CCAO-F Using the Claude API Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Embed the manual into chunks, retrieve the top-matching chunks for each question, and include only those excerpts in the request.

  Why this is correct

  Chunking plus similarity retrieval places the passages most likely to contain the answer into the context window while keeping total tokens bounded, which is the standard retrieval-augmented pattern for documents larger than the context limit. It scales to manuals of any size because only the retrieved excerpts are sent. Accuracy depends on chunk sizing and retrieval quality, but this directly addresses both the size constraint and the grounding requirement.

• ✗

  Truncate the manual to the first N tokens that fit and always send that same prefix with every question.

  Why it's wrong here

  A fixed prefix ignores where the relevant policy actually lives, so questions about later chapters receive no grounding and Claude will answer from parametric knowledge or refuse. Truncation is cheap but degrades accuracy in exactly the cases that matter most for a large manual. It also wastes context on early sections that may be irrelevant to the current question, crowding out genuinely useful text.

• ✗

  Increase the max_tokens parameter on each request so the model can internally hold more of the manual.

  Why it's wrong here

  max_tokens caps how many tokens the model may generate in its reply; it has no effect on how much input the model can attend to. Raising it lengthens possible answers but does nothing about the oversized document, so the request would still fail or omit content. Input capacity is governed by the model's context window, not by the output token ceiling, making this setting irrelevant to the problem.

• ✗

  Summarize the entire manual once with Claude, cache the summary, and send the summary with every question instead of the source text.

  Why it's wrong here

  A whole-manual summary necessarily discards specifics such as exact clause numbers, thresholds, and exceptions, which are precisely what policy questions ask about. Answers generated from a lossy summary can sound authoritative while being subtly wrong, and there is no way to trace a statement back to the source. Summarization compresses size but sacrifices the fidelity that grounded question answering requires.