docker-compose.yml

# Self-host the scoring service ("brain") next to the local proxy.
#
#   docker compose up -d            # brain on :8090, calibrated tau, in-process bge
#   PARSEC_BRAIN_URL=http://127.0.0.1:8090 scripts/proxy_dev.sh   # point the proxy at it
#
# The proxy stays a HOST process on purpose: your agent's auth headers never
# enter a container (docs/routing-and-liveness.md). The brain needs a curator
# checkpoint — either mount one (default: ~/.parsec/brain/curator_v4_prod.pt)
# or pull the released base model from the Hugging Face Hub:
#
#   PARSEC_CKPT=hf://<org>/<repo>/<file>.pt docker compose up -d
#
# packages/brain/README.md covers the knobs; docs/environment-variables.md §6
# lists every variable.
services:
  brain:
    build: ./packages/brain
    image: parsec-brain:local
    ports:
      - "8090:8080"
    volumes:
      - ${HOME}/.parsec/brain:/bundle:ro
      # Hugging Face cache: the bge weights (~1.3 GB) and any hf:// checkpoint
      # download once and survive container rebuilds.
      - parsec-hf-cache:/root/.cache/huggingface
    environment:
      # local (default): the bge-large encoder in-process — the trained
      # vectors, honest scores. hash: deterministic test vectors that exercise
      # the wiring but say NOTHING about curation quality (for smoke tests).
      PARSEC_EMBED_BACKEND: ${PARSEC_EMBED_BACKEND:-local}
      PARSEC_CKPT: ${PARSEC_CKPT:-/bundle/curator_v4_prod.pt}
      PARSEC_TARGET_COV: ${PARSEC_TARGET_COV:-0.70}
      # Leave unset for the checkpoint's calibrated tau. A forced value (e.g.
      # 0.999) makes the machinery visibly trim regardless of the model's
      # opinion — demo/testing only.
      PARSEC_SERVE_TAU: ${PARSEC_SERVE_TAU:-}
      # Optional shared bearer secret; the proxy sends it as PARSEC_BRAIN_KEY.
      PARSEC_BRAIN_KEY: ${PARSEC_BRAIN_KEY:-}
      PARSEC_BRAIN_LOG: ${PARSEC_BRAIN_LOG:-INFO}
      PARSEC_BRAIN_LOG_JSON: ${PARSEC_BRAIN_LOG_JSON:-}
    restart: unless-stopped

volumes:
  parsec-hf-cache: