CCDV-F practice questions

Question 1· Select all that apply

Applications and Integration

A service caches a 30,000 token system prompt and reads it back on nearly every call. Two routes were added last month and neither one shows a cache read.

Route B sends the identical system prompt to a smaller model. Route C sends the same prompt with the customer's display name appended to the end of it, so the reply can greet them by name.

What each of the three routes sends ahead of the cache breakpoint

Select the two statements that explain what these routes see.

Select all that apply.

Route B reads nothing because an entry belongs to the model that wrote it.
Route B reads nothing because a smaller model has no cache of its own.
Route C reads nothing because the appended name changes a prefix that has to match exactly.
Route C reads nothing because per customer text may not be cached.
Both read nothing because an entry is written by one route and readable only by that route.

Question 2· Multiple choice

Model Selection and Optimization

Two models are candidates for two routes. One is cheaper per input token and dearer per output token, and the other is the reverse. One route sends a long document and asks for a two line verdict. The other sends a short brief and asks for a page of copy. Both teams pick the same model, calling it the cheaper one.

What each route sends and gets back, and what the two models charge

What decides which model each route should use?

The lower headline price per million tokens, which covers both directions.
That route's own mix of input and output tokens, priced at both rate cards.
The model with the lower output rate, since output is the dearer half.
The total tokens each route moves in a month, whichever direction they run.

Question 3· Select all that apply

Applications and Integration

A service returns a classification result to two internal consumers it does not own. A developer wants to rename a field for clarity and add a new one, and cannot deploy to the consumers.

{ "cat": "billing", "conf": 0.91 }

Select the two changes that are safe for the existing consumers.

Select all that apply.

Add the new field alongside the existing ones.
Rename the field and keep emitting the old name as well.
Rename the field and announce the change in the release notes.
Rename the field and keep the old one populated with null.
Rename the field and version the whole payload as a new shape.

Question 4· Multiple choice

Agents and Workflows

A team is asked to build an agent that decides whether an incoming email is a complaint, a question or a receipt. The decision is made from the email text alone. There is nothing to look up, and nothing happens afterwards except storing the label.

What should they build?

An agent with a loop and one classification tool.
One model call that returns the category.
An orchestrator with one worker for each of the three categories.
A chain that classifies and then verifies the classification.

Question 5· Multiple choice

Applications and Integration

An internal tool sends the same long system prompt with every request and has prompt caching turned on. It is used about four times an hour, spread across the working day. After a month the input bill is higher than it was before caching was enabled.

Why?

The prompt is above the size a cached prefix is allowed to hold.
Cached prefixes expire between calls at this spacing, so each one writes again.
Only the message list is eligible, so a system prompt is never stored.
Reads are priced above writes, so every hit costs more than a miss.

Question 6· Multiple choice

Model Selection and Optimization

A support assistant answers about 90 percent of questions well on the small model. The remaining tenth needs the large one, and sending everything there costs more than the team wants to spend. They plan to try the small model first and escalate.

Traffic entering a small model, with a check that sends some requests on to the large one

What does this design need to work?

A second account, so the two models are billed separately.
A signal, available per request, that decides which path it takes.
A shared cache, so both models read the same prompt prefix.
A queue, so escalated questions are handled after the easy ones.

Question 7· Multiple choice

Applications and Integration

A product bills customers per conversation from the token counts on each reply. The figures were right until the chat moved to streaming. Since then every streamed reply is recorded at zero output tokens, while the account's own invoice still shows them.

The events one streamed reply produces, and where the client stops reading

What is the client doing wrong?

It takes the counts from the opening event, which carries the input side only.
It takes the counts from the response headers, which a streamed response does not carry.
Streamed replies report no counts at all, so they have to be estimated from the text.
It stops reading at the last text delta, and the output count arrives after that.

Question 8· Multiple choice

Model Selection and Optimization

A support router sorts every incoming message into one of six queues. Two thirds of the traffic is automated mail, delivery receipts and out of office replies, each carrying one of a dozen fixed subject lines. All of it goes to the model, which classifies it correctly every time.

What is the cheapest change that keeps the accuracy?

Move the whole route to the smallest model in the family.
Match the fixed subject lines in code and call the model on what is left.
Classify the fixed subject lines overnight rather than as they arrive.
Shorten the classification prompt that every one of these calls carries.

Question 9· Multiple choice

Applications and Integration

A design review is told the system must handle 20,000 classifications a day. One engineer sizes the integration by dividing that figure by the hours in a day. Another says the number cannot be used that way yet.

What is the second engineer asking for?

The cost of a classification, so the daily total can be priced.
The average length of the text being classified each time.
Confirmation that 20,000 is a daily figure rather than a monthly one.
How the 20,000 arrive across the day, and what the busiest hour holds.

Question 10· Multiple choice

Agents and Workflows

A file cleanup agent pauses before anything is deleted and asks the user to approve. The user reads what the dialog shows and approves. A second later the delete tool runs, and it takes this quarter's exports with the ones the user had in mind.

Two entries from the agent's run log, a second apart

What is wrong with the approval step?

It shows the model's description of the call instead of the call's arguments.
It asks for approval once instead of asking again for each file.
It runs after the model has already decided what to delete.
It lets the model write the text that the user reads.