Concepts

Uni-Xervo is organized around four core concepts:

• Model catalog: declarative alias-to-provider mapping.
• Runtime loading: safe lazy/eager/background model initialization.
• Reliability: timeout, retry, and remote circuit breaker controls.
• Capability-driven APIs: 13 task kinds with typed trait handles — dense embed, rerank, generate, raw tensor, sparse embed, multi-vector (ColBERT) embed, image/audio/multimodal embed, NLP, document extract, transcribe, and OCR. Every task trait shares the ModelInfo supertrait (model_id(), active_execution_providers()).

Request lifecycle

1. App asks for alias handle (runtime.embedding("embed/default") or runtime.raw_tensor_model("raw/model")).
2. Runtime resolves alias to ModelAliasSpec.
3. Runtime computes a ModelRuntimeKey (task + provider + model + revision + options hash).
4. Existing loaded instance is reused, or provider load is coordinated under a per-key mutex.
5. Returned handle is wrapped with instrumentation (metrics + timeout/retry).

Why this matters

• Application code stays provider-agnostic.
• Equivalent aliases share loaded instances.
• Cold start and failure handling are explicit and configurable.