Flexible Extraction

Flexible Extraction is a Supervision task in which a keyer reviews, corrects, or enters field values and table cells for a document. Hyperscience creates the task when a human must provide or verify data before the document can continue through its processing flow.

Use Flexible Extraction for a manually matched page, data selected for review by flow logic, or output from a specialized ORCA VLM that requires human review. After the keyer completes the task, Hyperscience saves the reviewed values in the standard document output and continues processing the submission.

Benefits of Flexible Extraction

Flexible Extraction provides a controlled path when extraction results need human verification before they can be accepted. It helps teams:

• Control automation risk

  • Thresholding can send specialized model results that do not meet the target accuracy to a human for verification.

• Keep exceptions in the processing flow

  • Keyers resolve them in Hyperscience instead of using a separate manual process.

• Focus review

  • The task can show only the fields or table cells selected for attention. Filters can further limit fields to manually matched pages or the selected layout variation.

• Preserve consistent output

  • Hyperscience saves reviewed values in the same document output used for automatically extracted values.

• Reduce rekeying

  • When enabled, machine predictions can prefill values, and text segment selection can copy visible document text into a field or table cell.

Best-fit scenarios

Use Flexible Extraction when a document can continue through the flow after a human supplies or verifies specific extraction results.

Flexible Extraction is not a fit when processing must remain machine-only, when a document contains more than 1 table, or when the keyer must select a different layout inside the task. See Requirements and limitations.

Flexible Extraction stages

A document moves through Flexible Extraction in 4 stages:

1. Selection

   • Classification, a Code Block, or the decision layer after ORCA VLM Extraction determines that the document or specific fields  require human review.

2. Routing

   • The Is Flexible Extraction needed? step sends the document through the Flexible Extraction Block. Hyperscience creates a task for the document.

3. Review

   • A keyer confirms, corrects marked values, or enters missing values

4. Output

   • Hyperscience saves the reviewed values and continues the submission through the flow.

   • In v43.2 and later, if address parsing is turned on, Hyperscience splits address transcriptions into consistent components, such as house number, street, city, state, and postal code. For configuration settings and supported providers, see the Address Parsing section in Document Processing Flow.

Specialized model output control

Training creates a specialized model on top of the ORCA base model. Hyperscience uses the adapted and base models together. The specialized model changes how the base model behaves for the use case; it does not replace or modify the base model. The specialized model produces an extracted value and one confidence score for each field. To learn more, see Training a Specialized Model.

ORCA does not flag a field, ask for clarification, or decide that it needs human help. After extraction, thresholding compares the confidence score with the Threshold for the configured Target Accuracy. The decision layer accepts results that meet the threshold for automation and sends results below the threshold to Flexible Extraction.

A required field that the model predicts is missing also goes to Flexible Extraction. The same applies when the field's transcription Supervision mode is Always or Consensus. Fields set to Auto-transcribe do not go to Supervision.

Together, thresholding and Flexible Extraction provide the operational control layer for a specialized model:

• thresholding determines how much automation risk to accept

• Flexible Extraction lets a human verify or correct the remaining results.

• Flexible Extraction does not improve model accuracy or change model behavior; training does.

Machine-only processing

A submission sent through the API with the machine_only parameter does not create Supervision tasks, including Flexible Extraction tasks.

Required permissions

Completing a Flexible Extraction task requires the following permissions. The flow's Default Task Restrictions setting can further limit which users receive its Flexible Extraction tasks.

Prerequisites

Complete the prerequisites for the way that you plan to use Flexible Extraction.

Use Flexible Extraction as a standard Supervision task

Confirm the following:

Use Flexible Extraction with a specialized ORCA VLM

Complete the standard prerequisites and confirm the following:

• Required: The instance has a supported GPU and an installed ORCA base model permitted by the current license.

• Required: A Semi-structured layout contains the fields that the specialized model must extract. The latest layout version is locked and is not linked to another ORCA VLM model definition. Do not rely on table definitions because ORCA VLMs do not extract table data.

• Required: An ORCA VLM model definition is linked to the layout. One layout can be linked to only 1 model definition.

• Required: The model definition has enough eligible, annotated training documents to meet the training requirement shown for the selected ORCA base model.

• Required: An adapter has been trained against the selected ORCA base model, and the resulting specialized model has been deployed as the live model.

• Required: Thresholds have been calculated for the specialized model. Recalculate them after training a new adapter or switching adapters because the model's confidence distribution can change.

• Required: In the ORCA Extraction settings, select the correct base model, turn on Use Live Model, and set ORCA Target Accuracy. Running the flow with Use Live Model turned on fails when no live model is available for the corresponding model definition.

• Required: The flow routes the output of ORCA Extraction through Is Flexible Extraction needed? to the Flexible Extraction Block.

• Recommended: Validate the candidate specialized model and its projected automation before deploying it.

• Recommended: Turn on Show Machine Predictions in Supervision Tasks so keyers can review the specialized model's predicted values.

Open a Flexible Extraction task

1. Go to Tasks > Perform Tasks.

2. In the Supervision Tasks table, click Perform Tasks in the Flexible Extraction row.

3. Review the task introduction, and then click Continue.

The task opens with the Document pages panel, the page view, and the Transcribe Fields sidebar. Press F2 to open the keyboard-shortcuts panel.

• Document pages lists the pages in the order set during Classification. An icon identifies a manually matched page. Select a page to view it.

• The page view shows the selected page. Use its controls to move between pages, rotate the page, or change the zoom. When text-segment search is enabled, the Search segment text field also appears.

• The Transcribe Fields sidebar contains the document details, fields, and table.

