Jevper is an independent implementation of OpenAI's Decisions API schema and the TypeSafe System One wire format, designed to provide a consistent interface for structured outputs from any OpenAI-compatible language model. As described in its documentation, jevper is not affiliated with or endorsed by OpenAI or TypeSafe AI, but it implements their documented schemas to enable structured classification over text and image inputs.
This image shows the Jevper project's visual identity, which corresponds to the description of Jevper as a unified interface for structured LLM outputs across models and modalities.
The wrapper accepts a shared input—either text or images—and processes three types of questions: predicate (yes/no), choice (multiple choice), and score (scale). For each question, jevper returns a typed answer that includes the selected option, associated probabilities or scores, and a confidence value. This allows developers to obtain structured outputs such as intent classification or sentiment scoring without needing to customize prompts for each model or provider, as demonstrated in the example where jevper wraps an OpenAI client to classify user intent.
To ensure compatibility across diverse LLM backends, jevper employs an adaptive method selection system. By default, it attempts to use logprobs where available to access the model's real token distribution, falling back to structured JSON output when logprobs are not supported. For choice questions with many options, jevper automatically switches to methods that use JSON schema to handle up to 255 options, and it supports few-shot examples and reasoning configurations to improve accuracy and capture the model's analysis process, as detailed in the documentation on methods and reasoning.