ace-docker-simulator

Fuuz Industrial Operations Platform — ACE Docker Simulator Version: 1.0.0 | Docs: https://accelerators.fuuz.com/ace-docker-simulator/

A running plant, a real industrial historian and a Fuuz-shaped GraphQL API, on your laptop, in one command. Part of ACE (Auto Contextualization Engine) from the Fuuz Industrial Operations Platform.

git clone https://github.com/Fuuz-Platform/ace-docker-simulator
cd ace-docker-simulator
docker compose up -d           # 5 containers, no configuration
open http://localhost:8081

No account, no tenant and no config file are needed for the simulator. Docker is the only prerequisite.

---

What runs

Service	Port	Does
ace-sim	4840, 4841	OPC UA server — 8 units × 11 signals (88 tags), process/discrete/counter/enum/string, with rotating COMMS/STALE/FLATLINE/DRIFT faults that emit real OPC UA status codes
ace-mongo	27017	MongoDB with mongot bundled — the time-series store, and $vectorSearch
ace-graphql	8098	Fuuz-shaped GraphQL API over Mongo, plus the historian query surface
ace-bridge	4842	OPC UA subscription → historian: deadband, plus a sample on every quality edge
ace-ui	8081	Operations console (React behind nginx, which proxies every service under /api/*)
ace-loadgen	—	Opt-in (--profile load) — high-rate writer for sizing work
ace-orchestrator	8099	Opt-in (--profile fuuz) — ACE match + classify against a live Fuuz tenant

The two opt-in services need something only you can supply — a tenant and a token for the orchestrator, and a deliberate decision to generate load for the loadgen — so neither starts by default.

Why the historian is not just a table of numbers

tagValue is a native MongoDB time-series collection. What makes it a historian, in graphql/historian.js:

Capability	Why it is not optional
Quality on every sample	Bad samples are stored, not dropped — discarding them hides outages and makes coverage a lie. Aggregates exclude them and count them separately.
Two timestamps (ts source, rt receive)	The difference is late arrival. Store-and-forward replays hours-old samples out of order after a WAN outage; latencyMs makes that visible.
Typed values (v / vb / vs)	Running=1 and Speed=1.0 are not the same kind of fact.
Interpolation mode per tag	Continuous interpolates, discrete holds. Linear interpolation of a state signal invents states that never happened.
Time-weighted aggregates	Exception-based samples are irregular, so an arithmetic mean over-weights whatever was sampled often. Discrete tags get state durations and transition counts instead.

Verified on simulator data:

PROCESS   Plant/PKG/TEM-001/Temp     arithmetic 121.332   TIME-WEIGHTED 121.589  <- preferred
DISCRETE  Plant/UTL/RUN-001/Running  ON 143s  OFF 37s  on-fraction 79.7%  transitions 6

Throughput and sizing

Measured on an Apple M5 Max with --profile load:

20,000 tags @ 25,000/s target  ->  24,415/s sustained, 0 errors
14.45M samples stored          ->  6.1 bytes/sample on disk (2.1x compression)
quality mix: 99.18% Good · 0.22% Bad/NotConnected · 0.22% Uncertain/Stale · 0.10% Bad/OutOfRange

Tune with LOADGEN_TAGS, LOADGEN_RATE, LOADGEN_BATCH_MS.

Pointing ACE at a live Fuuz tenant

cp .env.example .env      # fill in FUUZ_HOST, FUUZ_TENANT, FUUZ_TOKEN
docker compose --profile fuuz up -d
curl localhost:8099/health

Route	Does
GET /health	probes every dependency (Fuuz data API, embeddings, vectors, LLM)
POST /match	Tier-1 deterministic pass over PENDING candidates
POST /classify	measurement type + role for every candidate
POST /embed	embeds tag paths into MongoDB (the Tier-2 recall lane)
POST /similar	{"q":"chiller vibration"} — semantic search over live tag paths
POST /pipeline	Tier 1, then the LLM on the ambiguous band only

The last row is the important one. NO_MATCH is a Tier-1 rejection, and forwarding a rejection to a language model lets the model overturn it — which it will, confidently. Only the genuinely ambiguous band is routed, and every run reports what was withheld and why. The guardrail is not in the model, it is in what you route to it.

Inference

Defaults assume Docker Model Runner, which runs the model host-side on llama.cpp and so gets the GPU a Docker Desktop container never can:

docker model pull hf.co/ggml-org/bge-m3-Q8_0-GGUF
docker model pull ai/qwen3

Measured on 44 strings, same 1024-d model: Docker Model Runner (Metal) 6.4 ms/string · host LM Studio 9.7 · CPU container 20.9.

Do not substitute multilingual-e5-small. Every GGUF build of it has a broken tokenizer that returns bit-identical vectors for distinct CamelCase inputs (cosine 1.000000 for MotorTemp vs BatchId). It does not error — it returns well-formed unit-norm vectors — so it silently poisons the index. CamelCase is exactly what every OPC UA tag leaf is.

Repository layout

docker-compose.yml     the stack
docker/                Dockerfiles + the nginx config that fronts the console
simulator/             OPC UA plant server
graphql/               Fuuz-shaped GraphQL API + historian
bridge/                OPC UA subscription -> historian
loadgen/               high-rate writer
orchestrator/          ACE HTTP surface (opt-in)
core/kernel/           ACE match + classify kernels — dependency-free
docdrop/               unstructured-document ingestion used by the orchestrator
site/                  the published documentation site

Licence

No open-source licence is granted. © Fuuz. Published for Fuuz customers, partners and evaluators.