Argus

The all-seeing tagger. Argus is the auto-tagging service of the Codex media constellation — and its one non-JVM member. It watches uploaded media and suggests tags, turning tagging from "type from scratch" into "review a pre-filled set." Suggestions are never auto-applied; a human reviews them in artemis-ui's /review queue.

  Artemis ──TagJob──▶ HermesMQ (media.tag) ──▶ Argus ──▶ read sample (Apollo)
                                                  │        wd-tagger v3 + RAM++
                                                  │        threshold · normalize · merge
                                                  ▼
  Artemis ◀──TagSuggestions── HermesMQ (media.tags.suggested) ◀──┘
     alias-resolve · dedup(max) · flag needs-review ──▶ artemis-ui /review

Why Python

ML inference is Python's ecosystem. Isolating it as a dedicated service keeps that dependency out of the Scala/Pekko fleet. Argus speaks two protocols outward — HermesMQ REST (consume/ack) and Apollo gRPC (read the sample) — everything else is local CPU inference.

Models (CPU, no GPU)

Runtime is chosen per model:

Run-both-and-merge (rather than classify-and-route) is simple and robust to misclassification — any noise is filtered by the human review anyway. Argus normalizes only surface form; meaning-level alias merging is Artemis's job.

RAM++ note. No official ONNX export exists; torch-CPU is the pragmatic path. An ONNX export stays a later optimization (its 4,585 tag embeddings are frozen, so inference is a static graph) and is gated on a latency spike on the target node.

Develop

uv sync                 # core deps only — no ML stack needed for tests/lint
uv run pytest           # the pure pipeline, Hermes client, and consumer are fully tested
uv run ruff check .
uv run mypy

# End-to-end against a real Hermes, no weights required:
ARGUS_STUB_MODELS=1 uv run argus

Install the heavy inference stack only on the target node:

uv sync --extra ml --extra grpc

Configuration

All via env (prefix ARGUS_): HERMES_BASE_URL, SUBSCRIPTION_ID, SOURCE_TOPIC (media.tag), RESULTS_TOPIC (media.tags.suggested), APOLLO_ENDPOINT, PULL_MAX, ACK_DEADLINE_SECONDS, WD_MODEL_PATH, WD_TAGS_PATH, RAM_MODEL_PATH, GENERAL_THRESHOLD, CHARACTER_THRESHOLD, RAM_THRESHOLD, NUM_THREADS, HEALTH_PORT, STUB_MODELS.

Deploy

Own Docker image → Docker Hub (matching siblings). Deployed by Codex, pinned to a CPU-headroom node; model weights mounted from the NAS at /models. One replica is plenty — Hermes buffers if Argus falls behind.

Set a dead-letter topic on Hermes (HERMESMQ_DEAD_LETTER_TOPIC, e.g. media.tag.dlq) or exhausted poison jobs are silently dropped.

Status

Scaffold from the design-argus OpenSpec change. The pure pipeline, the Hermes REST client, and the consumer are implemented and tested; model inference is wired but needs weights + the ml extra on the target node (see openspec/ and the task list for what remains — the RAM++ latency spike gates the two-model design).