Enterprise use case and target architecture: request walkthrough

Customer request

1. Client and regional API Gateway

   Enter through a tenant-approved regional boundary; caller scope is verified.

2. Identity/Policy Service

   Current tenant, user and resource policy; mandatory disclosure audit.

3. Admission and fair budgets

   Bound waiting work, deadlines, request/token quotas and active slots.

4. AI App Service

   Run the supplier task and preserve its scoped operation identity.

5. Authorised retrieval

   Resolve current permitted source passages before model disclosure.

6. Model Gateway and approved serving

   Pin the approved bundle/profile; hosted route or justified self-hosted/GPU option.

7. Durable operations and result/history

   Save known completion and original evidence; inspect incomplete/unknown outcomes.

Retrieval reads the source/index state below. Model Gateway uses the approved release below. Result commit precedes saved success.

Enterprise document ingestion

1. Approved enterprise sources

   Tenant-approved source scope and versioned change/delete events.

2. Document Ingestion, Queue and Workers

   Durable jobs prepare bounded parsing/chunking and publish progress.

3. Immutable source objects and bounded cache

   Keep versioned passages and provenance; cache reuse still needs current access.

4. Metadata DB and Search Index

   Current versions, tombstones, access metadata and published searchable passages.

Publishes the current source, version and search state used by authorised retrieval; old events cannot undo a newer deletion.

Quality and model lifecycle

1. Optional datasets and Fine-Tuning Jobs

   Separately approved data, pinned recipe/base, experiments and verified checkpoints.

2. Model Registry

   Immutable candidate lineage; route-friendly aliases do not replace version evidence.

3. Evaluation Service and trusted approval

   Frozen task/safety cases and candidate-bound evidence before promotion.

4. Versioned release bundle

   Bind prompt, retrieval, model and policy configuration; eligible canary and rollback.

Optional training and registry feed evaluation; an existing approved model needs no new training job. Evaluation and approval supply the immutable release identity used by Model Gateway.

Platform, region and operations

• Versioned app manifests and migration controller

  Named app owner, validated desired state, read-only shadow and controlled cutover.

• Placement and regional recovery

  A/B EU cells; C US-home; one writer and verified recovery eligibility.

• Serving and capacity operations

  Measure ready replicas/GPU options and reserve actual regional dependency headroom.

• Audit, Telemetry and usage evidence

  Trace decisions, protect logs, inspect unknown effects and count all workflow spend.

Configures eligible regional execution and records evidence. Global control does not process customer prompts.