Google Vertex language model provider options
Google Vertex models support provider options passed via providerOptions.vertex. Available options: cachedContent (string), structuredOutputs (boolean, default true), safetySettings (array of objects with category and threshold), audioTimestamp (boolean, default false), labels (object), imageConfig (object), streamFunctionCallArguments (boolean, default false), sharedRequestType (string: 'priority', 'flex', 'standard'), requestType (string: 'shared').
Google Vertex safetySettings categories
Safety setting categories: HARM_CATEGORY_UNSPECIFIED, HARM_CATEGORY_HATE_SPEECH, HARM_CATEGORY_DANGEROUS_CONTENT, HARM_CATEGORY_HARASSMENT, HARM_CATEGORY_SEXUALLY_EXPLICIT, HARM_CATEGORY_CIVIC_INTEGRITY.
Google Vertex safetySettings thresholds
Safety setting thresholds: HARM_BLOCK_THRESHOLD_UNSPECIFIED, BLOCK_LOW_AND_ABOVE, BLOCK_MEDIUM_AND_ABOVE, BLOCK_ONLY_HIGH, BLOCK_NONE.
Google Vertex imageConfig aspectRatio options
Image aspect ratios: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.
Google Vertex imageConfig imageSize options
Image sizes: 1K, 2K, 4K. Defaults to 1K.
Google Vertex imageConfig personGeneration options
Person generation options: PERSON_GENERATION_UNSPECIFIED, ALLOW_ALL, ALLOW_ADULT, ALLOW_NONE.
Google Vertex imageConfig prominentPeople options
Prominent people options: PROMINENT_PEOPLE_UNSPECIFIED, ALLOW_PROMINENT_PEOPLE, BLOCK_PROMINENT_PEOPLE. When set together with personGeneration, personGeneration takes precedence.
Google Vertex audioTimestamp option
The audioTimestamp option (default false) enables timestamp understanding for audio files, useful for generating transcripts with accurate timestamps.
Google Vertex streamFunctionCallArguments option
When streamFunctionCallArguments is true, function call arguments will be streamed incrementally in streaming responses, enabling tool-input-delta events to arrive as the model generates function call arguments. Defaults to false. Only supported on Vertex AI API (not Gemini API) with Gemini 3+ models.
Google Vertex sharedRequestType option
The sharedRequestType option selects a pay-as-you-go (PayGo) tier: 'priority' for consistent low-latency performance at premium cost, 'flex' for 50% discount with longer expected latency. Both are supported only on the global endpoint and subset of Gemini models. When not set with Provisioned Throughput allocated, request consumes PT quota first.
Google Vertex thinkingConfig for reasoning tokens
Enable thinking tokens by setting `thinkingConfig: { includeThoughts: true, thinkingBudget: 2048 }` in providerOptions.vertex. When includeThoughts is true, parts marked with thought: true are processed as reasoning. In generateText, these contribute to reasoningText (string) and reasoning (array) fields. In streamText, these are emitted as reasoning stream parts.
Google Vertex embedding model options
Embedding model options: outputDimensionality (number, optional reduced dimension for output embedding), taskType (string: SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING, RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, QUESTION_ANSWERING, FACT_VERIFICATION, CODE_RETRIEVAL_QUERY), title (string, optional document title for RETRIEVAL_DOCUMENT taskType), autoTruncate (boolean, default true, whether to truncate if input exceeds max length).
Google Vertex Imagen model options
Imagen model options: negativePrompt (string), personGeneration (string: allow_adult, allow_all, dont_allow, defaults to allow_adult), safetySetting (string: block_low_and_above, block_medium_and_above, block_only_high, block_none, defaults to block_medium_and_above), addWatermark (boolean, default true), storageUri (string for Cloud Storage URI).
Google Vertex Imagen edit provider options
Edit provider options available under providerOptions.vertex.edit: mode (EDIT_MODE_INPAINT_INSERTION, EDIT_MODE_INPAINT_REMOVAL, EDIT_MODE_OUTPAINT, EDIT_MODE_CONTROLLED_EDITING, EDIT_MODE_PRODUCT_IMAGE, EDIT_MODE_BGSWAP), baseSteps (number 35-75, default higher = better quality but slower), maskMode (MASK_MODE_USER_PROVIDED, MASK_MODE_DEFAULT, MASK_MODE_DETECTION_BOX, MASK_MODE_CLOTHING_AREA, MASK_MODE_PARSED_PERSON), maskDilation (number 0-1, recommended 0.01).
Google Vertex video model options
Video model options: generateAudio (boolean, whether to generate audio with video), personGeneration (string: dont_allow, allow_adult, allow_all), negativePrompt (string), gcsOutputDirectory (string for Cloud Storage URI), referenceImages (array of objects with bytesBase64Encoded or gcsUri), pollIntervalMs (number, polling interval in milliseconds), pollTimeoutMs (number, maximum wait time in milliseconds, recommend 600000ms for longer videos or higher resolutions).
Google Vertex speech model voice parameter
The voice argument accepts one of Gemini's 30 prebuilt voices (e.g. Kore, Puck, Zephyr). Defaults to Kore.
Google Vertex transcription model options
Transcription model options: languageCodes (string[], BCP-47 language codes or ['auto'] to detect, defaults to ['auto'], multiple explicit codes require multi-region endpoint), enableAutomaticPunctuation (boolean, default true), enableWordTimeOffsets (boolean, default true, can reduce transcription quality and speed), region (string, Speech-to-Text region, defaults to provider location).
Google Vertex Node.js provider settings
Optional provider settings for Google Vertex in Node.js (applies to both MaaS and xAI):
- **project**: Google Cloud project ID. Defaults to GOOGLE_VERTEX_PROJECT environment variable.
- **location**: Google Cloud location (e.g. us-east5 or global). Defaults to GOOGLE_VERTEX_LOCATION environment variable or 'global'.
- **googleAuthOptions**: Authentication options for Google Auth Library.
- **headers**: Resolvable<Record<string, string | undefined>> for custom headers.
- **fetch**: Custom fetch implementation.