vlmparse

A unified wrapper for Vision Language Models (VLM) and OCR solutions to parse PDF documents into Markdown.

Features:

• ⚡ Async/concurrent processing for high throughput
• 🐳 Automatic Docker server management for local models
• 🔄 Unified interface across all VLM/OCR providers
• 📊 Built-in result visualization with Streamlit

Supported Converters:

• Open Source Small VLMs: LightOnOCR-1B-1025, LightOnOCR-2-1B, MinerU2.5-2509-1.2B, HunyuanOCR, PaddleOCR-VL-1.5, granite-docling-258M, olmOCR-2-7B-1025-FP8, dots.ocr, dots.ocr-1.5, dots.mocr, chandra, chandra-ocr-2, DeepSeek-OCR, DeepSeek-OCR-2, Nanonets-OCR2-3B, GLM-OCR, FireRed-OCR, OCRVerse, Qianfan-OCR
• Open Source Generalist VLMs: such as the Qwen family.
• Pipelines: docling
• Proprietary LLMs: gemini, gpt

Installation

Simplest solution with only the cli:

uv tool install vlmparse

If you want to run the granite-docling model or use the streamlit viewing app:

uv tool install vlmparse[docling_core,st_app]

If you prefer cloning the repository and using the local version:

uv sync

With optional dependencies:

uv sync --all-extras

Activate the virtual environment:

source .venv/bin/activate

CLI Usage

Note that you can bypass the previous installation step and just add uvx before each of the commands below.

Convert PDFs

With a general VLM (requires setting your api key as an environment variable):

vlmparse convert "*.pdf" -o ./output --model gemini-2.5-flash-lite

Convert with auto deployment of a small vlm (or any huggingface VLM model, requires a gpu + docker installation):

vlmparse convert "*.pdf" -o ./output --model nanonets/Nanonets-OCR2-3B

Deploy a local model server

Deployment (requires a gpu + docker installation):

• You need a gpu dedicated for this.
• Check that the port is not used by another service.

vlmparse serve lightonocr2 --port 8000 --gpu 1

then convert:

vlmparse convert "*.pdf" -o ./output --uri http://localhost:8000/v1

You can also list all running servers:

vlmparse list

You can get a list of registered models and their capabilities (ocr, ocr_layout, table, image_description) and inline image description with:

vlmparse registry

Show logs of a server (if only one server is running, the container name is not needed):

vlmparse log <container_name>

Stop a server (if only one server is running, the container name is not needed):

vlmparse stop <container_name>

View conversion results with Streamlit

vlmparse view ./output

Configuration

Set API keys as environment variables:

export GOOGLE_API_KEY="your-key"
export OPENAI_API_KEY="your-key"

Python API

Client interface:

from vlmparse.registries import converter_config_registry

# Get a converter configuration
config = converter_config_registry.get("gemini-2.5-flash-lite")
client = config.get_client()

# Convert a single PDF
document = client("path/to/document.pdf")
print(document.to_markdown())

# Batch convert multiple PDFs
documents = client.batch(["file1.pdf", "file2.pdf"])

Docker server interface:

from vlmparse.registries import docker_config_registry

config = docker_config_registry.get("lightonocr")
server = config.get_server()
server.start()

# Client calls...

server.stop()

Converter with automatic server management:

from vlmparse.converter_with_server import ConverterWithServer

with ConverterWithServer(model="mineru25") as converter_with_server:
    documents = converter_with_server.parse(inputs=["file1.pdf", "file2.pdf"], out_folder="./output")

Note that if you pass an uri of a vllm server to ConverterWithServer, the model name is inferred automatically and no server is started.

Credits

This work was realised by members of Probayes and OpenValue, two subsidiaries of La Poste.