Recipes / Chain
From documents to checked data
Pull facts out of a set of documents, check them, and only then use the table. The order matters: each step catches mistakes the one before can't.
- 1Extract a first sample
Follow the court filings recipe: describe the fields, preview them on three filings, and save a run on those three. Start small, so a badly worded field costs you three rows, not three hundred.
- 2Check known answers
Score the sample against expected.csv, the answers filled in by hand. If a field misses, rewrite its description, or compare another model on the same three rows, until it scores well.
- 3Review against the documents
Run the settings on every filing, then open the run in Review. Go through a random sample, checking each value against its cited passage. Correct what's wrong.
- 4Download the checked table
Export the checked sheet as CSV or Excel to share or publish. Keep the project, so you can trace any value back to its document.
Use the same order with your own documents.