Provider Capabilities

Functions for querying and validating provider capabilities.

Each service endpoint declares which API features it supports. Request functions validate capabilities before dispatching, giving clear errors instead of HTTP 404s.

What "validate before dispatch" means

Validation comes in two strengths, and which one applies depends on the verb:

Platform and lifecycle verbs — files, vector stores, conversations, moderations, audio, batch, fine-tuning, containers, uploads, video, realtime, FIM, prefix completion, image edits — validate strictly. An endpoint that declares no capabilities at all is rejected, because these surfaces are OpenAI-shaped and a backend that has not declared them would simply 404.
The four primary verbs — chatrequest! (:chat), embeddingrequest! (:embeddings), respond (:responses or :agentic, since the two agentic wires name the same surface differently), and generate_image (:images) — validate only endpoints that declare their capabilities. An endpoint with no provider_capabilities method passes through unvalidated.

That asymmetry is deliberate. Defining a ServiceEndpoint subtype is the documented way to reach an OpenAI-compatible backend this package does not ship, and such a backend cannot declare anything — refusing to dispatch it would be a false negative about a server the package knows nothing about. "Undeclared" is therefore no applicable method, never an empty capability set. Declaring capabilities is opt-in strictness: once your endpoint declares, the four primary verbs hold it to that declaration like any built-in.

A rejection throws ArgumentError naming the feature and listing what the provider does support.

Functions

UniLM.provider_capabilities — Function

provider_capabilities(service) -> Set{Symbol}

Return the set of capabilities supported by the given service endpoint.

Standard capability symbols include:

Core: :chat, :responses, :agentic, :tools, :streaming, :json_output
Embeddings & images: :embeddings, :images, :image_edits
Completions: :fim, :prefix_completion
Platform APIs: :files, :vector_stores, :conversations, :moderation, :audio, :batch, :fine_tuning, :containers, :uploads, :video, :realtime

source

UniLM.has_capability — Function

has_capability(service, cap::Symbol) -> Bool

Check whether the service endpoint supports a given capability.

source

Capability Symbols

Symbol	Description
`:chat`	Chat Completions API
`:responses`	Responses API
`:agentic`	Agentic `respond` verb (OpenAI Responses / Gemini Interactions)
`:tools`	Tool / function calling
`:streaming`	Server-sent-event token streaming (native providers)
`:json_output`	JSON / structured output mode
`:embeddings`	Embeddings API
`:images`	Image generation
`:image_edits`	Image editing
`:fim`	FIM (fill-in-the-middle) completion
`:prefix_completion`	Prefix completion (continue a partial assistant message)
`:files`	Files API
`:vector_stores`	Vector Stores API
`:conversations`	Conversations API
`:moderation`	Moderations API
`:audio`	Audio (TTS / transcription / translation)
`:batch`	Batch API
`:fine_tuning`	Fine-tuning API
`:containers`	Containers API
`:uploads`	Resumable Uploads API
`:video`	Video generation (Sora)
`:realtime`	Realtime API

Capabilities by Provider

using UniLM

for (name, svc) in [
    ("OpenAI",                 OPENAIServiceEndpoint),
    ("Azure",                  AZUREServiceEndpoint),
    ("Gemini (native)",        GEMINIServiceEndpoint),
    ("Gemini (OpenAI-compat)", GEMINIOpenAIServiceEndpoint),
    ("Anthropic",              ANTHROPICServiceEndpoint),
    ("DeepSeek",               DeepSeekEndpoint("k")),
    ("Generic",                GenericOpenAIEndpoint("http://x", ""))
]
    caps = join(sort(collect(provider_capabilities(svc))), ", ")
    println("$name: $caps")
end

OpenAI: agentic, audio, batch, chat, containers, conversations, embeddings, files, fine_tuning, image_edits, images, json_output, moderation, realtime, responses, tools, uploads, vector_stores, video
Azure: chat, tools
Gemini (native): agentic, chat, streaming, tools
Gemini (OpenAI-compat): chat, embeddings, json_output, tools
Anthropic: chat, json_output, streaming, tools
DeepSeek: chat, fim, json_output, prefix_completion, tools
Generic: chat, embeddings, fim, responses, tools