# Copy to .env and adjust. All values have sane defaults; the service runs with
# none of these set (in offline mode).

# Run the deterministic offline stub (1) or load the real model (0).
# Offline needs no network and no torch/transformers; it's the default.
OFFLINE=1

# Hugging Face repo (or local path) used only when OFFLINE=0.
MODEL_ID=LaelaZ/distilbert-emotion

# Port the API listens on.
PORT=8000

# Logging.
LOG_LEVEL=INFO

# Max items accepted in one /predict call and coalesced into one forward pass.
MAX_BATCH_SIZE=64

# Max characters allowed in a single input (cheap guard before tokenization).
MAX_TEXT_LENGTH=2000

# How long the micro-batcher waits (ms) to fill a batch before flushing.
# Smaller -> lower latency; larger -> higher throughput under concurrency.
BATCH_MAX_DELAY_MS=5