JevT++ documentation
JevT++ maps runtime text, JSON and serialized application objects to your C++ enums, boolean decisions and rubric scores. The vocabulary is defined at compile time; the input changes on every call.
These pages track main, including unreleased hardening changes. For a pinned
release, use the documentation in that release’s Git tag and review the
changelog.
Start with your integration
Section titled “Start with your integration”| Need | Start here |
|---|---|
| Try the API without model downloads | Core quickstart |
| Run a local ONNX model | Laya ONNX setup |
| Use original weights with ggml/CUDA | Native laya.cpp backend |
| Use hosted inference with explicit credentials | Remote backend |
| Inject JSON and ask several typed questions | System One API |
| Integrate with a service or event loop | Concurrency and Asio |
| Test decisions without inference or network access | Scripted testing backend |
| Inspect p95 and recent outcomes | Observability |
| Evaluate latency or release readiness | Measurements, production checklist |
What runs today
Section titled “What runs today”- C++20 core with explicit success, abstention and technical-error paths.
- Optional Laya backend using ONNX Runtime CPU or explicit CUDA and native tokenization.
- Optional native laya.cpp/ggml backend using safetensors, CPU or explicit CUDA.
- Optional remote HTTP backend with injected transports or libcurl.
- Multi-field batch inference with metadata and shared input on either backend.
- Diagnostics snapshots, Prometheus export and an optional loopback HTTP view.
An optional bounded worker pool provides owned future submission,
cross-request microbatching and length buckets. The optional
Boost.Asio adapter supports completion tokens and
co_await. Deadlines and stop tokens are checked at supported boundaries;
they cannot forcibly interrupt running local model inference.
JevT++ is independent of TypeSafe. Its open-source Laya model is not the proprietary Jev model. Type-safe output does not guarantee a correct business decision: evaluate quality and abstention on your domain.
Useful links
Section titled “Useful links”Repository and examples · Releases · Report an issue
The core is MIT-licensed. Model artifacts have separate provenance and licenses documented in Laya ONNX setup and the native backend guide.