Deployment Reference#
Use this page while reviewing a deployment or recording its handoff. The public catalog namespace is nvidia; the three application images use the unprefixed NGC tag 1.1.1. The release bundle’s env.example and the tagged Helm values.yaml remain the detailed inventories of supported settings.
Release Images#
Image |
Role |
Release rule |
|---|---|---|
|
Generator backend and APIs |
Pin with the frontend and proxy |
|
React frontend served by nginx |
Pin with the backend and proxy |
|
GitHub source and blueprint intelligence |
Pin with the backend and frontend |
An organization-approved mirror may replace nvcr.io; keep the three images on the same release or approved immutable digests. Phoenix uses a separate pinned upstream or mirror image, and PostgreSQL is a separate database dependency.
The source tag for this documentation is v1.1.1. POC_FACTORY_MCP_VERSION, GITHUB_MCP_PROXY_VERSION, and FRONTEND_VERSION override the shared Compose VERSION; keep them aligned or unset them. Internal registry namespaces and mirror tag conventions are deployment-specific.
Ports and Access#
Component |
Container or service port |
Exposure |
|---|---|---|
Frontend |
8080 |
Public through HTTPS ingress; local Compose maps host port 80 |
Generator backend |
8000 |
Internal behind frontend proxy; local health is available on host port 8000 |
GitHub MCP proxy |
8080 in Compose; 8081 in Helm |
Internal only |
Phoenix |
6006 HTTP; 4317 OTLP |
Optional UI and internal telemetry |
PostgreSQL |
5432 |
Private service or managed database |
Durable State#
State |
Container path |
Operating rule |
|---|---|---|
Generated POCs |
|
Persist and back up with the database |
NAT job store |
|
Persist and restore with POC records |
Fine-tuning artifacts |
|
Persist when customization workflows are used |
Fine-tuning datasets |
|
Persist when datasets must survive restarts |
Logs |
|
Collect centrally for incident review |
PostgreSQL |
Managed service or database volume |
Use database-native backup and restore |
The production Helm example uses one backend replica and enables PVCs for generated POCs, job state, fine-tuning artifacts, and datasets. Compose uses named volumes and host-mounted directories. Verify the actual mounted paths in the rendered deployment.
Configuration Priorities#
Priority |
Values to verify |
|---|---|
Release |
Three POC Factory image tags or digests, registry access, architecture |
Inference |
Base URL, API key, five model roles, user-override policy, optional Cursor key |
Security |
Stable |
State |
PostgreSQL, PVCs or volumes, backups, retention period |
Validation |
Static and test gates, Docker socket policy, deployment and runtime gate settings |
Optional services |
GitHub MCP token, Phoenix, NeMo Microservices endpoints |
The validator acceptance threshold is configured in the selected NAT YAML as functions.code_validator.quality_threshold (default 85), not by a QUALITY_THRESHOLD environment override. If an administrator sets CURSOR_API_KEY, also set MODEL_CURSOR, for example auto.
For procedures, see Docker deployment, Kubernetes deployment, and Configure and validate.