Extract Document Data

Prerequisites

Input options

Supported languages for generative models

Models used by the activity

The Extract Document Data activity uses the following:

• Public endpoints for out-of-the-box models.
• Custom ML models deployed in Document Understanding App projects.
• Generative extraction model.

Designer panel

• Project - Requires you to select your Document Understanding project from the drop-down menu. The available options are:
  • Predefined - The default project
  • You can create a custom project by going to Document Understanding.

• Extractor - Requires you to select the Extractor from the selected project. For the Predefined Project, the available options are:
  • Either one of the ML Packages found here
    Note: The Extract Document Data activity overrides the document type with the selected extractor. This is not applicable for generative models.
  • Generative
    Note: The data sent to the Generative Extractor will be sent to an LLM Model instance which is not publicly available, will not leave it, and once processed, it will not be stored or used for training.
• Prompt - this field appears if you choose the option Generative. Prompt to identify the fields to be extracted, provided as key-value pairs, where the key represents the name of the field and the value a description for it, helping the extractor identify the corresponding value. Click on the field, and you will get a prompt with the following options, provided as pairs:
  • Field name - Requires you to input the field name to be extracted (Ex. Due date) (30-character limit)
  • Generative prompt - Requires you to provide the prompt as input for the Generative Extractor. The maximum number of characters allowed is 1000. The response, extraction result, also called Completion, has a word limit of 700. This is limited to 700 words. This means that you can't extract more than 700 words from a single prompt. If your extraction requirements exceed this limit, you can divide the document into multiple pages, process them individually, and then merge the results afterwards.
  Tip: For good practices on how to use generative prompts, check the Generative extractor - Good practices page.
• Input - Requires you to specify the file itself, or Document Data, in case you have used other Document Understanding Activities before in your workflow, (for example, Classify Document).
  Important: The maximum numbers of pages a file can have is 500. Files exceeding this limit fail to extract.

Properties panel

• Timeout (seconds) - Maximum execution time (in seconds) for the call to the generative model. If the operation exceeds this timeout, it is automatically terminated to prevent delays or hangs. This property is only displayed if the Generative Extractor is selected as an extractor.
• Auto-validation - Use this option to enable automatic validation, a capability that helps validate the results obtained for data extraction against a Generative model. The default value for the Auto-validation field is False.
  • Confidence threshold - This field becomes visible once you enable Auto-validation. Extraction results falling below the threshold are compared to the generative extraction model. If they match, the system adjusts the extraction confidence to meet the threshold value. Possible threshold values range from 0 to 100.

    If the value is set to 0, no validation is applied. However, if you set a specific value (from 0 to 100), the system checks all extraction results below this value. For example, if you set a confidence threshold of 80%, the system will apply the generative validation for fields with confidence below 80%.

    Note: Auto-validation is available only for specialized extraction models.
• Generate Data Type (Preview) - If set to True, indicates that the output should be generated based on the selected extractor, resulting in an IDocumentData<ExtractorType> object. Alternatively, if set to False, indicates that the data generation should be skipped, resulting in a generic IDocumentData<DictionaryData> object.

  Visit Document Data for additional details and limitations available for the two object types.

Output

• Document Data - All the extracted field data from the file. Information can also be received from Classify Document.

  Visit Document data to learn how Document Data works and how to consume the extracted results for single and multi-value fields.