Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.aperium.apps.hillspire.com/llms.txt

Use this file to discover all available pages before exploring further.

The gates below are the last line of defence before declaring an on-prem deployment ready. Treat each one as blocking — a release that skips a gate is a release that has not been validated.

​
Pre-production validation gates

All images are mirrored to the private registry and deployed by digest.
Required secrets are materialized in-cluster.
Database migrations complete successfully.
Backend starts with MULTI_POD_ENABLED=true and shared-storage validation passes.
Every deployed aperium-mcp-<connector> service is healthy and ready.
Backend debug status reports each enabled connector’s transport as http.
Local LLM health check passes through the same local-provider configuration used by chat.
For every enabled connector, a representative read workflow succeeds end to end from the UI.
For every enabled connector, a representative write workflow succeeds in a non-production upstream environment.
A no-capability-routing prompt still exposes or hydrates the expected tool inventory for every enabled connector.

​
Production readiness gates

Cloud LLM fallback is disabled or formally approved for your environment.
Every enabled connector is configured only through HTTP MCP transport in the deployment overlay.
NetworkPolicy blocks external access to every aperium-mcp-<connector> service and to the local model endpoints.
Backups and restore procedure have been tested.
Release rollback is documented as redeploying a prior approved version, not silently switching to stdio transport or external LLM providers.

​
Open decisions

The following items are left open in the on-prem requirements contract and must be settled before signing off:
  • Which internal OpenAI-compatible model-serving endpoint will own the primary model traffic (for example, gemma-4).
  • Whether file storage remains RWX local storage or moves to an object-store replacement.
  • Whether Qdrant and Phoenix are mandatory in the initial on-prem release or disabled until needed.
  • The final write-tool allowlist and upstream service-account permissions for every enabled connector.
  • Local model SLOs for latency, context length, concurrency, and tool-call accuracy.