system-one

A vendor-neutral SDK for System One models. One contract — ask(state, questions) -> answers — over three question primitives (noul, choice, score). Switching providers is one config object, not a code change.

These are decision models, not chat models: no messages, no streaming, no temperature.

Installation

pip install "system-one[http]"   # hosted providers
pip install "system-one[onnx]"   # local, in-process

Usage

from system_one import SystemOne, TypesafeConfig

with SystemOne(TypesafeConfig()) as agent:  # reads $SYSTEM_ONE_API_KEY
    response = agent.ask(
        "Customer is furious about a double charge.",
        {
            "urgent": {"type": "noul", "instructions": "Does this need a human now?"},
            "topic": {
                "type": "choice",
                "instructions": "Which queue?",
                "criteria": ["billing", "technical", "other"],
            },
            "severity": {
                "type": "score",
                "instructions": "How severe?",
                "criteria": ["minor", "normal", "major", "critical"],
            },
        },
    )

print(response.nouls["urgent"].noul, response.nouls["urgent"].confidence)
print(response.choices["topic"].choice, response.choices["topic"].probabilities)
print(response.scores["severity"].score)

AsyncSystemOne has the same shape and the same method names, awaited: await agent.ask(...), await agent.close(), async with.

Configuration

All settings come from SYSTEM_ONE_* environment variables or a .env file, and any of them can be set in code instead: SystemOne(HTTPConfig(..., timeout=30)).

Vendor choice is a config class, not a string:

from system_one import (
    HTTPConfig,
    ONNXConfig,
    OpenRouterConfig,
    SystemOne,
    TypesafeConfig,
)

SystemOne(TypesafeConfig())  # hosted jev, $SYSTEM_ONE_API_KEY
SystemOne(OpenRouterConfig())  # same body, different endpoint
SystemOne(HTTPConfig(base_url="https://my-host", model="mine"))
SystemOne(ONNXConfig())  # local graph, no network, no key

Backend-specific calls stay reachable through agent.backend. Tests can inject one directly: SystemOne(using=FakeBackend()).

Running locally

The onnx backend runs a decision model in-process — no network, no API key, no torch at runtime. The weights come from laya (convaiinnovations/laya), in the layout the backend reads:

pip install "system-one[onnx,hub]"
system-one fetch --out-dir onnx   # --variant fp32|int8|fp16, --name sets the base name

That writes onnx/laya.onnx (≈1.6 GB fp32), onnx/laya.json and onnx/tokenizer/ — no torch, no tracing. To use another laya variant or your own model, write a spec and run system-one export examples/export/laya-english.yaml; that needs system-one[export]. Unset, SYSTEM_ONE_MODEL is whatever the backend serves — laya here — so a local run only sets it for a differently named graph. Nothing leaves the machine:

SYSTEM_ONE_BACKEND=onnx
SYSTEM_ONE_ONNX_DIR=onnx   # default
SYSTEM_ONE_MODEL=laya      # the file base name in that directory

with SystemOne(ONNXConfig(model="laya")) as agent:
    response = agent.ask(
        "The site is down.",
        {"outage": {"type": "noul", "instructions": "Is there an outage?"}},
    )

Full walkthrough — fetch vs export, variants, several models in one directory, verification, troubleshooting: docs/usage/local-model.md.

Errors

SystemOneError                 # catch this
├── APIError                   # non-2xx; .status .body .retry_after
│   └── AuthenticationError    # 401/403, never retried
└── APIConnectionError
    └── APITimeoutError

A malformed response body raises pydantic's ValidationError.

Development

uv sync --all-extras
uv run pytest
uv run pre-commit run --all-files

Tests needing the ONNX artifacts skip themselves when onnx/laya.onnx is absent (SYSTEM_ONE_ONNX_DIR and SYSTEM_ONE_MODEL override where they look). scripts/check_onnx_parity.py checks the graph against the reference laya implementation and needs the export extra plus the real weights.