Introduction

Extract data from documents in UiPath Maestro™ Flow, using IXP models that you build, evaluate, and run in the Extract node.

The Extract node brings document processing onto the UiPath Maestro™ Flow canvas as a step in your flow. It pulls structured fields out of a document using an IXP extraction model, and returns the result for downstream nodes to act on. Because the flow calls the node at a fixed point with the inputs you configure, extraction behaves deterministically.

This guide covers the full lifecycle of the model behind the Extract node:

• Building the model: uploading sample documents, defining the taxonomy, configuring the model, and validating predictions.
• Evaluating model performance before you deploy the model to production.
• Deploying and managing models.
• Consuming models via Flow, via automations, or via the Document Understanding API.

Warning:

The Validate extraction node works only with Extract nodes created in the same flow, with the Extract document option. It does not yet work with Extract nodes bound to a published model or to a model from another flow. For details, check Consuming models via Flow.

Warning:

Only the user who generated the predictions that a trace belongs to can view that trace. For details, check Maintenance and monitoring.

Data extraction challenges

Depending on the stage you are at, some of the common challenges when extracting data from documents are:

During solution design

• Difficult to build a document processing solution from scratch that integrates well with the rest of the automation in an organization.
• Difficult to create a structured extraction schema. Custom solutions usually have no user interface to build and visualize the taxonomy.

During model building and testing

• Access required to different LLMs to test on various document types, which are use cases.
• Difficult to quickly iterate on a user prompt and view prediction results.
• Access required to more parameters to optimize prediction results.

During evaluation

• Difficult to validate predictions and provide ground truth.
• Time-consuming to assess model and field-level metrics.
• Required to compare the impact of adjusting model parameters, for example, model, user prompts, chunk size, zero-shot, or few-shot.
• Required to build a visualization of performance metrics, for example, precision, recall, F1 score.

During production deployment

• No easy way to modify and maintain data schema for a use case.
• Little model governance, for example, the ability to revert to a previous version, or to select a particular version for production.

Document processing in Flow

The Extract node in Flow helps you solve the challenges outlined in the previous section, and integrates document processing with the rest of your flow.

Use the Extract node for document processing in general, from simple, documents to advanced scenarios where data is not consistently formatted, for example, when:

• Documents contain paragraphs of free-form text or complex elements, such as:
  • Complex tables
  • Graphics
  • Charts
  • Checkboxes
  • Call-out boxes
  • Signatures
  • Handwriting, and so on.
• You need to extract inferred values, which are information that is not stated directly but must be derived from context.
• There is high variation in layout or structure between documents or within fields to be extracted.
• You are dealing with stacks of multiple document types combined as one file and need to extract data without splitting them up first.

Benefits

These are some of the benefits of using the Extract node in Flow:

• The Extract node runs alongside agents, automations, and human tasks in the same flow.
• You can build a model directly in the node or reuse one deployed from IXP.
• Complex data, such as tables and long documents, is better supported.
• You control prompts, LLMs, and model settings.
• Evaluation runs are versioned.
• Models and taxonomies deploy together with the solution, from development to production, without being re-created by hand.
• Every extraction run is recorded in the flow trace, for end-to-end debugging and monitoring.
• Every LLM request passes through the AI Trust Layer and is audited.
• Solution packages are versioned, so you can upgrade or roll back, and fields keep a change history.
• Models built in Flow can be consumed from public APIs and activities, without changing your automations.