docker-compose.semantic.yml

Rama main · réplica de lectura del repositorio original.

# polyrouter -semantic variant override (add-semantic-dashboard 4.1).
#
# Layer it over the base stack to run the batteries-included semantic image —
# the ONNX runtime and the reference embedding model are pre-baked, so L2 is
# available on first boot with zero manual steps (the same module powers the
# semantic WORKLOAD source — research / writing detection — no extra files):
#
#   docker compose -f docker-compose.yml -f docker-compose.semantic.yml up -d
#
# `docker compose ... build` builds the `runtime-semantic` Dockerfile target;
# to run a PUBLISHED variant instead, set POLYROUTER_IMAGE to the `-semantic`
# tag (e.g. ghcr.io/izzoa/polyrouter:latest-semantic) and skip the build.

services:
  app:
    build:
      context: .
      target: runtime-semantic
    image: ${POLYROUTER_IMAGE:-polyrouter:latest-semantic}
    environment:
      # Turn the L2 capability on instance-wide so `semanticAvailable` is true
      # out of the box (the baked model satisfies the embedder half). Tenants
      # still opt in per-account via the dashboard.
      ROUTING_AUTO_LAYERS: ${ROUTING_AUTO_LAYERS:-structural,semantic,cascade}
      # The image bakes SEMANTIC_MODEL_PATH=/app/models/reference. Override it to
      # a mounted path for bring-your-own-model (see the volume below).
      SEMANTIC_MODEL_PATH: ${SEMANTIC_MODEL_PATH:-/app/models/reference}
      SEMANTIC_HIGH_THRESHOLD:
      SEMANTIC_LOW_THRESHOLD:
      SEMANTIC_CONCURRENCY:
      SEMANTIC_TIMEOUT_MS:
      SEMANTIC_MAX_INPUT_CHARS:
      # Semantic WORKLOAD rails (add-semantic-workloads): research/writing detection.
      SEMANTIC_WORKLOAD_MARGIN:
      SEMANTIC_WORKLOAD_MIN_SIM:
      # Optional L2 learning loop (default off; opt-in PER TENANT in the UI —
      # there is no global enable env). These tune the sweep + its bounds.
      SEMANTIC_LEARNING_ALPHA:
      SEMANTIC_LEARNING_MAX_DRIFT:
      SEMANTIC_LEARNING_MIN_SAMPLES:
      SEMANTIC_LEARNING_MIN_COHORT:
      SEMANTIC_LEARNING_MAX_COHORTS:
      SEMANTIC_LEARNING_COOLDOWN_H:
      SEMANTIC_LEARNING_STATE_TTL_D:
      SEMANTIC_LEARNING_SCHED_ENABLED:
      SEMANTIC_LEARNING_SCHED_CRON:
    # Bring-your-own-model: mount a bundle (model.onnx + vocab + manifest.json)
    # and point SEMANTIC_MODEL_PATH at it in .env to override the baked default:
    #   SEMANTIC_MODEL_DIR=/abs/path/to/your/bundle
    #   SEMANTIC_MODEL_PATH=/app/models/custom
    # then uncomment:
    # volumes:
    #   - ${SEMANTIC_MODEL_DIR}:/app/models/custom:ro