Create Public Alias

Use this form to create a stable, environment-specific model name and the policy contract that every route behind that name must satisfy. Applications use the Alias Name instead of a provider Deployment or physical model ID.

Despite the entity name, an Alias can be generally available (PUBLIC) or restricted to an agent or workload identity. Add its Routes, Credential, and Pricing records, then use Publish to review the complete configuration.

Important: operations, requiredCapabilities.embeddingSpace, requireExpectedEmbeddingSpace, and embeddingWorkloadLane form an immutable routing contract. To change any of them, create a new Alias revision instead of updating the existing Alias.

Fields

Operations and Required Capabilities are structured editors. After changing their JSON or YAML value, choose Apply before submitting the form.

Visibility and identity rules

Use exactly one of these shapes:

A non-standard Knowledge Base lane (kb_query or kb_index) additionally requires all of the following:

• operations is exactly ["embed"];
• Require Expected Embedding Space is enabled; and
• Alias Visibility is INTERNAL_WORKLOAD with a non-empty workload principal.

The Alias names kb-query and kb-index use hyphens. The corresponding lane identifiers kb_query and kb_index use underscores. They are distinct contracts and must not be substituted for each other.

Embedding space

An embedding Alias must declare exactly these six fields under requiredCapabilities.embeddingSpace:

Matching dimensions alone do not make two embedding spaces compatible. Both Knowledge Base Aliases and every eligible primary or fallback Deployment must publish exactly the same six-field contract.

NVIDIA Nemotron Knowledge Base example

For nvidia/nemotron-3-embed-1b, NVIDIA currently documents a native 2048-dimensional float embedding, a 4096-token NIM limit, and no reduced dimension support. NVIDIA’s model card describes the output as L2-normalized, so the example uses L2 normalization with cosine distance. See the NVIDIA NIM support matrix and NVIDIA model card.

Create the indexing Alias first:

{
  "environment": "loc",
  "aliasName": "kb-index",
  "operations": ["embed"],
  "requiredCapabilities": {
    "embeddingSpace": {
      "spaceId": "nvidia-nemotron-3-embed-1b-float-v1",
      "revision": 1,
      "dimension": 2048,
      "normalization": "l2",
      "distanceMetric": "cosine",
      "documentInputTransformVersion": "document-v1"
    }
  },
  "requireExpectedEmbeddingSpace": true,
  "embeddingWorkloadLane": "kb_index",
  "maxInputTokens": 4096,
  "maxRequestBytes": 1048576,
  "dataClassification": "public",
  "loggingMode": "METADATA",
  "piiMode": "DENY",
  "aliasVisibility": "INTERNAL_WORKLOAD",
  "boundWorkloadPrincipal": "knowledge-indexer"
}

Then create the query Alias with the same embedding-space object and policy, changing only:

{
  "aliasName": "kb-query",
  "embeddingWorkloadLane": "kb_query",
  "boundWorkloadPrincipal": "knowledge-service"
}

The two workload-principal examples assume that the indexing and query bearer tokens resolve to knowledge-indexer and knowledge-service. If your tokens use different subjects, enter those exact resolved principal IDs instead.

NVIDIA retrieval models distinguish document (passage) input from query input. The Alias records the immutable vector-space and document-transform identity, but provider-specific request transformation remains a Deployment and gateway responsibility. Verify that the qualified provider path applies the corresponding passage and query behavior before activating the Routes.

General generation example

For a generally available generation Alias:

{
  "environment": "prod",
  "aliasName": "governed-chat",
  "operations": ["generate"],
  "requiredCapabilities": {
    "tools": true,
    "streaming": true
  },
  "maxInputTokens": 128000,
  "maxOutputTokens": 8192,
  "maxRequestBytes": 1048576,
  "dataClassification": "internal",
  "loggingMode": "METADATA",
  "piiMode": "REDACT",
  "aliasVisibility": "PUBLIC"
}

The backend creates the Public Alias Id and aggregate version. The active state is backend-managed through soft delete and is not part of this form.