google.vertex
Package
Section titled “Package”@ai-sdk/google-vertex
Factories
Section titled “Factories”AIProviders.googleVertex
Capabilities
Section titled “Capabilities”- Language
- Embedding
- Image
- Transcription
- Speech
- Video
Current Vertex behavior
Section titled “Current Vertex behavior”Vertex inherits the Google structured JSON plus function-tool handling, error/denial envelopes and incremental JSON response parsing. Generic embedding dimensions are forwarded when provider-specific output dimensions are absent. nano-banana-2.1 selects the generateContent image path and preserves binary or gs:// image inputs, IMAGE response modality, typed aspect ratio and provider imageConfig, and returned usage/metadata. Masks and multiple image inputs are rejected. Future model IDs remain open strings. Vertex’s Anthropic adapter recognizes the published Haiku 5.5 model family through its shared native implementation.
Gemini long decode
Section titled “Gemini long decode”Vertex Gemini shares GenerateContent automatic CONTINUATION handling with the Google provider, including appended content, one stream lifecycle, summed billing usage and modality details, and the first request’s initialPromptTokenCount metadata. MAX_TOKENS and failed streaming continuation expose a resumable continuationToken under providerMetadata["google"]; pass it using the supported vertex, googleVertex, or google provider-options namespace with the original prompt and a larger maxOutputTokens.
Automatically continued calls also expose groundingMetadataPerRequest under providerMetadata["google"], preserving grounding from each successful request in order and null for requests without it. Generation’s existing groundingMetadata reflects the final response; streaming retains the latest non-null payload. Calls without automatic continuation omit the per-request array.
Explicit thinking budgets or levels override shared reasoning settings. Dotted routed Gemini IDs retain their model capabilities, and Interactions model requests with background: true poll until terminal.
Vertex’s implemented Gemini language, tuned endpoint, embedding, Gemini TTS speech, Cloud Speech-to-Text transcription, Veo video, and Interactions surface was audited against upstream Google Vertex 5.0.122. The current pass exposes the published embedding limits: 250 inputs for standard models and one input for Gemini multimodal embeddings. Imagen image generation/editing factories remain locally available for source compatibility after their removal from the current upstream package. Shared Gemini request behavior includes local JSON Schema $ref expansion with boolean-schema references, the Gemini 3.7 Flash low thinking floor, forward-compatible future-model capabilities, and response IDs; standard Vertex function-call and function-response IDs are omitted while server-tool IDs remain intact. Video supports first/last frame images, reference images, and top-level generateAudio precedence. Speech routes chirp* model IDs to authenticated Cloud TTS text:synthesize with Chirp voice/language defaults, LINEAR16/WAV handling, speed mapping, warnings, and metadata; other IDs keep Gemini TTS generateContent with WAV-wrapped PCM by default, raw PCM opt-in, multi-speaker provider options, sample-rate metadata, and upstream warnings. Google Batch, Gemini 3.5 transcription, and the broader shared Google language/safety/usage/request changes remain deferred to a dedicated Vertex pass. Node google-auth-library options, workflow serialization, ProviderV4 type names, and ESM-only packaging remain JS-runtime/package-surface concerns. The 5.0.76 pass adds an exact Maverick/Scout MaaS default of max_tokens: 8192 without overwriting caller values and inherits the audited Google, Anthropic, and OpenAI-compatible shared behavior. The 2026-09-27 pass adds gs:// tool-result media and the shared current TTS behavior. The 2026-10-04 pass validates generated location/Chirp region hosts, reports Gemini image capabilities, and retains speech usageMetadata; Vertex Gemini transcription remains deferred. The 2026-10-11 pass inherits JSON-with-tools conversion and shared embedding dimensions; native Gemini transcription remains a declared gap. The 2026-10-18 pass: Inherited GenerateContent continuation/usage/AUDIO/thinking and background Interactions behavior is current. Vertex Anthropic labels are covered by its separate adapter.
Factory argument requirements are defined by the public Swift factory signatures. Use Public symbols when you need the exact initializer or factory declaration.
Return to the provider matrix.