docker-compose.onyx.yml

This page is part of the documentation for Orleans.Lattice 9.9.0 (release line 9.9), built 2026-10-04. It is also published as markdown, with every table and list, at docker-compose-onyx-yml.md, and llms.txt lists every page.

Part of RepoContextContainer source.

# Optional override: run the sample against the Onyx embedding companion
# (apps/embedding) instead of the default ONNX Runtime one (apps/embedding-onnx).
#
#   docker compose -f docker-compose.yml -f docker-compose.onyx.yml up -d
#
# This is opt-in and additive. The base compose file is unchanged, so plain
# `docker compose up` still brings up the ONNX companion.
#
# You almost certainly do not need this. The ONNX companion is the default
# because it is better on every measured axis - a ~10x smaller image (1.34 GB vs
# 13 GB), a markedly faster cold start, lower query latency, and no model
# download on first run - while emitting numerically identical vectors. This
# override exists for one reason: to fall back quickly if the ONNX image ever
# misbehaves on a particular host, without editing the base file.
#
# Only the embedder IMAGE changes. The service name, the port, and therefore the
# host''s LATTICE_EMBEDDING_ENDPOINT (http://embedder:9000) are all identical,
# because both images serve the same contract on the same port. The vectors are
# numerically identical too (fp32, same pinned model revision, same tokenizer and
# pooling), so switching either way does NOT require re-indexing an existing
# volume - see apps/embedding-onnx/README.md.

services:
  embedder:
    build:
      context: ../..
      dockerfile: apps/embedding/Dockerfile
      # The ONNX flavour argument means nothing to this image.
      args: !reset {}
      # Nor does the ONNX image's cgroup-reader named context.
      additional_contexts: !reset {}
    environment:
      # The ONNX image selects its accelerator with this variable; the Onyx image
      # does not read it. Reset so it cannot look like live configuration.
      EMBED_PROVIDER: null
    volumes:
      # Persist the HuggingFace model cache so a restart needs no re-download.
      # The ONNX image bakes its weights into an image layer and needs no cache,
      # so the base file declares no volume here and this override adds one.
      - hf-cache:/app/.cache/huggingface

volumes:
  # Only this override mounts it, so the declaration lives here rather than in
  # the base file where it would be an unused, misleading leftover.
  hf-cache: