PDF4vllm

Provides a PDF reading MCP server that returns text, tables, and images in an ordered, compact format for vision-enabled LLMs.

python

0

GitHub Stars

python

Language

6 months ago

First Indexed

3 months ago

Catalog Refreshed

Documentation & install

Readme and setup notes from the catalogue, plus a client-ready config you can copy for your MCP host.

Installation

Add the following to your MCP client configuration file.

Configuration

View docs

{
  "mcpServers": {
    "pyjudge-pdf4vllm-mcp": {
      "command": "/path/to/python",
      "args": [
        "/path/to/pdf4vllm-mcp/src/server.py"
      ],
      "env": {
        "PDF_MAX_PAGES": "20",
        "PDF_PAGE_IMAGE_DPI": "150"
      }
    }
  }
}

PDF4vllm MCP is a specialized server that reads PDFs and prepares blocks of text, tables, and images for downstream vision-enabled language models. It automatically detects corrupted text data, preserves reading order, and avoids overloading the model by limiting pages per request while filtering extraneous images. This makes it practical to integrate PDF content into vision-assisted workflows without manual pre-processing.

How to use

You will run the MCP server locally or point your MCP client to a running server. The server is designed to be invoked through standard MCP client patterns and will return structured blocks for each requested page. Use the auto mode to get text and tables when clean, or images when text is too noisy for reliable extraction.

Key usage points:

Choose an extraction mode: auto, text_only, or image_only. In auto, the server starts with text extraction and switches to image blocks if text appears corrupted.
Rely on preserved reading order so text, tables, and images appear in sequence for each page.
Handle page limits to avoid context overflow; typical values cap the number of pages per request.
Benefit from automatic filtering of non-content images like logos, lines, and headers/footers.