krite

The Krite decision runtime in process: Protocol v1 answers from Python, with the same runtime, the same bytes, and the same errors as krite serve, and no server.

pip install krite
hf download skyoo2003/krite-0.15b-v1 --local-dir krite-0.15b-v1

import krite

k = krite.Krite("krite-0.15b-v1")
r = k.decide(
    {
        "state": [
            {"from": "customer", "text": "Hi, I was charged twice for my March subscription."},
            {"from": "agent", "text": "Sorry about that. Can you share the invoice number?"},
        ],
        "questions": {
            "route": {"type": "choice", "criteria": {"billing": None, "technical": None, "other": None}},
            "refund_requested": {"type": "noul"},
        },
    }
)
print(r["answers"]["route"]["choice"], r["answers"]["refund_requested"]["noul"])

decide accepts a dict, a JSON string, or JSON bytes and returns the response as a dict. The model alias jev-latest resolves to the loaded model, as on the server.

Errors raise krite.KriteError with status, type, message, and param. A missing or broken model directory raises RuntimeError when Krite is constructed.

device="auto" uses Metal on Apple silicon and the CPU elsewhere; device="cpu" forces the CPU. Calls run one at a time on one Krite and release the GIL, so threads may share it. See docs/runtime.md.