Docker Compose stack that runs OpenWebUI with Ollama (GPU-backed) and custom tool services: Time UI (time/timezone utilities) and HereYouAre (vision-based object detection).
Upload image → processing → bounding boxes and labels (Mouse, Keyboard, Headphones, Deskmat).
| Service | Description |
|---|---|
| ollama | Local LLM runtime; runs vision/language models (e.g. qwen2.5-vl). Uses GPU when available. |
| openwebui | Web UI for chat and tools; connects to Ollama and to time-ui / hereyouare. |
| time-ui | OpenAPI tool: UTC/local time, formatting, timezone conversion. Used by OpenWebUI as a connection. |
| hereyouare | Object detection on images via Ollama vision models; returns JSON or an annotated image. |
- Docker and Docker Compose
- NVIDIA Container Toolkit (for GPU support with Ollama)
# Start all services (Ollama will pull images on first run)
docker compose up -d
# Open Web UI
open http://localhost:3000
# Pull a vision model for HereYouAre (e.g. qwen2.5-vl)
docker exec -it ollama ollama pull qwen2.5vl:7b-q4_K_M- OpenWebUI: http://localhost:3000
- Ollama API: http://localhost:11434 (bound to 127.0.0.1)
- HereYouAre API: http://localhost:8082
FastAPI service that detects physical objects in images using an Ollama vision model (e.g. qwen2.5vl). Images are resized (max side 1024 px), sent to Ollama with a structured prompt, and the response is parsed into bounding boxes.
Для интерактивной работы с детекцией доступно веб-приложение HereYouAre UI (React): загрузка изображения, отображение рамок и миниатюр обнаруженных объектов. Пример интерфейса:
| Method | Path | Description |
|---|---|---|
| GET | /health |
Health check; returns {"status": "ok"}. |
| GET / POST | /echo |
Echo request (query or body); useful for connectivity checks. |
| POST | /v1/detect |
Upload an image; returns JSON with image dimensions and a list of detections (label + pixel position [x1, y1, x2, y2]). |
| POST | /v1/detect/image |
Same detection, but returns a PNG image with corner-only bounding boxes and labels drawn (same colour tone). |
{
"image": { "width": 1024, "height": 768 },
"detections": [
{ "label": "keyboard", "position": [360, 375, 1022, 568] },
{ "label": "mouse", "position": [192, 383, 337, 570] }
],
"provider": "qwen2.5vl",
"latency_ms": 3500
}| Input | Annotated result (corner boxes + labels) |
|---|---|
 |
 |
Original image and the same scene with detected objects marked (HereYouAre /v1/detect/image).
# JSON result
curl -X POST http://localhost:8082/v1/detect \
-F "[email protected]"
# Annotated image (PNG)
curl -X POST http://localhost:8082/v1/detect/image \
-F "[email protected]" \
--output annotated.png- Swagger UI: http://localhost:8082/docs
- ReDoc: http://localhost:8082/redoc
Edit hereyouare/settings.py to change:
- Ollama URL, model name, timeout, temperature
- Prompt and FORMAT_SCHEMA (guided JSON output)
- Image max side, JPEG quality
- OLLAMA_CONCURRENCY (default 1 to avoid overloading the GPU)
The stack includes a commented nginx service for HTTPS (e.g. for remote microphone or production). Uncomment it in docker-compose.yaml, then create certificates (see nginx/ssl/README.txt or nginx/ssl/gen-selfsigned.sh) and adjust nginx/conf.d as needed.
.
├── docker-compose.yaml # Ollama, OpenWebUI, time-ui, hereyouare (optional nginx)
├── hereyouare/ # Object detection API (FastAPI, Pillow, orjson)
│ ├── main.py
│ ├── settings.py
│ ├── Dockerfile
│ └── requirements.txt
├── time-ui/ # Time tool (OpenAPI)
├── nginx/ # Optional reverse proxy config and SSL
└── README.md
See repository or project-specific files for license information.