map.extract

Extract fields

Reads each row's text, images or PDFs and fills in columns you describe, such as a case number, a date or the parties. You say what each column should hold; a model finds it in the source.

Find it under Analyze › Extract › Extract fields

Analyze ribbon with Extract fields highlighted

Settings

Source: The column, or columns, the model reads. Usually the text from an import.
Engine: The model. Start with a cheap one and test it on a few rows.
Columns it creates: The columns it creates. Give each a name, a type and a description of exactly what to copy and how.
Instruction: Rules for every field, such as “Copy names exactly” or “Leave a field blank if the document doesn't say”.
Include Confidence: Adds a column with the model's own estimate, from 0 to 1, of how sure it is about the first field. A rough guide to which values to check first, not a measure of accuracy.
Citations: Citations. Cite sources records the passage each value came from. Require citations leaves a value blank when it can't cite one. PDFs sent directly can't be cited.
Citation sources: Under Advanced, when citations are on: columns to check citations against. Leave it empty to use the action's own source columns.
Context: Under Advanced: optional background about the dataset, such as where the documents come from, to help the model read them.

Writes

One column per field you describe.
With Include Confidence, a confidence score for the first field.
With citations on, the passage behind each value, shown in Review and the row's details.

Limits

The model can copy a value wrongly or make one up. Check a sample against the source before relying on a column.
Sending PDFs directly works only with OpenAI and Anthropic models that read PDFs, and can't be cited.
Each run is charged by the model's provider unless the model runs with Frisket.