Quick start

Docker Compose is the supported deployment path. CPU mode is ideal for evaluation; NVIDIA CUDA is recommended for real-time inference.

Requirements

• Git
• Docker Engine or Docker Desktop
• Docker Compose v2
• At least 8GB RAM; 16GB recommended
• At least four CPU cores for CPU evaluation
• An NVIDIA driver with CUDA container support for GPU mode

Clone the project

git clone https://github.com/kkive/video-search-local.git
cd video-search-local

CPU mode

docker compose up --build -d

NVIDIA CUDA mode

docker compose -f compose.yaml -f compose.gpu.yaml up --build -d

The first start prepares the YOLO26s and D-FINE-S ONNX models and takes longer than later starts. Open the application at:

http://127.0.0.1:8080

First-time configuration

1. Enter and test the RTSP or other stream URL under Settings → Video source.
2. Select YOLO26s or D-FINE-S and CPU/CUDA under Inference.
3. Enable ByteTrack under Tracking.
4. Create a rule and select the model labels to watch.
5. Configure clip duration, output frame rate, and annotation style.
6. Configure and test a VLM API if semantic review is required.
7. Test Qdrant, save all settings, and start inference.

If a stream runs on the Docker host, remember that 127.0.0.1 inside a container refers to the container. Docker Desktop users should normally use host.docker.internal.

See the complete English user guide for every setting and troubleshooting procedure.