Quickstart

Ollaya runs open decision models on your own machine. You give a model a state — a message, an email, a ticket, a JSON object — plus a few typed questions, and it returns typed answers with calibrated probabilities in a single forward pass.

1. Install

On Linux and macOS:

curl -fsSL https://ollaya.dev/install.sh | sh

The script downloads the latest release from GitHub and checks its sha256. On Linux it also fetches the CUDA libraries when it finds an NVIDIA GPU, and, where systemd runs and it has root rights, sets up a service that serves the API on 127.0.0.1:11435. See Download for requirements and the Docker images.

2. Run a model

ollaya run laya --preset triage "I was charged twice for my subscription this month and want a refund."

intent            refund                                ████████████████ 1.00
is_urgent         no                                    ██████████████░░ 0.88
frustration       1.76 / 3  clearly annoyed             ██████░░░░░░░░░░ 0.36
refund_requested  yes                                   ██████████████░░ 0.90
churn_risk        no                                    ██████████░░░░░░ 0.61

ollaya run starts the server if it isn't running, pulls the model on first use and loads it. laya is a router: it sends English text to laya:en and other languages, Turkish for example, to laya:multilingual. Pulling laya pulls both.

--preset NAME asks a built-in question set: triage, email, guard, moderation or router.
--verbose adds every option's probability, the routing decision and timings.
--format json prints the full API response.
Without a state, ollaya run reads piped stdin, or opens a prompt on a terminal.

3. Ask your own questions

Write the questions to a file:

{
  "topic": {
    "type": "choice",
    "instructions": "What is this message about?",
    "criteria": {
      "billing": "Payments, invoices and refunds",
      "access": "Login, passwords and permissions",
      "other": "Anything else"
    }
  },
  "urgency": {
    "type": "score",
    "instructions": "How urgent is this?",
    "criteria": ["Can wait", "Needs attention this week", "Needs attention today"]
  },
  "angry": {
    "type": "noul",
    "instructions": "Is the customer angry?"
  }
}

ollaya run laya --questions questions.json "Hi, I cannot log in since this morning and I have a demo at 3pm."

Or send the same questions to the API:

curl http://localhost:11435/api/decide -d '{
  "model": "laya",
  "state": "Hi, I cannot log in since this morning and I have a demo at 3pm.",
  "questions": {
    "angry": {"type": "noul", "instructions": "Is the customer angry?"}
  }
}'

Every answer comes back typed: a choice with a probability per option, a score as the expected level, and a noul as the probability that the statement holds. See the API reference.

4. Use an existing TypeSafe client

Ollaya serves TypeSafe's API too. The official TypeSafe Python SDK works unchanged:

export TYPESAFE_BASE_URL=http://localhost:11435
export TYPESAFE_API_KEY=local        # any value; the SDK needs one
export TYPESAFE_DEFAULT_MODEL=laya

See TypeSafe compatibility.

5. Bake your questions into a model

A Modelfile turns a question set into a model you can run by name:

FROM laya
QUESTIONS ./questions.json
DESCRIPTION Support inbox triage