infinity-parser2

name

infinity-parser2

description

Parse PDFs, scanned documents, and document images (screenshots, invoices, paper documents, whiteboard photos) into structured layout JSON (labeled regions with bounding boxes) and Markdown via Infinity-Parser2 through an existing vLLM server. Use when the user wants to extract text, tables, formulas, or structured data from visual documents; mentions OCR, text recognition, or document parsing; or asks to parse, digitize, or extract content from a document. Requires a vLLM OpenAI-compatible API URL and API key.

Infinity-Parser2 Skill

Prerequisite: pip install infinity_parser2

Use only the vllm-server backend. The vLLM server must already be running. The CLI reads the server URL and API key directly from the INFINITY_PARSER2_* process environment variables.

CLI

The parser command is available after installation. Set the environment variables first:

export INFINITY_PARSER2_API_URL=<full /v1/chat/completions URL>
export INFINITY_PARSER2_API_KEY=<api-key>

parser /path/to/document.pdf \
  --backend vllm-server \
  --output-dir /path/to/output \
  --output-format md,json \
  --pages 5-7 \
  --model-name inf-mllm

Key flags:

• -o/--output-dir <dir> — write results to disk (<dir>/<basename>/result.md|json). Omit it and results print to stdout, which floods the model's context — prefer writing to disk.
• --output-format md|json|md,json — md,json keeps both files; default md.
• --pages 1-3,5 — 1-based physical PDF pages (same syntax as the Python API).
• --backend vllm-server — required; the default backend is vllm-engine.
• Overrides if the env vars are not set: --api-url <full /v1/chat/completions URL> --api-key <key>.

Python API

import os

from infinity_parser2 import InfinityParser2

parser = InfinityParser2(
    model_name="inf-mllm",
    backend="vllm-server",
    api_url=os.environ["INFINITY_PARSER2_API_URL"],
    api_key=os.environ["INFINITY_PARSER2_API_KEY"],
)
parser.parse(
    "/path/to/document.pdf",
    output_dir="/path/to/output",
    output_format="md,json",
)
print("saved to /path/to/output")

Tasks

• Default (Markdown + layout JSON): parser.parse(path, output_dir="/path/to/output", output_format="md,json").
• Write only Markdown: parser.parse(path, output_dir="/path/to/output", output_format="md").
• Write only layout JSON: parser.parse(path, output_dir="/path/to/output", output_format="json").
• Convert directly to Markdown: parser.parse(path, task_type="doc2md", output_dir="/path/to/output").
• Parse several paths: parser.parse([path1, path2], output_dir="/path/to/output").

output_format="json" or "md,json" is only valid when task_type="doc2json" (the default). doc2md and custom tasks only support output_format="md"; passing json/md,json with them raises a ValueError.

Accept PDF files and supported document images. For a directory, pass its path directly to parser.parse().When both result.md and result.json are needed, explicitly set output_format="md,json". Each input gets its own subdirectory: <output_dir>/<input-basename>/result.md and result.json.

PDF Pages

Pass pages to select physical PDF pages with 1-based numbering:

parser.parse("/path/to/document.pdf", pages="1-3,5,8-10", output_dir="/path/to/output")
parser.parse("/path/to/document.pdf", pages=[1, 3, 5], output_dir="/path/to/output")

Omit pages to parse every page. The selection applies to each PDF in a multi-file call and is ignored for images.

Before running, verify that INFINITY_PARSER2_API_URL is exported in the process environment. If the endpoint requires authentication, also verify INFINITY_PARSER2_API_KEY. Never request, print, or expose API keys.