Before you dive in
What you’ll find in the original
- Send the documented model, state, and questions object to `/v1/systemone` rather than an ordinary chat-completions route.
- Use the same noul, choice, and score schema across the wrapper and compatible clients.
- Run the repository's comparison benchmark before relying on the wrapper, because API-shape compatibility does not establish calibration or quality.
Worth knowing
The live Vercel service and Upstage model are external dependencies. Benchmark results are repository-reported snapshots, and a prompted LLM wrapper is not the same model or training method as Jev.