@llumiverse/drivers/lib/types/vertexai/index.d.ts

LLM driver implementations. Currently supported are: openai, huggingface, bedrock, replicate.

import { AnthropicVertex } from "@anthropic-ai/vertex-sdk";
import { PredictionServiceClient, v1beta1 } from "@google-cloud/aiplatform";
import { type Content, GoogleGenAI, type Model } from "@google/genai";
import { type AIModel, AbstractDriver, type Completion, type CompletionChunkObject, type DriverOptions, type EmbeddingsOptions, type EmbeddingsResult, type ExecutionOptions, type LlumiverseError, type LlumiverseErrorContext, type ModelSearchPayload, type PromptSegment } from "@llumiverse/core";
import { FetchClient } from "@vertesia/api-fetch-client";
import { GoogleAuth, type GoogleAuthOptions } from "google-auth-library";
import { type ImagenPrompt } from "./models/imagen.js";
export interface VertexAIDriverOptions extends DriverOptions {
    project: string;
    region: string;
    googleAuthOptions?: GoogleAuthOptions;
}
export interface GenerateContentPrompt {
    contents: Content[];
    system?: Content;
}
export type VertexAIPrompt = ImagenPrompt | GenerateContentPrompt;
export declare function trimModelName(model: string): string;
export declare class VertexAIDriver extends AbstractDriver<VertexAIDriverOptions, VertexAIPrompt> {
    static PROVIDER: string;
    provider: string;
    aiplatform: v1beta1.ModelServiceClient | undefined;
    anthropicClient: AnthropicVertex | undefined;
    fetchClient: FetchClient | undefined;
    googleGenAI: GoogleGenAI | undefined;
    googleGenAIRegion: string | undefined;
    googleGenAIFlex: boolean | undefined;
    llamaClient: FetchClient & {
        region?: string;
    } | undefined;
    modelGarden: v1beta1.ModelGardenServiceClient | undefined;
    imagenClient: PredictionServiceClient | undefined;
    googleAuth: GoogleAuth<any>;
    private authClientPromise;
    constructor(options: VertexAIDriverOptions);
    private getAuthClient;
    getGoogleGenAIClient(region?: string, flex?: boolean): GoogleGenAI;
    private buildGoogleGenAIClient;
    getFetchClient(): FetchClient;
    getLLamaClient(region?: string): FetchClient;
    getAnthropicClient(region?: string): Promise<AnthropicVertex>;
    getAIPlatformClient(): Promise<v1beta1.ModelServiceClient>;
    getModelGardenClient(): Promise<v1beta1.ModelGardenServiceClient>;
    getImagenClient(): Promise<PredictionServiceClient>;
    validateResult(result: Completion, options: ExecutionOptions): void;
    protected canStream(options: ExecutionOptions): Promise<boolean>;
    protected isImageModel(model: string): boolean;
    createPrompt(segments: PromptSegment[], options: ExecutionOptions): Promise<VertexAIPrompt>;
    requestTextCompletion(prompt: VertexAIPrompt, options: ExecutionOptions): Promise<Completion>;
    requestTextCompletionStream(prompt: VertexAIPrompt, options: ExecutionOptions): Promise<AsyncIterable<CompletionChunkObject>>;
    /**
     * Build conversation context after streaming completion.
     * Reconstructs the assistant message from accumulated results and applies stripping.
     * Handles both Gemini (Content[]) and Claude (ClaudePrompt) formats.
     */
    buildStreamingConversation(prompt: VertexAIPrompt, result: unknown[], toolUse: unknown[] | undefined, options: ExecutionOptions): Content[] | unknown | undefined;
    /**
     * Build conversation for Claude streaming.
     * Creates assistant message with tool_use blocks in Claude's ContentBlock format.
     */
    private buildClaudeStreamingConversation;
    requestImageGeneration(_prompt: ImagenPrompt, _options: ExecutionOptions): Promise<Completion>;
    getGenAIModelsArray(client: GoogleGenAI): Promise<Model[]>;
    listModels(_params?: ModelSearchPayload): Promise<AIModel<string>[]>;
    validateConnection(): Promise<boolean>;
    generateEmbeddings(options: EmbeddingsOptions): Promise<EmbeddingsResult>;
    /**
     * Cleanup Google Cloud clients when the driver is evicted from the cache.
     */
    destroy(): void;
    /**
     * Format VertexAI errors by routing to model-specific error handlers.
     * Each model definition (Gemini, Claude, Llama) can provide custom error parsing
     * based on their specific SDK error structures.
     *
     * @param error - The error from the VertexAI/model SDK
     * @param context - Context about where the error occurred
     * @returns A standardized LlumiverseError
     */
    formatLlumiverseError(error: unknown, context: LlumiverseErrorContext): LlumiverseError;
}
//# sourceMappingURL=index.d.ts.map