GovOn

AI-powered agentic CLI for Korean public-sector civil complaint workflows.

Government employees in South Korea spend hours drafting responses to civil complaints — looking up laws, finding precedents, and formatting official documents. GovOn automates that entire workflow. You type a natural-language request, and an AI agent retrieves relevant data, searches legal references, and generates a draft response in official document format.

The CLI is lightweight (~10 MB). The heavy lifting happens on a remote server running EXAONE 4.0-32B with domain-specific LoRA adapters for civil and legal tasks.

Quick Start

pip install govon
export GOVON_RUNTIME_URL=https://umyunsang-govon-runtime.hf.space
govon

Installation

Prerequisites: Python 3.10+ or Node.js 18+ or Homebrew

pip (recommended)

pip install govon

npm

npm install -g govon

Homebrew (macOS / Linux)

brew tap govon-org/govon && brew install govon

Self-hosted runtime (GPU required)

You can host your own GovOn runtime on Hugging Face Spaces or on-premises.

Option A: Hugging Face Spaces (recommended)

1. Duplicate the reference Space: umyunsang/govon-runtime
2. Select A100 GPU (80 GB) or higher hardware
3. Wait for the Space to start, then set your CLI to point at it:

export GOVON_RUNTIME_URL=https://<your-hf-username>-govon-runtime.hf.space
govon

Tip: Add the export to your shell profile (~/.zshrc, ~/.bashrc) so it persists across sessions.

Option B: Docker (on-premises)

govon server pull
govon server start

Option C: pip extras (bare-metal)

pip install govon[server]

Hardware requirement: The runtime requires an NVIDIA GPU with ≥ 80 GB VRAM (A100 80 GB or higher). Smaller GPUs will fail to load the EXAONE 4.0-32B model.

Usage

Interactive mode (REPL)

# Point the CLI to your runtime (HF Space, Docker, or localhost)
export GOVON_RUNTIME_URL=https://<your-hf-username>-govon-runtime.hf.space
govon

govon> Draft a response for a road damage complaint

┌─ Approval Request ────────────────────┐
│  Type: Draft response                 │
│  Goal: Road damage complaint response │
│  Tasks:                               │
│   - Look up legal basis               │
│   - Search similar cases              │
│                                       │
│  ● Approve  ○ Reject                  │
└───────────────────────────────────────┘

The agent proposes a plan and waits for your approval before executing any tools.

One-shot mode

govon "Show me road damage complaint statistics for this month"

Multi-turn conversations

govon --session my-session

govon> What are the most common complaint types?
→ (response with statistics)

govon> Draft a response for the top one
→ (uses conversation context to generate a draft)

API call

curl -X POST $GOVON_RUNTIME_URL/v3/agent/run \
  -H "Content-Type: application/json" \
  -d '{"query": "Show complaint statistics", "session_id": "demo-1"}'

Features

• ReAct agent with 7 tools -- the agent autonomously selects and chains tools based on your request
• Human-in-the-loop approval -- every tool execution requires explicit user approval before running
• Multi-turn conversations -- session-based context management with extractive summarization
• Multi-LoRA inference -- domain-specific adapters (civil complaints + legal references) on a single base model
• Streaming responses -- real-time SSE streaming with per-node progress display
• Three installation methods -- pip, npm, and Homebrew for maximum accessibility

Tools

Architecture

graph LR
    subgraph Client
        CLI["govon CLI<br/>pip / npm / brew"]
    end

    subgraph Server ["Server (HF Space or Docker)"]
        API["FastAPI"]
        AGENT["ReAct Agent<br/>+ 7 Tools"]
        LLM["EXAONE 4.0-32B<br/>+ civil LoRA<br/>+ legal LoRA"]
    end

    CLI -- "HTTP / SSE" --> API --> AGENT --> LLM

    style Client fill:#e0f2fe,stroke:#0284c7
    style Server fill:#fef3c7,stroke:#d97706

The CLI is thin, the server is powerful.

• CLI: httpx, rich, prompt-toolkit (~10 MB, no GPU needed)
• Server: EXAONE 4.0-32B + vLLM + Multi-LoRA on A100 80 GB

Server Management

Manage the Docker-based backend with govon server:

Configuration

export GOVON_RUNTIME_URL=https://umyunsang-govon-runtime.hf.space
export API_KEY=your-api-key

See deploy/env/.env.example for the full list of server-side configuration options.

Documentation

Resources

Contributing

git clone https://github.com/GovOn-Org/GovOn.git
cd GovOn
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for guidelines.

About

GovOn is an industry-academia project from the Department of Computer Engineering at Dong-A University. It assists local government employees with civil complaint response workflows. The AI does not replace human judgment -- it automates repetitive tasks so public servants can focus on decisions that matter.

License

The GovOn source code is licensed under MIT.

Important: The runtime uses EXAONE 4.0-32B-AWQ, which is distributed under the EXAONE AI Model License Agreement — non-commercial use only. Commercial deployment of the model requires a separate license from LG AI Research.