Shared Services Architecture & Health

Verified integration contracts, timeout behavior, and fallback mechanisms for shared enterprise services.

Shared AI Platform Service

LOCAL_ADAPTER

Provides embeddings, vector indexing (Qdrant), RAG retrieval, and LLM-as-a-judge evaluation suite.

http://localhost:5000

Verified API Contracts

• POST /api/v1/rag/query — Vector similarity search with multi-tenant filtering
• POST /api/v1/ai/evaluation/run — Automated LLM-as-a-judge quality suite
• POST /api/v1/rag/documents — Document ingestion & chunking
• GET /health — Liveness probe

Authentication & Resilient Fallback

Auth: Header: x-api-key or Authorization: Bearer <token>

Resilience: When unreachable or unconfigured, the platform automatically switches to a deterministic local adapter yielding reproducible chunks, citations, and scores, tagged visibly as LOCAL_DETERMINISTIC_ADAPTER.

Model Regression Detection Service

LOCAL_ENGINE

Evaluates candidate models or RAG pipelines against accepted baselines to detect metric regressions.

http://localhost:4002

Verified API Contracts

• POST /api/v1/regression/check — PR & pipeline regression decision contract
• GET /api/v1/baselines — Historical active baselines catalogue
• GET /health — Service liveness probe

Authentication & Resilient Fallback

Auth: Header: x-api-key or Authorization: Bearer <token>

Resilience: When offline, local deterministic regression engine evaluates candidate deltas against baseline thresholds (5% drop flags regression), labeled as LOCAL_DETERMINISTIC_ENGINE.

Shared Notification Service

LOCAL_LOGGING

Dispatches high-priority email and push notifications for regression alerts, stale knowledge warnings, and reviewed document proposals.

http://localhost:3001

Verified API Contracts

• POST /v1/notifications — Multi-channel notification ingestion
• GET /health — Service liveness probe

Authentication & Resilient Fallback

Auth: Header: x-api-key: NOTIFICATION_RAG_OPS_API_KEY

Resilience: Operates with resilient try/catch timeouts; if notification dispatch fails, structured logs are recorded without interrupting benchmark execution or crashing the server.