jevper

Jevper offers a robust and type-safe classification wrapper built for OpenAI-compatible models, uniquely providing probabilistic outputs and confidence scores instead of mere text responses. This library serves as an interface between Python programs and hosted language models, enabling queries that yield structured answers while preserving the power of LLMs.

Key Features

• Typed Questions and Probabilities: Inputs are structured as state with typed output responses such as noul, choice, and score, which include clear probabilities and confidence levels.
• Independence from TypeSafe API: Unlike other implementations, jevper operates independently from the hosted TypeSafe API and can interface seamlessly with any model exposing the required methods, including self-hosted models like llama.cpp.
• Concurrent Question Handling: Each question is treated as an independent call, allowing parallel execution and efficient processing. The library defaults to a maximum concurrency of 8 but can be adjusted as needed.

Usage Example

Here’s a brief demonstration of how to use jevper with an OpenAI model:

from openai import OpenAI
from jevper import Choice, SystemOneClient

client = SystemOneClient(OpenAI(), model="gpt-5.6-terra", method="logprobs")

response = client.system_one(
    state="I was charged twice for the same subscription this month.",
    questions={
        "intent": Choice(
            instructions="Pick the intent of the message.",
            criteria={
                "billing": "money, invoices, refunds, charges",
                "technical": "errors, crashes, login or performance problems",
                "sales": "pricing, plans, purchasing, upgrades",
            },
        )
    },
)

answer = response.answers["intent"]
print(answer.choice)        # Outputs: "billing"
print(answer.probabilities) # Outputs: {"billing": 0.88, "technical": 0.08, "sales": 0.03}
print(answer.confidence)    # Outputs: 0.83

Classification Types

jevper supports three classification types that correspond to the Jev API:

• Noul: Outputs a yes/no response with an associated probability.
• Choice: Selects from multiple predefined options with detailed probabilities and confidence metrics.
• Score: Provides ratings on a scale with associated legend and probabilities.

Flexible Methods of Analysis

The library features various methods for decision-making, including:

• auto: Automatically decides the best method between logprobs and structured JSON based on provider capabilities.
• logprobs: Fetches probability distributions over labels for nuanced response generation.
• grammar: Allows for GBNF grammar structures in requesting data from the provider.

Verification and Testing

jevper includes comprehensive testing features, ensuring functionality against local stub servers and easy verification of responses. The codebase is maintained with strict testing protocols, ensuring reliability in production environments.

Documentation

Further documentation is available, covering API specifications, method details, reasoning configurations, and examples. More insights can be found in the API Docs, Methods Guide, and other relevant sections of the repository.