Adding document extractors

Choosing a tool option for extracting fields from documents

You can choose from two tool options to extract fields from documents: Structured and Unstructured. These tools differ in how they process information, but the steps that you need to follow in both options are almost the same.

• Unstructured: Uses a text-based language model to extract top-level fields from documents. It works best for text-heavy content, but does not extract column values from tables.

• Structured: Uses a vision-based, multimodal language model to extract both top-level fields and table data. Because it relies on a larger model, it usually runs slower. Structured extraction is designed to identify and return these types of values:

  • A single, atomic value per field from a document. Typical examples include values such as an invoice number, form date, or total amount.
  • Structured tuples of values with a fixed structure that are grouped within the document. Typical examples include line items in an invoice, or items in a receipt.

Configuring a document extractor after selecting the Unstructured option

To configure the document extractor to identify and extract a set of fields from unstructured documents:

1. Select the document extractor that you added to the agentic workflow.

2. Click the Edit fields icon .

   A dialog to upload the documents and add fields is displayed.

3. Select a model to use for the document extractor from the Model list.

   From the Models list, click View all foundation models to open the model selection dialog, which lists all available models. To select a model, search for it or choose one from the list. After you select a model, click Save. Any notices associated with the selected model, such as deprecation notices or third‑party license requirements, are displayed.

   Certain models include a status tag in the dialog to indicate states such as Recommended or Third party. A warning icon indicates that a model might be withdrawn or deprecated in a subsequent release.

   Note:

   You can upload documents or add fields in any order. Extraction begins when the system has at least one document and one field.

4. Do the following:

   • Upload your documents

     The uploaded sample documents help in creating the fields. These documents do not train the model, and are also not a part of the agent that is being configured.

     It might take some time to upload the documents.

     When the upload is complete, you can see a document preview where you can do the following actions:

     • Evaluate document extraction accuracy. For more information, see Evaluating and improving document extraction.

     • Select an uploaded document to view from the list of documents

     • Add more documents and delete documents. You can click Add documents to upload more documents. You can also delete documents from the document manager. To delete a single document, click the Delete icon when you hover over the document name. To delete multiple documents, click Select all, then select the checkboxes next to the documents to delete, and click Delete documents.

     • Browse the pages of a document that you are viewing

     • Fit the page width and height to view

     • Zoom in and zoom out a page to view

     • Search for a field in a document

     

     In the previous example, a user uploads sample documents and the document preview is displayed for these documents.

   • Add fields

     Click Add field to add fields for the information that you want to extract from the documents such as dates, names, and others.

     The document extractor searches for values that are related to the fields across documents.

     

     In the previous example, a user has added certain fields, which are searched and shown on the document preview panel.

     To edit the field details, click the Options icon and select Edit. You can edit the field name, description, and data type. You can add examples and options for the field to help the model understand what information you want to extract. You can also delete the field if it is no longer needed.

5. Configure the settings that trigger a user review. For more information, see Configuring a user review.

6. After you verify and get the expected results, close the document extractor dialog.

Configuring a document extractor after selecting the Structured option

The ability to extract all line items from structured documents such as purchase orders and invoices reduces manual efforts and improves accuracy in document processing.

To configure the document extractor to identify and extract a set of fields from structured documents:

1. Select the document extractor that you added to the agentic workflow.

2. Click the Edit fields icon .

   A dialog to upload the documents and add fields is displayed.
3. Select or specify a model to use for the document extractor from the Models list.

   From the Models list, click View all foundation models to open the model selection dialog, which lists all available models. To select a model, search for it or choose one from the list. After you select a model, click Save. Any notices associated with the selected model, such as deprecation notices or third‑party license requirements, are displayed.

   Certain models include a status tag in the dialog to indicate states such as Recommended or Third party. A warning icon indicates that a model might be withdrawn or deprecated in a subsequent release.

   Note:

   You can upload documents or add fields in any order. Extraction begins when the system has at least one document and one field.

4. Do the following:
   • Use a schema from a list of predefined schemas

     You can choose a document type and add multiple fields that are associated with the document type, instead of manually adding the fields one at a time. Several document types such as bank statements, invoices, insurance claims, and others are available for selecting schemas.

     To add fields from a predefined schema, do the following:

     1. Click Define schema.
     2. Select a type of document from Predefined schemas.

        A description about the selected document type and all the associated fields that are available to extract data are displayed.

     3. Click Create.

        The document extractor searches for values that are related to the fields across documents.

        In the previous example, the fields that are extracted for an invoice document are shown in the dialog.

        To edit the field details, click the Options icon and select Edit. You can edit the field name, description, and data type. You can add examples and options for the field to help the model understand what information you want to extract. You can also delete the field if it is no longer needed.

   • Upload your documents

     The uploaded sample documents help in creating the fields. These documents do not train the model, and are also not a part of the agent that is being configured.

     It might take some time to upload the documents.

     When the upload is complete, you can see a document preview where you can do the following actions:

     • Evaluate document extraction accuracy. For more information, see Evaluating and improving document extraction.

     • Select an uploaded document to view from the list

     • Add more documents and delete documents. You can click Add documents to upload more documents. You can also delete documents from the document manager. To delete a single document, click the Delete icon when you hover over the document name. To delete multiple documents, click Select all, then select the checkboxes next to the documents to delete, and click Delete documents.

