Vision Tool

Image recognition using Ollama + qwen3.5:4b with think=False for reliable content extraction.

Image recognition using Ollama + qwen3.5:4b. Uses /api/chat endpoint for direct content extraction.

Features

✅ Direct content extraction - Uses /api/chat endpoint for clean output
✅ Simplified architecture - No complex thinking field processing needed
✅ English prompts - Optimized for English language analysis
✅ Multi-channel support - Works in WeChat, Telegram, Discord, etc.
✅ Error handling - Full error recovery and reporting

Installation

Prerequisites

1. Ollama service: ollama serve (must be running)
2. qwen3.5:4b model: ollama pull qwen3.5:4b
3. Python 3.8+: Required for running the skill

Install the skill

bash

clawhub install vision-tool

Development Setup (For Contributors)

If you want to contribute or modify the skill, see CONTRIBUTING.md for detailed development instructions.

Basic setup:

bash

# Clone the repository
git clone https://github.com/HuRuilizhen/vision-tool
cd vision-tool

# Set up development environment
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

# Run tests
python3 -m pytest tests/

Usage

Basic usage

bash

# From any OpenClaw channel
exec: python3 /path/to/vision-tool/main.py /path/to/image.jpg

# With custom prompt
exec: python3 /path/to/vision-tool/main.py /path/to/image.jpg --prompt "Describe this image"

# Debug output
exec: python3 /path/to/vision-tool/main.py /path/to/image.jpg --debug

Channel-specific examples

WeChat Channel:

bash

# When receiving an image
exec: python3 /path/to/vision-tool/main.py "$IMAGE_PATH"

Telegram Channel:

bash

# Reply to photo messages
exec: python3 /path/to/vision-tool/main.py "/path/to/telegram_photo.jpg"

Discord Channel:

bash

# Process attachments
exec: python3 /path/to/vision-tool/main.py "./discord_attachment.jpg"

Example Output

text

Analysis (30.7s):
------------------------------------------------------------
The user wants a description of the image provided.
**1. Overall Composition:**
- It's a top-down view of a meal served on a white tray.
- There are six distinct dishes/bowls arranged...
**2. Detailed Breakdown of Dishes:**
- **Top Left:** A small white rectangular dish...
- **Top Middle:** A small white rectangular dish...
------------------------------------------------------------

How It Works

1. Image reading: Reads and Base64 encodes the image
2. API call: Calls Ollama /api/chat endpoint with qwen3.5:4b
3. Direct extraction: Gets analysis directly from content field
4. Fallback handling: Simple cleanup if thinking field is used
5. Output formatting: Generates clean analysis results

Performance

• Average processing time: 25-35 seconds per image (hardware dependent)
• Image size support: 100KB-500KB recommended
• Token consumption: ~2000 tokens per image
• API endpoint: Uses /api/chat for direct content access

Troubleshooting

Common Issues

1. Ollama not running: Run ollama serve first
2. Model not installed: Run ollama pull qwen3.5:4b
3. Image path incorrect: Use absolute paths or correct relative paths
4. Timeout: Model may take 30+ seconds for complex images

Performance Tips

• Compress images to under 300KB for faster processing
• Use clear, concise prompts
• Ensure Ollama has sufficient system resources

API Reference

Python API

python

from vision_core import VisionAnalyzer

analyzer = VisionAnalyzer()
result = analyzer.analyze_image("image.jpg", "Describe this image")
print(result["analysis"])

Command Line

bash

# Basic analysis
python3 main.py image.jpg

# Custom prompt
python3 main.py image.jpg --prompt "What objects are in this image?"

# Debug mode
python3 main.py image.jpg --debug

Development

File Structure

text

vision-tool/
├── SKILL.md          # This documentation
├── main.py           # Main skill script
├── scripts/
│   └── vision_core.py  # Core analysis engine
└── tests/
    └── test_basic.py   # Basic tests

Testing

bash

# Test with example image
python3 main.py /path/to/test.jpg --prompt "Test analysis"

# Run unit tests
python3 -m pytest tests/

Changelog

v1.1.0 (2026-04-13)

• Uses /api/chat endpoint for direct content extraction
• Simplified architecture without complex thinking field processing
• Default English prompt "Describe this image"
• Removed regex dependencies for cleaner code

v1.0.0 (2026-04-12)

• Initial release

Contributing

Issues and pull requests are welcome. Please ensure tests pass before submitting.

License

This skill is part of the OpenClaw ecosystem.

---

Ready to use in all OpenClaw channels! 🚀