Review and transcribe fields

A warning icon marks each field that you must review. The icon's message is “Field must be reviewed to complete the task.” When available, the following filters appear and are turned on by default:

• Show fields from manually matched pages only appears when the document contains a manually matched page.

• Show fields from selected variation only limits the list to fields from the selected layout variation.

When both filters are on, the list includes only fields that meet both conditions. Turn off a filter to show more fields.

Data types

Transcribe exactly what you see. When Flexible Extraction Transcription Masking is turned on, which is the default, the task does not accept characters that are invalid for the field's data type. If a value fails validation, press F9 to override the validation. The field then shows Formatting Overridden. For supported characters, see Default Data Types.

1. Select a field. If the field has a location, the page view scrolls to and highlights it.

2. Compare the field value with the document.

3. If the value is correct, press Tab to move to the next field.

4. If the value is missing or incorrect, enter the value as it appears in the document. For a checkbox or signature field, mark the field as filled or not filled.

5. If the document does not contain a value, leave the reviewed field blank.

6. (Optional) If you cannot read the value, click Mark Illegible. The field shows Marked Illegible.

To move to the next field that requires review, press Ctrl + E on Windows or Command + E on Mac.

Add or remove a field occurrence

When the field supports multiple occurrences:

1. Click Add occurrence in the field, or click Add another [field name] on the field's bounding box.

2. To delete an occurrence, click Remove occurrence.

Clear a field value and location

Click Clear location on the bounding box. Hyperscience removes the field value and its location.

Transcribe table cells

If the document has a table, the sidebar lists its name and columns. A warning icon with the message “Column cells must be reviewed” marks a column that contains cells you must review. Table cells follow the same data-type rules as fields.

Deleting all rows cannot be undone

Delete All Rows removes every row from the table. In the “Delete all rows?” dialog, click Delete all rows only when you want to transcribe the table again from the beginning.

1. Select a column name. The table editor opens with the first cell in the column selected.

2. Press Enter to edit the cell.

3. Enter the value.

4. Press Enter to save the value.

5. Press Tab to move to the next cell, or use the arrow keys to move between cells.

To manage rows:

1. If the table has no rows, click Start Table.

2. If the table has rows, right-click a row.

3. Select Insert Row Above, Insert Row Below, Delete Row, or Delete All Rows.

Search for text segments

Text-segment search availability

Text-segment search and selection are available in v43.2 and later. A flow builder must turn on Enable Text Segment Search; the setting is turned off by default.

When Enable Text Segment Search is turned on, use Search segment text to find text on every page of the document. Search is not case-sensitive.

1. Enter all or part of the text in Search segment text.

2. Press Enter. Matching segments are highlighted, and the page view moves to the first match.

3. Click Previous Segment or Next Segment to move between matches.

The counter shows the current match and total number of matches, such as 2 of 5. If the search has no matches, it shows No results. Press Shift + Enter to move to the previous match or Esc to clear the search. To focus the search field, press Ctrl + Alt + F on Windows or Command + Option + F on Mac.

Populate a value from text segments

1. Select the field or table cell that you want to populate.

2. Click one text segment, or draw a box around several text segments.

Hyperscience adds the selected text and its location to the field or cell. Selecting another segment replaces the current value. Transcription masking applies the field's or column's data-type rules. For example, selecting Dunlop2026 for a numeric field adds 2026. You can also select a detected checkbox or signature segment for the corresponding field type.

Mark an incorrect layout

If the document was matched to the wrong layout, you cannot select another layout in the Flexible Extraction task. The document-details section shows the Submission, Layout Group Name, and Layout Variation. A manually matched variation also shows (Manually Matched).

Marking the layout incorrect can end processing

If the flow does not contain an enabled Reprocessing Block, Hyperscience ends processing for the submission and does not extract data from the document.

1. In the Transcribe Fields sidebar, expand the document details.

2. Click Mark Layout Incorrect. For a task created by an older flow, click Mark Layout Variation Incorrect.

3. Review the consequence in the confirmation dialog.

4. Click Continue.

If the flow contains an enabled Reprocessing Block, Hyperscience removes the incorrect layout from all pages in the document and creates a Manual Classification task.

Reject a document

Reject a document when it cannot be processed, for example, when it is not in good order. Rejecting a document dismisses its remaining Supervision tasks.

1. In the Transcribe Fields sidebar, expand the document details.

2. Click Reject Document.

3. In the Confirm Document Rejection dialog, select a reason from Reason for rejection.

4. Click Reject Document.

Hyperscience marks the document as rejected. An administrator configures the available reasons; the default reason is Not in good order.

Complete the task

1. Confirm that you reviewed every field marked for review.

2. Click Complete Task. You can also press Ctrl + Enter on Windows or Command + Enter on Mac.

If a marked field is still unreviewed, the Fields Need Review dialog opens. Click Return to Task, and then review the remaining fields. A reviewed field can remain blank when the document does not contain a value. If a value fails validation, correct it or override the validation before completing the task.

Hyperscience saves the reviewed values and continues the submission through the flow. If address parsing is turned on, Hyperscience parses address-typed fields after the Flexible Extraction task.

Page numbers in the JSON output

For a manually matched page processed through Flexible Extraction, layout_page_number and layout_variation_page_number are 1. Use document_page_number to determine the page's order in the document.

Requirements and limitations

• A Flexible Extraction task supports no more than 1 table in a document. A document with more than 1 table does not receive a Flexible Extraction task.

• You cannot select a different layout in the task. Mark the current layout incorrect instead.

• Text-segment search requires v43.2 or later and the Enable Text Segment Search setting.

• Submissions sent with the machine_only API parameter do not receive Flexible Extraction tasks.

Next step

To review the extracted values, see Submission Output Page.