Triton

GPU-accelerated media-to-text REST API. Accepts YouTube URLs, audio/video files, PDFs, and images, converts them to text using faster-whisper and PaddleOCR, and stores results in PostgreSQL.

Architecture

┌───────────┐    ┌────────────┐    ┌───────────────┐
│  Client   │───▶│  FastAPI   │───▶│  PostgreSQL   │
│           │    │ (REST API) │    │ (tasks, docs) │
└───────────┘    └─────┬──────┘    └───────────────┘
                     │
                ┌────▼─────┐
                │  Redis   │
                │ (broker) │
                └────┬─────┘
           ┌─────────┴─────────┐
     ┌─────▼──────┐       ┌────▼─────┐
     │ GPU Worker │       │CPU Worker│
     │ -whisper   │       │ -yt-dlp  │
     │ -PaddleOCR │       │ -download│
     └────────────┘       └──────────┘

• API: FastAPI REST endpoints for tasks, documents, schedules, health
• GPU Worker: Celery worker running faster-whisper (transcription) and PaddleOCR (text extraction)
• CPU Worker: Celery worker running yt-dlp (YouTube/podcast downloads), chains to GPU worker
• Two queues: gpu and cpu to prevent download tasks from blocking GPU processing

API Endpoints

Method	Endpoint	Description
POST	/tasks	Create task (auto-dispatches to worker)
POST	/tasks/upload	Upload file for processing
GET	/tasks/{id}	Get task status and result
GET	/tasks	List tasks (filter by type, status)
DELETE	/tasks/{id}	Delete task
POST	/documents	Store text document
GET	/documents/{id}	Get document
GET	/documents	List documents (filter by type)
POST	/schedules	Create recurring schedule
GET	/schedules	List schedules
PUT	/schedules/{id}	Update schedule
DELETE	/schedules/{id}	Delete schedule
GET	/health	GPU status, VRAM, queue lengths

Task Types

Type	Source	Worker
youtube	URL	CPU → GPU (download then transcribe)
podcast	URL	CPU → GPU (download then transcribe)
video	File/URL	GPU (transcribe)
audio	File/URL	GPU (transcribe)
pdf	File/URL	GPU (OCR)
image	File/URL	GPU (OCR)

Tech Stack

• Python 3.11, FastAPI, Celery, Redis, PostgreSQL, SQLAlchemy, Alembic
• faster-whisper (large-v3, CTranslate2) — ~5GB VRAM
• PaddleOCR (chinese_cht) — ~2GB VRAM
• yt-dlp for YouTube/podcast downloads
• Docker Compose (5 services)

Quick Start

Local Development (no GPU)

docker compose -f docker-compose.dev.yml up --build -d
curl http://localhost:8000/health

Production (with NVIDIA GPU)

docker compose up --build -d

Run Tests

uv venv && uv pip install -e ".[dev]"
uv run python -m pytest tests/ -v

Environment Variables

See .env.example for all available settings.

Variable	Default	Description
DATABASE_URL	postgresql://triton:triton@localhost:5432/triton	PostgreSQL connection
REDIS_URL	redis://localhost:6379/0	Redis broker
WHISPER_MODEL	large-v3	faster-whisper model
WHISPER_DEVICE	cuda	cuda or cpu
UPLOAD_DIR	/data/tmp	Temp file storage