Why AI Models Return Incomplete Arrays and How to Fix Structured Output Bugs

A common structured-output bug causes AI models to return fewer items than expected — such as seven out of ten invoice line items — without any error being reported. The problem most often stems from token limit truncation, incomplete input chunks, or silently dropped schema constraints like minItems, rather than a genuine model failure. Developers are advised to log finish_reason and token usage on every extraction call to quickly rule out the first two causes. Schema design also plays a role: placing a total_items count field before the array can cause the model to commit to a number it then fails to match, while placing it after provides a free consistency check. The most reliable fix involves numbering input rows and requiring the model to reference each line number in its output, converting a vague recall task into a series of concrete per-item decisions.