Skip to main content

Getting Started with Agentic Table Extractor

Prerequisites​

Before using Agentic Table Extractor, ensure you have:

  1. Two LLM adapters configured in your Unstract workspace:

    • An advanced LLM — set as the default LLM in your project's LLM Profile. This powers Agents 0, 2, 3, and 4. Use a capable model such as GPT-4o, Claude Sonnet, or Gemini Pro.
    • A lite LLM — configured in the prompt's settings modal. This powers Agent 1 (page screening). A smaller, faster model works well here (e.g., GPT-4o-mini, Claude Haiku).

  2. An X2Text adapter configured — set in your project's LLM Profile. This is the text extraction adapter used to process uploaded documents.

  3. A Prompt Studio project — either an existing project or a new one.

Step-by-Step Setup​

1. Open or Create a Project​

Open an existing Prompt Studio project or create a new one. Make sure the project has an LLM Profile with both an LLM adapter (advanced) and an X2Text adapter configured.

2. Upload a Document​

Upload the document you want to extract tables from using the document manager. Supported formats include PDF, Excel (.xlsx/.xls), and images.

3. Create a New Prompt​

Add a new prompt card in your project. In the Enforce Type dropdown, select agentic_table.

4. Write Your Prompt​

Write a prompt describing the table you want to extract. The prompt has three components — a required target table description, a required JSON structure defining the output schema, and optional instructions. See Writing Prompts for details.

The prompt card displays a readiness checklist with three checkboxes that validate your prompt as you type. All three should be checked before running extraction.

5. Configure Settings​

Click the gear icon on the prompt card to open the settings modal. Configure the lite LLM adapter and adjust page range and parallelism settings as needed. See Configuring Settings for details.

6. Run Extraction​

Click the Run button on the prompt card. The pipeline will process through all six agents and display the extracted table data in the output area.

The output is a JSON array of objects, where each object represents a table row with column headers as keys.