Models Block

Configuring dynamic model listings.

The models block (available inside defaults blocks only in v0.1) controls how ONR queries and parses the GET /v1/models endpoint of an upstream provider to populate its internal registry or Admin Page proxy.

Global reusable presets can now be declared in config/modes/models_modes.conf and included from config/onr.conf:

models_mode "openai" {
  path "/v2/models";
}

provider "demo" {
  defaults {
    models {
      models_mode openai;
    }
  }
}

Built-In Handlers

You can rely on built-in parser presets for standard providers. In this repository’s default config, the same names are declared in config/modes/models_modes.conf as global models_mode presets:

models {
  models_mode openai;    # Default query path: /v1/models
}

models {
  models_mode gemini;    # Default query path: /v1beta/models
}

models {
  models_mode vertex;    # Default query path: /v1beta1/publishers/google/models
}

The vertex preset lists Google publisher base models. It requires credential metadata from the selected upstream key. Configure credential_file and location in keys.yaml; use base_url_override for regional Vertex hosts.

Custom Model Parsers

For providers exposing bizarre schemas or prefixed namespaces, build a custom parser to extract and clean the model lists explicitly.

models {
  models_mode custom;
  method GET;
  path "/v1beta/models";
  
  # Locate the array of models
  id_path "$.models[*].name";
  
  # Perform a regex rewrite on the returned model names
  # (e.g., stripping off 'models/' prefix from the string)
  id_regex "^models/(.+)$";
  
  # Only allow models out that begin with "gemini-"
  id_allow_regex "^gemini-";
}

Directives specific to models_mode custom

• id_path (repeatable): Extracts variables into a deduplicated union list.
• id_regex: If it includes a Capture Group, group 1 becomes the final ID. Otherwise, the full Regex Match is used. Non-matching values are dropped.