Extract
Category: Processing
The Extract activity pulls structured field values out of each document according to the fields defined for its class. This is where the transaction's data is produced.
Configuration
The fields, their descriptions and the guidelines belong to the project: every Extract step reads them. A step can extract only some of the fields, or send its own text for a field, for a document class's guidelines or for the general guidelines, shown as Custom in this step with Reset to project text. Every other setting, Memory's Use examples in this step included, belongs to this step alone, so two Extract steps can use different models.
| Setting | Description | Default |
|---|---|---|
| Fields | Fields to extract from all document types. | (project fields) |
| Per-class Fields | Fields to extract only for specific document classes. | (optional) |
| Guidelines | Additional extraction instructions for the model. | (optional) |
| Auto-discover Fields | Let the model find and extract additional fields. | Off |
| Include Reasoning | Ask the model to explain each extracted value. | Off |
| Include Coordinates | Return bounding-box coordinates for each value. | On |
| Send Images | Send page images to the model. | On |
| Send Text | Send OCR text to the model. | On |
| Send Images of Pages Without Text | When only text is sent, still send the image of a page that has no text (a blank page, a photo). Not available for text-only models. | Off |
With Auto-discover Fields on, the first document of a class that has no fields yet is extracted on its own. The other documents of that class wait for it and reuse the fields it found, instead of each naming the same data differently ("Invoice Date" and "Inv Date"). Documents of classes that already have fields, and documents of different classes, are extracted in parallel.
Pages and documents without text
A page with no text (a blank page, a photo) is a normal page, not an error.
- Some pages of a document have text: extraction works as usual. Each page without text is marked to the model as "(no text on this page)".
- No page has text, and no image is sent: there is nothing to extract from, so the model is not called. The document's fields stay empty, the step carries on, and the Activity Log names the document. Validation rules that check those fields then report them as usual.
- A page that never went through OCR, while the step sends text and no image of that page, fails the step at once with the page numbers. Add an OCR step before Extract, or send page images.
Re-running a single document from the transaction viewer follows the same rules.
Document Examples
The Extract step's configure panel also holds the Document Examples section, which controls the project's Memory: whether reviewed transactions are captured as samples, and how those samples are matched and injected into extraction prompts. See Memory for the full list of options.
Notes
- Bounding-box coordinates power click-to-locate highlighting in the review interface.
- Fields are defined per document class in project settings; see Projects.