TypeSafe (JEV)

TypeSafeClient evaluates text or structured state against named Noul, Choice, and Score questions using the System One API. It has its own evaluation interface; use it directly rather than LLMClient or AgentSession.

Configuration

Set TYPESAFE_API_KEY in your environment or pass api_key= explicitly. The client defaults to model="jev-latest", timeout=60, and base_url="https://api.typesafe.ai/v1/". The client does not load .env itself.

Requests use niquests' built-in retry handling: up to two retries for connection errors, timeouts, HTTP 408/429, and 5xx responses, respecting Retry-After. Exhausted rate limits raise TooManyRequestsError; other provider errors raise LLMError with provider="typesafe".

Evaluate several questions

from padwan_ai import TypeSafeClient
from padwan_ai.typesafe import ChoiceQuestion, NoulQuestion, ScoreQuestion

async with TypeSafeClient() as client:
    response = await client.system_one(
        state={"ticket": "I was charged twice. Please fix this today."},
        questions={
            "billing": NoulQuestion(
                type="noul", instructions="Is this ticket about billing?"
            ),
            "tone": ChoiceQuestion(
                type="choice",
                instructions="What is the customer's tone?",
                criteria={"calm": None, "frustrated": None, "angry": None},
            ),
            "urgency": ScoreQuestion(
                type="score",
                instructions="How urgent is this ticket?",
                criteria=["can wait", "this week", "today"],
            ),
        },
    )

print(response["answers"])
print(response["model"], response["usage"])

Responses are typed dictionaries preserving the API's model, answers, and usage fields. Noul answers contain a probability, Choice answers contain a selected label and probabilities, and Score answers contain a score, legend, and probabilities. Score legend and probability keys remain strings ("0", "1", etc.). Usage exposes input_tokens and output_tokens.

TypeSafeModel and TYPESAFE_MODELS track the jev-latest and jev-preview aliases. A pinned model ID is also accepted through the constructor or the per-request model= argument. Responses identify the model that answered; aliases can move between releases.