Part 2 : Generative Extractor in IXP

⏱ ~55 min📖 Theory + Hands-On🎯 Expert

🎯 Learning objectives

  • Explain what IXP Generative Extraction is and which document types it is designed for
  • Identify when to use IXP Generative Extraction vs. Document Understanding or Communications Mining
  • Build a Generative Extraction project: create, upload documents, define taxonomy, validate predictions
  • Evaluate model performance using precision, recall, and F1 metrics
  • Publish and consume a Generative Extraction model in a UiPath Studio workflow
  • Apply prompting best practices at project, field group, and field levels

Theory : Understanding IXP Generative Extraction

What is IXP Generative Extraction?

The Unstructured and Complex Documents capability in UiPath IXP uses Generative Extraction to process documents that are too complex or variable for traditional rule-based or template-based methods. It is an inference-first experience that provides zero-shot and few-shot extraction — meaning it can extract data from documents it has never seen before, guided purely by natural language prompts.

This capability converts unstructured document content into structured, actionable data for automation and analytics — without requiring ML model training or positional rules.

📌 Examples of documents IXP handles
Contracts, mortgage applications, medical records, paystubs, insurance policies, loan applications, and any document with high layout variability or complex elements.
Six types of complex documents IXP can handle: complex tables, charts and graphs, handwriting and signatures, packet documents, inferred values, and checkboxes

IXP Generative Extraction handles a wide range of complex document elements

When to Use IXP Generative Extraction

This capability is ideal for advanced document processing scenarios where data is not consistently formatted. Use it when one or more of the following apply:

Complex Document Elements

Documents contain paragraphs of free-form text or complex elements such as:

  • Complex tables — nested, merged cells, multi-level headers
  • Graphics and Charts — bar charts, pie charts, embedded images with data
  • Checkboxes and Call-out boxes — form-style elements with contextual values
  • Signatures, Stamps, and Handwriting — non-digital content that varies per document

Inferred Values

You need to extract inferred values — information that is not stated directly in the document but must be derived from context. For example, calculating net pay from gross pay and deductions, or determining account type from surrounding text.

High Layout Variation

There is high variation in layout or structure between documents, or within the fields to be extracted. Different employers, vendors, or agencies may use entirely different formats for the same document type.

Packet Documents

You are dealing with stacks of multiple document types combined as one file and need to extract data without splitting them up first. IXP can process these packet documents in a single request.

💡 Zero-shot vs. Few-shot Extraction
Zero-shot extraction uses only your field instructions — no annotated examples needed. Few-shot adds validated annotation examples alongside instructions to improve accuracy on complex or ambiguous fields.

Supported File Formats

PDF PNG JPG / JPEG TIF / TIFF GIF

User Roles in Generative Extraction

Role Permissions Access
IXP Service Admin
Manage service settings, create projects, assign all roles Full Access
IXP Project Admin
Manage specific project, assign project-level roles Project-level
IXP Developer
Build, annotate, train, and publish models Build & Publish
IXP Model Trainer
Annotate documents and validate predictions Annotate
IXP Analyst
View project stats and performance metrics View Stats
IXP Viewer
View-only access to projects and results View Only

Generative Extraction Prompting Best Practices

The quality of your prompts directly determines extraction accuracy. Well-written instructions unlock the full potential of the generative model.

Project-Level Instructions

  • Briefly describe the industry and document type (e.g., "This is a US payroll paystub from the HR/Finance industry")
  • Mention any special characteristics of the document (multi-page, scanned, handwritten sections)
  • If the document may contain multiple instances of the same data, include a unique identifier field in every field group

Field Group-Level Instructions

  • Choose clear, recognizable names — use the same terminology as in the documents
  • Group related data points together logically
  • Explain how the field group contributes to the document's overall meaning
  • Indicate the likely location or section in the document (e.g., "Found in the top-right header section")
  • Treat each field group as a table — each field is a unique column
  • Use > to create parent-child relationships (e.g., Paystub > Deductions)

Field-Level Instructions

  • Start every instruction with "Extract…"
  • Specify the exact format and structure of the expected output (e.g., "Extract as DD/MM/YYYY")
  • Provide example inputs and expected outputs for complex or ambiguous fields
  • Focus on one piece of data per instruction — don't overload a single field
  • Select the appropriate data type: Text, Number, Date, or Boolean
  • For inferred fields, describe the derivation logic (e.g., "Extract net pay — calculated as Gross Pay minus total deductions")

General Prompting Rules

RuleGuidance
Clarity & SimplicityUse clear, direct, unambiguous language. Avoid complex nested sentences.
ConsistencyUse the same terminology across project, field group, and field instructions.
Provide ContextGive the model relevant industry and document-type context at the project level.
IterateWrite → Test → Edit → Save. Repeat until extraction quality is satisfactory.
Avoid Negative InstructionsInstead of "Don't skip sections", say "Ensure all earnings sections are captured".
Avoid RepetitionRepeating instructions across levels causes confusion and reduces accuracy.
Avoid ContradictionsEnsure project, field group, and field instructions do not conflict with each other.
Positive ExamplesInclude examples of correct responses to guide the model (few-shot).
⚠️ Important
Exact vs Inferred Fields: Generative Extraction can extract both exact fields (values explicitly stated in the document) and inferred fields (values derived or calculated from document content). Always be explicit in your field instruction about which type each field is.
📖 Official Documentation
For the full best practices reference, see: Best Practices · IXP User Guide →
📥 Prefer a written copy?
Download this best-practices guide as a handout for future reference:

📝 Recap

  • IXP Generative Extraction provides zero-shot and few-shot extraction for complex, unstructured documents — no ML training or positional rules required.
  • Best suited when documents have complex elements (nested tables, charts, handwriting), inferred values, high layout variation across vendors/agencies, or are packet documents combining multiple document types in one file.
  • Six user roles govern access: Service Admin, Project Admin, Developer, Model Trainer, Analyst, and Viewer — each with a different scope of permissions.
  • Prompting best practices apply at three levels — project (industry/document context), field group (clear names, logical grouping), and field (start with "Extract…", specify format) — plus general rules: be clear and consistent, iterate (write → test → edit), and avoid negative instructions, repetition, or contradictions.

✅ Knowledge check

1. What is the difference between zero-shot and few-shot extraction in Generative Extraction?

2. Which of the following is a field-level prompting best practice in Generative Extraction?

3. How do you consume a published Generative Extraction project in UiPath Studio?