TypeSafe plugin for LiveKit Agents

Support for TypeSafe Jev decision models. Jev evaluates probability questions, named choices, and ordered scores over a conversation in one request.

Installation

pip install livekit-plugins-typesafe

Pre-requisites

Set TYPESAFE_API_KEY for the TypeSafe API, or OPENROUTER_API_KEY to use Jev through OpenRouter's Decisions API.

Usage

from livekit.agents import llm, utils
from livekit.plugins import typesafe

async with utils.http_context.open():
    model = typesafe.DecisionModel(model="jev-latest")
    # or: typesafe.DecisionModel.with_openrouter()
    context = llm.ChatContext.empty()
    context.add_message(role="user", content="Please let me speak to a person.")
    response = await model.evaluate(
        input=context,
        decisions={
            "handoff": llm.Probability("The caller requests a human."),
            "intent": llm.Choice(
                "What does the caller want?",
                options={"booking": "Make a reservation", "other": "Anything else"},
            ),
            "frustration": llm.Score(
                "How frustrated is the caller?", levels=["Calm", "Annoyed", "Angry"]
            ),
        },
    )
    for name, result in response.results.items():
        print(name, result.value)

input also accepts a string, which is sent as-is. Each requested decision appears in either response.results or response.errors. Request failures and malformed responses raise.

Models

jev-latest is the most recent stable release and jev-preview the most recent release overall; TypeSafe moves both aliases as new versions ship. Pin jev-1.14.0 or jev-1.13.0 for a fixed version.

Notes

• Jev rounds probabilities and scores to two decimal places. Results keep the provider's values.
• Score answers must return a legend that matches the requested levels.
• provider_data["confidence"] describes how concentrated the distribution is, not the probability that the answer is correct.
• A conversation is sent as a role: text transcript of user and assistant messages. System and developer messages, agent instructions, and non-text content are never sent.