> ## Documentation Index
> Fetch the complete documentation index at: https://pathlit.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# VLM Document Extraction Node

> Extract structured data from PDF documents using vision language models

## Overview

The VLM Document Extraction Node uses vision language models to extract content from PDF documents. This node enables you to:

* Extract text, tables, and structured data from PDFs
* Process multiple PDF pages using configurable chunking (pages per chunk and page overlap)
* Customize extraction prompts for specific data formats
* Handle large documents by tuning chunk size and overlap
* Support various vision-capable models

<Note>
  This node is best for **deterministic data extraction** (invoices, forms, structured fields). For more consistent results, use **Temperature 0** in Advanced Settings.
</Note>

## Configuration Parameters

### Node Configuration

* **Upload PDF**:

  Upload one or more PDF documents to extract content from

* **Extraction Prompt**:

  Describe what to extract and the desired output format

  > *Extract all text content from this document. Format the output as markdown with proper headings and sections.*

  > *Extract the table data from this invoice and return it as CSV format*

  > *Find all dates, amounts, and vendor names in this receipt*

<AccordionGroup>
  <Accordion title="Advanced Settings">
    <ul>
      <li>
        <strong>Vision Model</strong>: Select the vision language model to use (e.g., GPT-4o, Gemini Pro Vision)
      </li>

      <li>
        <strong>Temperature</strong>: Controls randomness in extraction (0.0-2.0). **Recommended: 0** for deterministic, repeatable extraction (invoices, forms, structured data).
      </li>

      <li>
        <strong>Max Tokens</strong>: Maximum output tokens per chunk (1000-100000)
      </li>

      <li>
        <strong>Pages per Chunk</strong>: Number of pages to process together in each chunk (1-20). Chunking is fixed by this value—there is no automatic or semantic chunking.
      </li>

      <li>
        <strong>Overlap</strong>: Number of pages that overlap between consecutive chunks (0-10). Use to preserve context across chunk boundaries (e.g. tables or paragraphs that span pages).
      </li>
    </ul>
  </Accordion>
</AccordionGroup>

## Expected Inputs and Outputs

* **Inputs**:

  * The node accepts PDF file uploads from the file upload input

* **Outputs**:

  * **content**: Extracted content in the requested format

## Use Case Examples

1. **Invoice Processing**:

   Extract structured data from invoice PDFs including vendor names, amounts, dates, and line items.

2. **Document Digitization**:

   Convert scanned documents or images in PDFs to searchable, structured markdown or JSON.

3. **Form Data Extraction**:

   Extract filled form data from PDF forms for database entry or further processing.

## Error Handling and Troubleshooting

* **Large Documents**:

  For very large PDFs, adjust **Pages per Chunk** and **Overlap** to balance context (e.g. overlap 1–2 for continuity) with processing speed and token usage.

* **Poor Image Quality**:

  Low-quality scans may produce poor extraction results. Ensure PDF pages are readable before processing.

* **Model Selection**:

  Different vision models have varying capabilities. Experiment with models if extraction quality is unsatisfactory.

If you encounter any issues not covered in this documentation, please reach out to our [support team](mailto:dev@pathlit.ai) for assistance.

## Relevant Nodes

<CardGroup>
  <Card title="Document Reader Node" href="/docs/nodes/document-reader">
    Read text from various document formats
  </Card>

  <Card title="Raw LLM Node" href="/docs/nodes/raw-llm">
    Direct LLM interface with file uploads
  </Card>

  <Card title="LLM Node" href="/docs/nodes/llm">
    Process extracted content with LLMs
  </Card>
</CardGroup>