49 lines
1.8 KiB
YAML
49 lines
1.8 KiB
YAML
version: '3.8'
|
|
|
|
services:
|
|
rag-service:
|
|
build:
|
|
context: .
|
|
dockerfile: Dockerfile
|
|
container_name: rag-service
|
|
restart: unless-stopped
|
|
# Порт открываем только для localhost (NGINX будет проксировать)
|
|
# Для прямого доступа используй: "8000:8000"
|
|
ports:
|
|
- "127.0.0.1:8000:8000"
|
|
volumes:
|
|
# Персистентность данных модели и векторов
|
|
- ./data/models:/app/data/models
|
|
- ./data/vectors:/app/data/vectors
|
|
environment:
|
|
- RAG_MODEL=${RAG_MODEL:-sentence-transformers/all-MiniLM-L12-v2}
|
|
- RAG_CACHE_DIR=/app/data/models
|
|
- RAG_VECTORS_PATH=/app/data/vectors/vectors.npz
|
|
- RAG_MAX_EXAMPLES=${RAG_MAX_EXAMPLES:-10000}
|
|
- RAG_SCORE_MULTIPLIER=${RAG_SCORE_MULTIPLIER:-5.0}
|
|
- RAG_BATCH_SIZE=${RAG_BATCH_SIZE:-16}
|
|
- RAG_MIN_TEXT_LENGTH=${RAG_MIN_TEXT_LENGTH:-3}
|
|
- RAG_API_HOST=0.0.0.0
|
|
- RAG_API_PORT=8000
|
|
# Безопасность
|
|
- RAG_API_KEY=${RAG_API_KEY}
|
|
- RAG_ALLOW_NO_AUTH=${RAG_ALLOW_NO_AUTH:-false}
|
|
# Автосохранение
|
|
- RAG_AUTOSAVE_INTERVAL=${RAG_AUTOSAVE_INTERVAL:-600}
|
|
- LOG_LEVEL=${LOG_LEVEL:-INFO}
|
|
healthcheck:
|
|
test: ["CMD", "python", "-c", "import urllib.request, os; key = os.getenv('RAG_API_KEY', ''); req = urllib.request.Request('http://localhost:8000/api/v1/health'); req.add_header('X-API-Key', key) if key else None; urllib.request.urlopen(req)"]
|
|
interval: 30s
|
|
timeout: 10s
|
|
retries: 3
|
|
start_period: 60s
|
|
# Ограничения ресурсов (рекомендуется для продакшена)
|
|
# deploy:
|
|
# resources:
|
|
# limits:
|
|
# cpus: '2'
|
|
# memory: 4G
|
|
# reservations:
|
|
# cpus: '1'
|
|
# memory: 2G
|