     • Browse the pages of a document that you are viewing

     • Fit the page width and height to view

     • Zoom in and zoom out a page to view

     • Search for a field in a document

       

       In the previous example, a user uploads sample invoice documents, and the document preview is displayed. The preview also shows the automatically detected tables in the document from which you can extract data.

   • Add automatically detected tables

   To add an automatically detected table from the document to extract data, click the Add table icon .

   The data from the table is added to the fields for extraction. To edit the table name, hover your mouse over the table name, and click the Edit icon . To delete the table, you can click the Delete icon .

   You can also reorder the columns in the extracted tables. For more information, see Reordering columns.

   • Add custom tables

     Do the following to add a custom table:

     a. Click Add table and specify a name for your custom table, and press Enter.

     b. Click Add column and provide the column name for your custom table, and press Enter. Repeat this step to add more columns to your custom table.

     The system searches for the column values in the document. The first value that is found for a column is displayed and all other values for that column are highlighted in the document preview. If a value is not found for a column, no value is displayed for the column.

     You can also reorder the columns in the extracted tables. For more information, see Reordering columns.

     c. To view your custom table with a preview of the extracted data, click View full table next to the table name.

   • Add fields

     To add more fields for the information that you want to extract from the documents such as dates, names, and others, click Add field.

     After you enter a field name, the document extractor searches for values that are related to the fields across documents.

     To edit the field details, click the Options icon and select Edit. You can edit the field name, description, and data type. You can add examples and options for the field to help the model understand what information you want to extract. You can also delete the field if it is no longer needed.

5. Configure the settings that trigger a user review. For more information, see Configuring a user review.

6. After you verify and get the expected results, close the document extractor dialog.

Mapping data to inputs

By default, auto-mapping is enabled. However, you can map values to the inputs.

To map values to inputs, complete the following steps:

1. Select the document extractor node and then click Edit data mapping.

2. Specify the input values for data mapping. For more information about data mapping, see Mapping data.

Mapping tables extracted from a document extractor node to other nodes in a workflow

You can now map a table type output from a document extractor node in a workflow. You can do any of the following:

• To display the entire table in a chat, map the extracted table in the workflow by using the list interaction type under a document extractor node.

• To display row data from a table in a chat, you can iterate through the rows of an extracted table using a For each loop in the workflow. To do this, you can combine a For each loop with a user activity node and Message output. Then, to display each row, you can select the columns in the For each loop data mapping.

Also, to use any extracted table output in a downstream node in the workflow, you can use this same process of For each loop data mapping. This enables further processing such as applying conditional logic in a workflow.

Configuring a user review

You can configure the settings that trigger a user review.

If the system extraction confidence score is lesser than the set extraction confidence threshold (for unstructured option), or when the extraction returns empty values (for structured option), a user review task is created and assigned to the user.

To configure a user review:

1. Select the document extractor in the agentic workflow.

2. Set the User review switch to on.

3. Do one of the following:

   • For unstructured option: Click the Edit icon in the If the extraction confidence is below field. Set the extraction confidence threshold for all fields or specific fields that can trigger a user review. To select specific fields that trigger a user review, click the Edit icon , select the specific fields, and click Done. Extraction confidence score reflects how closely the extracted results match the expected values.

   • For structured option: Click the Edit icon in the If the extraction returns any empty values field.

     1. In the Assign to list, select one of these options:

The system notifies the user in the chat that a review is required depending on the extraction confidence score or empty values.

The user can click the chat notification to view the review activity.

The user can then examine the extracted values, make any necessary changes, and submit the review. After submission, the workflow continues with the updated values.

Evaluating and improving document extraction

Evaluation workflow

Step 1: Upload test documents

Step 2: Define extraction fields

Step 3: Review and verify the ground truth

For each document:

1. Select the document from the document manager.
2. Review the extracted values displayed in the field editor.
3. Compare each extracted value with the source document displayed in the document viewer.
4. Correct any inaccurate values in the Correct Value column. Leave the field blank if the correct value should be empty or not available in the document.
5. Click Verify document after you have reviewed all fields.

After verification:

• The document receives a Verified status.
• An accuracy percentage appears next to the document name.
• The document is included in evaluation metrics.
• The Metrics Summary is automatically updated.

A verified document indicates that you have reviewed and confirmed all the ground truth values.

Characteristics include:

• An accuracy percentage is displayed beside the document name.
• A Verified badge is displayed.
• The Verify document button is disabled.
• The document contributes to the metrics.

The percentage displayed beside a verified document is a performance measurement indicating how accurately the model extracted values from that document.

By providing verified ground truth data, evaluating model predictions against that data, analyzing metrics, and iteratively refining the extractor, you can systematically improve the extraction accuracy over time.

Document extractor limits and restrictions

Document extractors have the following limits and restrictions.

The following limitations also apply to the document extractor:

• Confidence values for extracted fields

The structured document extractor does not provide confidence scores for extracted field values, unlike the unstructured document extractor. Therefore, confidence-based user review rules cannot be configured.

• Single-mention extraction

The document extractor captures only one instance of a field value in a document, usually the first occurrence. For example, if an invoice number appears multiple times, only one instance is extracted. To capture multiple values, define fields with distinct semantic labels, such as home phone number and mobile phone number.

Creating agentic workflows

An agentic workflow defines a set of linked activities and controls that are designed to achieve a specific business purpose or goal. Learn more about agentic workflows.