Model Discovery¶

Providers can list the models they serve at runtime through IModelDiscoveryProvider:

public interface IModelDiscoveryProvider
{
    bool SupportsModelDiscovery { get; }

    Task<IReadOnlyList<AIModel>> GetModelsAsync(
        CancellationToken cancellationToken = default);
}

All four built-in provider implementations implement the interface, but whether discovery actually works depends on the catalog definition: SupportsModelDiscovery mirrors ProviderDefinition.HasModelDiscoveryApi. When consumer-supplied model overrides are registered for a provider (see Consumer-Supplied Overrides below), the decorator also returns true — even if the inner provider has no discovery API.

Usage Pattern¶

Always check the flag first — calling GetModelsAsync on a provider without a discovery API throws AiException with code ai/model-discovery-not-supported:

if (provider is IModelDiscoveryProvider discovery
    && discovery.SupportsModelDiscovery)
{
    IReadOnlyList<AIModel> models = await discovery.GetModelsAsync();
}

AIModel¶

Property	Type	Description
`Id`	`string`	Model ID to pass to `ChatCompletionRequest.Model`
`DisplayName`	`string`	Human-readable name
`ProviderId`	`string`	Owning provider's catalog ID
`OwnedBy`	`string?`	Organization that owns the model
`Description`	`string?`	Model description, when the API provides one
`ContextWindow`	`int?`	Maximum context size in tokens
`Capabilities`	`EModelCapability`	Capability flags (see below)
`Modality`	`string?`	Input/output modalities, e.g. `"text+image->text"`
`PromptPrice`	`decimal?`	Price per million prompt tokens
`CompletionPrice`	`decimal?`	Price per million completion tokens
`PriceUnit`	`EModelPriceUnit`	Pricing unit, default `EModelPriceUnit.Per1M`

Not every provider API returns all of these — discovery fills what the provider exposes and leaves the rest at defaults.

EModelCapability¶

Capabilities is a [Flags] enum covering text and multimodal abilities:

Text & core: TextGeneration, StructuredOutput, ToolCalling, Embedding, Reranker
Vision: ImageRecognition, ImageGeneration
Audio: AudioRecognition, TextToSpeech, AudioGeneration
Video: VideoTranscription, VideoRecognition, VideoGeneration

bool supportsTools = model.Capabilities.HasFlag(EModelCapability.ToolCalling);

Discovery per Protocol¶

Protocol	Endpoint	Notes
OpenAI-compatible / HybridGateway	`GET {BaseUrl}{ModelsEndpoint}`	Parses the `data` array
Messages API	`GET {BaseUrl}{ModelsEndpoint}`	Parses the `data` array
KeyQuery	`GET {BaseUrl}models`	Authenticates with `x-goog-api-key`; strips the `models/` prefix from names
Catalog	`GET {BaseUrl}models`	Maps `id`, `name`, `publisher`

Dynamic Catalogs (Localhost Providers)¶

Self-hosted providers (Ollama, Jan, LM Studio, LocalAI, vLLM, TGI, TextGenWebUI, FastChat) have IsDynamicModelCatalog = true. Their model catalog depends on what the user has downloaded or loaded locally — no external method (API, web search, or page scraping) can determine the available model count.

SupportsModelDiscovery is false for these providers. Use ProviderCatalog.WithDynamicCatalog to retrieve the full list of providers with dynamic catalogs.

Consumer-Supplied Overrides¶

Many providers have no /models endpoint, and those that do often omit pricing or return stale data. You can supply model lists and token prices from your own source and have them applied transparently over the existing GetModelsAsync() — the call site does not change, and providers without overrides behave exactly as before.

Overrides are registered on the DI builder with OverrideModels. Each ModelOverride is a patch with all-nullable fields: only the fields you set are applied, so you never clobber a live value you did not intend to change. The Id field matches an existing AIModel.Id (case-insensitively); an override whose Id matches nothing is added as a new model.

Enrich live models with prices¶

services.AddAiProviders(o => o.OverrideModels("openai", new[]
{
    new ModelOverride { Id = "gpt-4o", PromptPrice = 2.50m, CompletionPrice = 10.00m },
}));

// Same call site as before — live list merged with your prices.
IReadOnlyList<AIModel> models = await discovery.GetModelsAsync();

Supply a full list for a provider with no API¶

services.AddAiProviders(o => o.OverrideModels("edenai", MyCatalog.Load("edenai")));

When the inner provider has no discovery API (or throws ai/model-discovery-not-supported), the base list is treated as empty and only your overrides are returned. Any other live-discovery error (network, auth) propagates to the caller.

Correct a field¶

services.AddAiProviders(o => o.OverrideModels("openai", new[]
{
    new ModelOverride { Id = "gpt-4o", ContextWindow = 128000, OwnedBy = "OpenAI" },
}));

Hide a model¶

services.AddAiProviders(o => o.OverrideModels("openai", new[]
{
    new ModelOverride { Id = "gpt-3.5-turbo", Hidden = true },
}));

Custom override source¶

By default overrides are held in memory. Implement IModelOverrideStore to back them with your own source (JSON, database, remote config):

services.AddAiProviders(o => o.UseModelOverrideStore(new JsonModelOverrideStore("models.json")));

Merging preserves the live order, patches matched models in place, and appends newly materialized models last.