Part 1 : Generative Classifier and Extractor

⏱ ~30 min📖 Theory🎯 Intermediate

🎯 Learning objectives

  • Explain what generative AI means in the context of UiPath Document Understanding
  • Configure and use the Generative Classifier for description-driven document classification
  • Configure and use the Generative Extractor activity with prompt-based field extraction
  • Apply good practices for writing clear, accurate extraction prompts

In UiPath Document Understanding, generative refers to the use of Large Language Models (LLMs) and Generative AI (like GPT-4) to process, classify, and extract data from documents. Instead of rigid templates or extensive training, it uses plain-language prompts to understand unstructured information — making it possible to handle a wide variety of document types with minimal setup.

🧠 Generative Classifier

The Generative Classifier is a classification tool that leverages generative AI to categorize documents using description-driven document type definitions. Instead of training on labeled examples, it classifies documents based on natural language descriptions you provide for each document type.

⚙️ How to Configure the Generative Classifier

  1. 📥 Insert a Classify Document Scope activity into your workflow
  2. 🔗 Place the Generative Classifier activity within that scope
  3. ✍️ Click Manage Prompts and enter clear natural language descriptions for each document type

🎯 When to Use the Generative Classifier

  • 📄 Documents are unstructured with varied content
  • 💬 You can describe the document types in natural language in a way that allows generative AI to distinguish between them
💡 Tip
Success depends on writing descriptions clear enough for the AI to differentiate between classification categories. If two document types have very similar descriptions, the classifier may struggle — make the distinctions explicit.

⚡ Generative Extractor

The Generative Extractor activity enables data extraction from documents using generative machine learning models through UiPath's Document Understanding platform. It uses a prompt-based wizard where you define field names and descriptions to guide the model.

✨ Key Capabilities

  • 🔍 Extracts document data using generative models
  • 📑 Supports multiple document types with customizable extractor types
  • 🗂️ Three extractor types available:
    • 📃 Long Document Simple Layout Extractor — for long documents with a straightforward layout
    • 📰 Long Document Complex Layout Extractor (preview) — optimized for long-form documents with complex layouts
    • 📄 Short Document Complex Layout Extractor (preview) — for short documents with complex layouts (maximum 10 pages)
  • 🎛️ Field descriptions directly influence model extraction accuracy (max 1,000 characters per description)
  • 🖥️ Compatible with Windows Legacy and Windows platforms
Generative Extractor Prompt dialog showing Extractor Details and Document Types

Generative Extractor Prompt — Extractor Details and Document Types configuration in UiPath Studio

📌 Note
The Generative Extractor does not support table field extraction defined in Taxonomy Manager. For complex table extraction, use the IXP Generative Extraction capability instead.

🏆 Generative Extractor — Best Practices

⚠️ Key Limits to Keep in Mind

  • 🔢 Maximum 50 prompts per extractor
  • 📏 Extraction result (Completion) has a maximum word limit of 700 words
PracticeGuidance
🎯 Use Precise Language Use clear, unambiguous phrasing. Imagine multiple people answering your question — if responses would differ significantly, refine the language.
📋 Standardize Output Format Specify the desired output structure explicitly — e.g. "return date in yyyy-mm-dd format" or "return numbers in ##,###.## format".
📝 List Expected Options When answers are limited to known choices, list them explicitly — e.g. "Possible answers: Married, Unmarried, Separated, Divorced, Widowed, Other."
🔢 Use a Step-by-Step Approach Break complex queries into sequential steps. Programming-style formatting (JSON or XML) improves accuracy by leveraging the model's code-interpretation capabilities.
🚫 Avoid Arithmetic & Logic Operations Do not ask the extractor to perform sums, multiplication, subtraction, or comparisons — handle these in your workflow with standard activities instead.
🗃️ Avoid Complex Table Extraction The Generative Extractor struggles with two-dimensional table data. Use the IXP capability for complex table extraction.
🔎 Assess Confidence Carefully Rather than relying solely on confidence scores, pose the same question multiple ways. Convergent answers indicate reliability; divergent responses suggest human review is needed.

📝 Recap

  • The Generative Classifier categorizes documents from natural-language descriptions of each document type — no labeled training examples needed.
  • The Generative Extractor uses a prompt-based wizard with field names and descriptions; three extractor types are available (Long Document Simple/Complex Layout, Short Document Complex Layout), and it does not support table field extraction.
  • Key limits: 50 prompts maximum per extractor and a 700-word completion limit per field.
  • Prompting best practices: use precise language, standardize the output format, list expected options, break complex queries into steps, avoid arithmetic/logic operations, avoid complex tables, and assess confidence by rephrasing the question.

🧠 Knowledge check

1. What is the key difference between the Generative Classifier and a traditional (ML-model) classifier?

2. Best practice: if you're unsure whether a Generative Extractor field result is reliable, what should you do?