Custom Extractor

Create Processor

Navigate to Processors → Custom Processors and click Create processor on the Custom Extractor card. Enter a descriptive name (e.g., “ACME Invoice Extractor”) in the slide panel and click Create.

Define Fields

Open your extractor and navigate to the Field Management tab. This is where you define what data to extract.

Click Add Field to configure each extraction target:

Fields support parent/child hierarchies for nested extraction (e.g., a “Line Items” parent with “Description”, “Quantity”, “Price” children).

Document Prompt: Optionally add a document-level prompt that provides context for the AI about what kind of document it’s processing and what to look for.

Generate Schema with AI

For faster setup, use the Schema Generator panel:

1. Click Generate Schema in the field management toolbar
2. Upload up to 5 representative documents (drag-and-drop or S3 import)

3. AI analyzes documents and suggests extraction fields
4. Review generated fields — each shows name, display name, data type, occurrence, and description
5. Select which fields to import using checkboxes
6. Preview the JSON schema of selected fields
7. Click Apply to create all selected fields at once

The Schema Generator is particularly useful when starting with a new document type — it bootstraps your field definitions from real documents.

Import Training Documents

Navigate to the Documents tab and click Import Documents. Upload your training documents:

• Supported formats: PDF, PNG, JPEG, TIFF
• Minimum: 10 examples recommended for initial training
• Best practice: Include diverse document variations (different vendors, layouts, edge cases)

Documents can be uploaded via file picker, drag-and-drop, or imported from S3 storage.

Label Examples

Click Start Annotating or select a document to open the labeling interface.

The labeling interface provides a two-panel layout:

Left panel — Document canvas:

• Multi-page document viewer with zoom (0.25x to 3x)
• Three drawing modes: Draw, Select, Move
• Draw bounding boxes around field values
• Page navigation with previous/next buttons

Right panel — Annotation panel:

• Field list with assigned colors and annotation counts
• Select a field, then draw a box on the document to annotate
• Keyboard shortcuts: 1-9 to quickly select fields
• Delete annotations with Delete/Backspace key

AI-assisted labeling:

• Auto-annotate — AI suggests field values from the document
• Auto-layout — AI positions annotations on the page
• View raw AI extraction results for comparison

Workflow per document:

1. Select a field from the annotation panel (or press 1-9)
2. Draw a bounding box around the field value on the document
3. Repeat for all fields on all pages
4. Click Complete & Next to save and move to the next document
5. Skip documents with no extractable data

Train Model

Navigate to Training Jobs and click Start Training. See Training for detailed configuration.

Evaluate Results

After training, the Evaluation dashboard shows:

Overall metrics:

• Accuracy, Precision, Recall, F1 Score (as percentages)

Per-field breakdown table:

Color-coded badges indicate metric strength: green (≥90%), yellow (≥70%), red (<70%).

Training progress:

• Epoch-by-epoch metrics: Train Loss, Validation Loss, Train Accuracy, Validation Accuracy, Learning Rate
• Use to assess convergence and detect overfitting

Deploy

Activate the trained version for production use. See Training — Production Deployment.