance requirement directly, since the client is tracked and upgraded like any other pinned dependency.
Why D is wrong: Tempting because connection reuse is a real performance concern, but any competent HTTP client can pool connections, so this is not an SDK-only capability.
Why E is wrong: Tempting because the SDK does give typed responses, but it knows nothing of the team's schema, so validating generated content remains the application's job.
Answers that look right on this material and are not. Each one is a distractor from a different question in the CCDV-F bank for this domain.
Keep the largest model and hold spend down by reducing the output limit on each request, since a tag is only a few tokens long and nothing longer is ever needed.
errorWhy it is wrong: Tempting because the output really is tiny, so a low limit looks free. It is wrong because a short output was never the cost driver, so the ceiling is unaffected, and any reply needing a little more room is cut off instead.
Instruct the assistant to check its arithmetic carefully before answering and to flag any total it is not fully confident in for a human reviewer to confirm.
errorWhy it is wrong: Tempting because a self-check reads like verification, but the model cannot audit its own generated total against the source figures, so a wrong total is asserted just as confidently as a right one.
Send each test input three times and assert that the three replies are identical to one another, treating a disagreement between them as the regression signal for that test case.
errorWhy it is wrong: Tempting because it acknowledges run to run variation, but identical wording is not what the test should require, so the suite still fails on harmless rewording and passes consistent wrong answers.