Step-by-step answer
Stretch 60 min7 stepsCompare long-document strategies
Builds on the verified-extraction project and compares two strategies on an 80-page synthetic lease with two near-identical rent clauses. One is a single sonnet call over the whole document. The other has a haiku locator name candidate pages, runs one sonnet call per field group over only those pages, and merges the results by written rules. Both are graded by the same verifier and normalisers. The dry run shows that locate-then-extract uses more tokens but costs less.
Back to the exercise