Laya System 1 on Unraid

Laya is a fast "System 1" decision engine. You give it some text and a set of typed questions. It answers every question in one pass, each with a calibrated probability:

• choice picks one option from a list
• score places the text on a scale
• noul answers yes, no or unsure

It serves the Jev-compatible /v1/systemone HTTP API. Use it for routing, triage, moderation and guardrails in front of a larger model.

Laya has no official Docker image. This template runs pikkonmg/laya-system-one, which is built from the upstream source and rebuilt when Laya has a new release. The Dockerfile and the build workflow are public at PikkonMG/laya-system-one-docker.

First start

1. Copy templates/laya-system-one.xml to /boot/config/plugins/dockerMan/templates-user/my-laya-system-one.xml on Unraid. Then open Docker > Add Container and pick laya-system-one from the user templates.

2. Pick a tag. Community Applications asks you when you click Install:

   Tag	Extra Parameters	Also needed
   cpu	Empty	Nothing
   nvidia	--runtime=nvidia, filled in for you	Unraid NVIDIA Driver plugin, driver 560 or newer

   From a user template, set Repository to pikkonmg/laya-system-one:nvidia and Extra Parameters to --runtime=nvidia yourself.

   If the nvidia container cannot see the GPU, it stops and the log tells you what is missing.

3. Optional: set API key. Leave it blank and the API needs no key. When you set one, clients must send it.

4. Apply the template. The first start downloads the model weights, a few GB, into /mnt/user/appdata/laya-system-one. The container shows healthy when the models are loaded.

5. Open http://YOUR_UNRAID_IP:8012/docs to see the API.

Call the API

Send the text as state and your questions as questions. Leave out the Authorization line when you did not set an API key.

curl -s http://YOUR_UNRAID_IP:8012/v1/systemone \
  -H 'content-type: application/json' \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  --data '{
    "state": {
      "body": "I was charged twice for my subscription. Please refund the duplicate charge."
    },
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which department should handle this request?",
        "criteria": {
          "billing": "invoices, payments, refunds",
          "technical": "bugs, outages, system errors",
          "sales": "pricing, new contracts"
        }
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent is this request?",
        "criteria": ["not urgent", "needs attention soon", "critical deadline or blocking issue"]
      },
      "refund_requested": {
        "type": "noul",
        "instructions": "Does the user explicitly request a refund?"
      }
    }
  }'

GET /health needs no key. It shows the device and the loaded models.

Models

Laya has three models. A router picks one for each request.

Name	Use
english	English text, up to 512 tokens
multilingual	100+ languages, up to 1024 tokens
typed-decisions	Structured decisions, up to 1024 tokens

Models to load picks which models load at start:

Choice	Memory, about
english,typed-decisions (default)	4 GB
english	2 GB
english,multilingual	3.5 GB
english,multilingual,typed-decisions	5.5 GB

Automatic routing only picks english or multilingual. A request uses typed-decisions when it names that model, or when Auto typed decisions is 1. Text in another language still loads multilingual on first use, even when it is not in the list.

Storage and updates

The appdata path maps to /data inside the container. It holds only downloaded model weights, so you can delete it and they download again. The container sets the folder owner to PUID:PGID at start.

Set Offline mode to 1 to stop all downloads once the weights are in the cache.

Sources

• Laya project
• Image build repository
• Laya documentation
• Laya HTTP serving settings