NCP-GENL Prompt Engineering 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.

• ✓

  Instruct the model to output fields in a fixed JSON schema and to use null for any field not present in the document.

  Why this is correct

  A fixed schema removes ambiguity about output shape and key names, and the null convention gives the model a defined behavior for missing data instead of inventing values. Together these constraints make outputs parseable and reduce hallucinated field values, which is essential for downstream automation.

• ✗

  Ask the model to summarize the invoice in prose first and then extract fields from its own summary in a second pass.

  Why it's wrong here

  Summarization loses exact values and can introduce paraphrase errors before extraction even begins. The second pass then works from degraded information, compounding inaccuracy. Direct extraction with few-shot guidance and a fixed schema is more reliable than routing through a lossy intermediate representation.

• ✗

  Raise the temperature to 0.9 so the model explores multiple interpretations of ambiguous invoice text before committing to values.

  Why it's wrong here

  Higher temperature increases variability in generated tokens, which is harmful for deterministic extraction tasks. Field values may change between runs, and hallucinated numbers become more likely. Extraction benefits from low-variance decoding combined with strong structural instructions, not from creative sampling.

• ✓

  Provide few-shot examples that cover several distinct invoice layouts, each showing the exact input-to-output field mapping.

  Why this is correct

  Few-shot examples with layout diversity teach the model to generalize extraction patterns rather than memorize one format. Showing the precise field mapping in the completion clarifies expected keys and value styles, which directly reduces variance when new vendor layouts appear. This is a core technique for improving structured extraction consistency.

• ✗

  Remove all formatting and concatenate the entire invoice text into a single lowercase string before sending it to the model.

  Why it's wrong here

  Stripping layout and case destroys spatial and typographic cues that distinguish labels from values, such as table alignment and header capitalization. This makes extraction harder, not easier. Preserving document structure, or converting it to a structured representation like Markdown, generally improves accuracy.