CCA-F Practice Questions — Page 17

Question 161

Scenario: Structured Data Extraction

You have several extraction tools (one per document type) and must guarantee the model returns a tool call producing structured output rather than a free-text answer, but you are fine with the model choosing which extraction tool fits.

Which `tool_choice` setting fits?

A.`{"type": "auto"}`
B.`{"type": "any"}`
C.`{"type": "tool", "name": "extract_invoice"}`
D.Omit `tool_choice` entirely.

Question 162

Open question ↗

Scenario: Structured Data Extraction

An extraction pipeline must always run `extract_metadata` as the very first tool call, before any enrichment step, with no exceptions allowed.

How do you guarantee that specific first action?

A.Use `tool_choice: "any"` and rely on the model to pick it first.
B.List `extract_metadata` first in the `tools` array.
C.Use `tool_choice: "auto"` with a strong system-prompt hint.
D.Set `tool_choice: {"type": "tool", "name": "extract_metadata"}` on the first request.

Question 163

Open question ↗

Scenario: Structured Data Extraction

An extraction returns a date in the wrong format and fails validation. You decide to retry with feedback to fix it efficiently.

What should the retry prompt include to be effective?

A.The original document, the incorrect extraction, and the specific validation errors to correct.
B.Only a terse instruction to "try again and be more careful."
C.A larger `max_tokens` budget but the same original prompt.
D.A request to raise the temperature so the model varies its output.

Question 164

Open question ↗

Scenario: Structured Data Extraction

A pre-merge CI check must extract and validate the changed configuration files on every pull request, blocking the merge until results return — typically within seconds. A teammate suggests routing it through the Batch API to cut costs.

Which API choice is correct here?

A.The Batch API, because it is always cheaper than synchronous calls.
B.The Batch API with a 1-hour submission cadence to approximate real time.
C.The synchronous API, because this is a blocking, latency-sensitive check.
D.Either one; the API choice has no effect on a pre-merge gate.

Question 165

Open question ↗

Scenario: Structured Data Extraction

A CI extraction reviewer flags many findings as "possible performance issue," and the team wants to analyze which findings are false positives so they can tighten that category's criteria over time.

Which schema addition best supports this analysis?

A.Add a numeric severity score from 1–100 to every finding.
B.Include a `detected_pattern` field recording exactly what triggered each finding, so false positives can be analyzed by pattern.
C.Add the model's free-text reasoning paragraph to each finding.
D.Add a `timestamp` field to each finding.

Question 166

Open question ↗

Scenario: Conversational AI Architecture Patterns -

A multi-turn assistant is told in turn 1 to "always respond in formal English and never reveal internal IDs." By turn 12 it begins using casual phrasing and once echoes an internal record ID. Investigation shows each request sends only the latest user message to the API.

What is the root cause and fix?

A.The model's context window is too small; switch to a larger-context model.
B.The temperature is too high; lower it so instructions persist.
C.Re-send the instruction as a user message every few turns to remind the model.
D.The full conversation history (with the system instruction) must be sent on every request; the API is stateless per call.

Question 167

Open question ↗

Scenario: Conversational AI Architecture Patterns -

A travel-booking assistant must follow a behavior rule — "never confirm a booking until payment is verified" — across every turn of a long conversation. The team currently embeds this rule only in the first user message of each session.

Where should this standing rule live?

A.In the first user message only, since the model remembers earlier turns.
B.Repeated inside each tool's result payload so the model sees it often.
C.In the system prompt, which defines persistent behavior across all turns of the conversation.
D.In a one-time assistant message the model writes to itself at the start.

Question 168

Open question ↗

Scenario: Conversational AI Architecture Patterns -

A user says "Cancel my dinner reservation for Friday," then two turns later says "Actually, keep Friday but move the Saturday one to 8pm." The assistant must resolve the conflicting intent without acting on the wrong booking.

What is the best way to handle this conflicting input?

A.Act on the earliest instruction, since it was stated first and is most deliberate.
B.Confirm the current intent with the user before acting — restate what will change (keep Friday, move Saturday to 8pm).
C.Average the two requests by canceling Friday and moving it to 8pm.
D.Silently apply both instructions in the order received.

Question 169

Open question ↗

Scenario: Conversational AI Architecture Patterns -

An assistant aggregates a long report from many sources into one large block before answering. Reviewers notice it reliably uses facts placed at the top and bottom but frequently overlooks key findings buried in the middle, even though they are present.

Which phenomenon is this, and how do you mitigate it?

A.Token starvation; simply raise `max_tokens` so the middle fits.
B.Schema drift; enforce a stricter JSON Schema on the report.
C.Tool overload; reduce the number of tools the assistant has.
D.The lost-in-the-middle effect; place key findings at the start and action items at the end of the aggregated input.

Question 170

Open question ↗

Scenario: Conversational AI Architecture Patterns -

A conversational ops assistant exposes a `delete_environment` tool. Policy says it must never run without an explicit human confirmation step, but on one occasion the model invoked it directly from an ambiguous user phrase.

What is the most reliable way to make this tool safe?

A.Add a strongly worded system-prompt warning to always confirm before deleting.
B.Lower the temperature so the model is less likely to call the tool by mistake.
C.Add a `PreToolUse` hook/precondition that intercepts `delete_environment` and requires an explicit confirmation before it can execute.
D.Rename the tool to `danger_delete_environment` so the model treats it cautiously.