rskit_inference/lib.rs
1//! Model-serving runtime adapter layer for rskit.
2//!
3//! `rskit-inference` owns runtime-neutral prediction contracts for Triton, vLLM raw, TGI, KServe v2,
4//! BentoML, ONNX Runtime Server, TFServing, and custom REST/gRPC serving runtimes.
5//! It intentionally does not model chat completions; chat belongs to `rskit-llm`
6//! and provider implementations under `rskit-llm-providers`.
7//!
8//! Adapters declare their executable authority through `rskit_tool::Envelope` on [`InferenceDescriptor`].
9//! Consumers
10//! and orchestrators enforce envelope intersection with the five-stage permission model (declaration, registry verification, activation, per-invocation enforcement, HITL elicitation).
11
12#![warn(missing_docs)]
13
14/// Lean default echo adapter.
15pub mod echo;
16/// Model-serving inference traits.
17pub mod inference;
18/// Explicit adapter registry.
19pub mod registry;
20/// Runtime-neutral request, response, value, descriptor, and error types.
21pub mod types;
22
23pub use echo::{Echo, register as register_echo};
24pub use inference::{Inference, StreamingInference};
25pub use registry::{Factory, Registry, RegistryError, default_registry};
26pub use rskit_ai::{Model, StreamEvent, StreamEventRef, Usage};
27pub use types::{
28 InferenceDescriptor, InferenceError, PredictRequest, PredictResponse, PredictStatus,
29 ServingProtocol, Tensor, TensorData, Value,
30};