This module owns the shared model runtime and the chat, embedding, and reranking client families.
| Type | Scope |
|---|---|
unit/ |
Configuration, normalization, routing, lifecycle, cache, telemetry, and client behavior with injected transports |
contract/ |
Reusable chat backend conformance harness |
chaos/ |
Deterministic provider, Redis-lock, circuit, and cancellation fault injection |
performance/ |
Correctness-oriented local concurrency and single-flight load |
smoke/ |
Small real chat, embedding, and reranking provider requests |
Run deterministic model coverage with:
python -m pytest packages/harborrag-adapters/tests/models
Live requests can consume paid quota and need provider-specific models, credentials, endpoints, and optional LiteLLM configuration. Follow the complete model smoke setup first.