CCARF Practice Question

Paper 1 · Question 28

D4 Prompt Engineering & Structured Output · Paper 1

An extraction pipeline processes 50-page legal contracts in a single API call. It captures parties and dates from the first few pages but misses key financial terms in later pages, and occasionally contradicts itself across fields. A colleague suggests simply increasing the context window. What is the correct diagnosis and fix?

• A.The context window is too small; upgrading to a larger context model will fix the inconsistency
• B.This is attention dilution; split into per-section extraction passes plus a cross-section reconciliation pass.
• C.Run the full extraction three times and take the most common value for each field
• D.Add more detailed instructions about the importance of financial terms to increase the model's focus on those sections

Show answer & explanations

Correct answer: B

A: If the document fits in the context window, a larger window does not improve attention quality. The problem is attention dilution over a long document, not context capacity.

B: Long documents cause attention dilution: the model attends thoroughly to early content but loses focus on later sections. Per-section passes ensure each part gets dedicated attention, and the integration pass catches contradictions and reconciles fields across sections.

C: Multiple passes on the full document suffer from the same attention dilution pattern. The model will repeatedly extract early-page data well and miss later-page data.

D: Instructions cannot override the attention dilution effect of processing a 50-page document in one pass. The structural fix is breaking the extraction into focused per-section passes.

Related Prompt Engineering & Structured Output questions

More: Paper 1 · All practice questions