Skip to content

Quickstart

From zero to first detection in under a minute — no GPU, no weights, no network.

1. Install

git clone https://github.com/Phoenix0531-sudo/Hokage_Vision_Agent.git
cd Hokage_Vision_Agent
pip install -e .[dev]

2. Run the one-command demo

python examples/quickstart.py

This walks the full pipeline with the deterministic mock backend and writes every artifact under runs/quickstart/:

  1. Detect one image (with rendered overlay and JSON output)
  2. Detect a folder
  3. Validate the example YOLO dataset
  4. Run a smoke training plan (safe dry-run by default)
  5. Ask the rule-based agent to detect a folder
  6. Generate a markdown report

Expected output ends with:

=== 5. Ask the rule-based agent ===============================
tool: detect_folder | status: success | count: 1
...
=== Done ======================================================
All artifacts are under: runs\quickstart

3. Try the CLI

hokage-vision --help
hokage-vision detect image examples/images/sample.jpg --backend mock
hokage-vision agent run "检测 examples/images 里的图片"

The mock backend returns deterministic detections for obito, naruto, and gaara, so every surface (CLI, API, GUI, agent) works before real model weights exist.

4. Real inference (optional)

Real weights stay external. After placing a reviewed weight file under models/ (a YOLO .pt or exported .onnx), run it explicitly:

hokage-vision detect image examples/images/sample.jpg --backend ultralytics --model-path models/your-model.onnx --device cpu

ONNX Runtime is used automatically for .onnx models; install the training extra (pip install -e .[train]) for .pt weights.

5. Next steps

  • Usage — full CLI, API, GUI, and agent reference
  • Data And Models — dataset governance and license guardrails
  • Architecture — backend factory and service layer design