# LLM API — environment (copy to <LLM_API_ROOT>/.env or export before `llm-api serve`)

APP_ENV=production
DOMAIN=www.llm-api.io
PUBLIC_URL=https://www.llm-api.io

# Bind address of the API gateway (the public reverse proxy talks to this)
HOST=127.0.0.1
PORT=8300
# Where the Next.js dashboard runs (proxied by the API for non-/api, non-/v1 paths)
WEB_URL=http://127.0.0.1:8301

# Data root: models/, data/ (SQLite + secrets), logs/
LLM_API_ROOT=/Users/USERNAME/llm-api
# Override the model directory (e.g. an external SSD)
# MODEL_ROOT=/Volumes/LLM/models

# Memory policy (GB). 64 GB machine: 8 macOS + 3 app + 4 KV + 4 safety = 45 GB for models
MAX_MODEL_MEMORY_GB=45
ABSOLUTE_MAX_MEMORY_GB=50
MACOS_RESERVE_GB=8
MIN_FREE_DISK_GB=100

MAX_SIMULTANEOUS_MODELS=1
MODEL_IDLE_TIMEOUT_MINUTES=30
SMALL_MODEL_RESIDENT_GB=3
LOAD_TIMEOUT_SECONDS=900
GENERATION_TIMEOUT_SECONDS=1800
PRELOAD_MODEL=none

ENABLE_MLX=true
ENABLE_GGUF=true
ALLOW_DOWNLOADS=true
LOG_PROMPTS=false

DEFAULT_MAX_TOKENS=2048
DEFAULT_CONTEXT=16384

# Hugging Face token for gated repositories (never exposed to the frontend)
HF_TOKEN=
LLAMA_SERVER_BIN=llama-server

# Sessions/CSRF secret (auto-generated into data/.secret when empty)
SECRET_KEY=
# First-run admin (used only when no user exists)
ADMIN_EMAIL=
ADMIN_PASSWORD=
SESSION_HOURS=336
SECURE_COOKIES=true

METRICS_INTERVAL_SECONDS=15
METRICS_RETENTION_DAYS=30