ai-sdk-ollama/dist/provider.d.ts

Vercel AI SDK Provider for Ollama using official ollama-js library

import { LanguageModelV4, EmbeddingModelV4, RerankingModelV4, ProviderV4 } from '@ai-sdk/provider';
import { Ollama, type Options as OllamaOptions, type ChatRequest, type EmbedRequest, type Config } from 'ollama';
import { z } from 'zod';
import { OllamaRerankingSettings } from './models/reranking-model.js';
import { OllamaEmbeddingRerankingSettings } from './models/embedding-reranking-model.js';
import { ollamaTools } from './ollama-tools.js';
import type { WebSearchToolOptions } from './tool/web-search.js';
import type { WebFetchToolOptions } from './tool/web-fetch.js';
import type { ObjectGenerationOptions } from './utils/object-generation-reliability.js';
export interface Options extends OllamaOptions {
    /**
     * Minimum probability threshold for token selection
     * This parameter is supported by Ollama API but missing from ollama-js TypeScript definitions
     */
    min_p?: number;
}
export type { Ollama, ChatRequest, EmbedRequest, Config, ToolCall, Tool, Message, ChatResponse, EmbedResponse, } from 'ollama';
/**
 * Settings for configuring the Ollama provider.
 * Extends from Ollama's Config type for consistency with the underlying client.
 */
export interface OllamaProviderSettings extends Pick<Config, 'headers' | 'fetch'> {
    /**
     * Base URL for the Ollama API (defaults to http://127.0.0.1:11434)
     * Maps to Config.host in the Ollama client
     */
    baseURL?: string;
    /**
     * Ollama API key for authentication with cloud services.
     * The API key will be set as Authorization: Bearer {apiKey} header.
     */
    apiKey?: string;
    /**
     * Existing Ollama client instance to use instead of creating a new one.
     * When provided, baseURL, headers, and fetch are ignored.
     */
    client?: Ollama;
}
export interface OllamaProvider extends ProviderV4 {
    /**
     * Create a language model instance
     */
    (modelId: string, settings?: OllamaChatSettings): LanguageModelV4;
    /**
     * Create a language model instance with the `chat` method
     */
    chat(modelId: string, settings?: OllamaChatSettings): LanguageModelV4;
    /**
     * Create a language model instance with the `languageModel` method
     */
    languageModel(modelId: string, settings?: OllamaChatSettings): LanguageModelV4;
    /**
     * Create an embedding model instance
     */
    embedding(modelId: string, settings?: OllamaEmbeddingSettings): EmbeddingModelV4;
    /**
     * Create an embedding model instance with the `textEmbedding` method
     */
    textEmbedding(modelId: string, settings?: OllamaEmbeddingSettings): EmbeddingModelV4;
    /**
     * Create an embedding model instance with the `textEmbeddingModel` method
     */
    textEmbeddingModel(modelId: string, settings?: OllamaEmbeddingSettings): EmbeddingModelV4;
    /**
     * Create a reranking model instance
     */
    reranking(modelId: string, settings?: OllamaRerankingSettings): RerankingModelV4;
    /**
     * Create a reranking model instance with the `rerankingModel` method
     *
     * NOTE: This uses Ollama's native /api/rerank endpoint which is NOT YET AVAILABLE.
     * Use `embeddingReranking()` for a working solution.
     * @see https://github.com/ollama/ollama/pull/11389
     */
    rerankingModel(modelId: string, settings?: OllamaRerankingSettings): RerankingModelV4;
    /**
     * Create an embedding-based reranking model (RECOMMENDED - working now)
     *
     * This is a workaround that uses embedding similarity for reranking
     * since Ollama doesn't have native reranking support yet.
     *
     * @param modelId - The embedding model to use (e.g., 'bge-m3', 'nomic-embed-text')
     * @param settings - Optional settings for the reranking model
     *
     * @example
     * ```ts
     * const result = await rerank({
     *   model: ollama.embeddingReranking('bge-m3'),
     *   query: 'What is machine learning?',
     *   documents: [...],
     *   topN: 3,
     * });
     * ```
     */
    embeddingReranking(modelId: string, settings?: OllamaEmbeddingRerankingSettings): RerankingModelV4;
    /**
     * Ollama-specific tools that leverage web search capabilities
     */
    tools: {
        webSearch: (options?: WebSearchToolOptions) => ReturnType<typeof ollamaTools.webSearch>;
        webFetch: (options?: WebFetchToolOptions) => ReturnType<typeof ollamaTools.webFetch>;
    };
}
export interface OllamaChatSettings extends Pick<ChatRequest, 'keep_alive' | 'format' | 'tools' | 'think'> {
    /**
     * Additional model parameters - uses extended Options type that includes min_p
     * This automatically includes ALL Ollama parameters including new ones like 'dimensions'
     */
    options?: Partial<Options>;
    /**
     * Enable structured output mode
     */
    structuredOutputs?: boolean;
    /**
     * Enable reliable tool calling with retry and completion mechanisms.
     * Defaults to true whenever function tools are provided; set to false to opt out.
     */
    reliableToolCalling?: boolean;
    /**
     * Tool calling reliability options. These override the sensible defaults used by the
     * built-in reliability layer (maxRetries=2, forceCompletion=true,
     * normalizeParameters=true, validateResults=true).
     */
    toolCallingOptions?: {
        /**
         * Maximum number of retry attempts for tool calls
         */
        maxRetries?: number;
        /**
         * Whether to force completion when tool calls succeed but no final text is generated
         */
        forceCompletion?: boolean;
        /**
         * Whether to normalize parameter names to handle inconsistencies
         */
        normalizeParameters?: boolean;
        /**
         * Whether to validate tool results and attempt recovery
         */
        validateResults?: boolean;
        /**
         * Custom parameter normalization mappings
         */
        parameterMappings?: Record<string, string[]>;
        /**
         * Timeout for tool execution in milliseconds
         */
        toolTimeout?: number;
    };
    /**
     * Enable reliable object generation with retry and repair mechanisms.
     * Defaults to true whenever JSON schemas are used; set to false to opt out.
     */
    reliableObjectGeneration?: boolean;
    /**
     * Object generation reliability options. These override the sensible defaults used by the
     * built-in reliability layer (maxRetries=3, attemptRecovery=true, useFallbacks=true,
     * fixTypeMismatches=true, enableTextRepair=true).
     */
    objectGenerationOptions?: ObjectGenerationOptions;
}
/**
 * Settings for configuring Ollama embedding models.
 * Uses Pick from EmbedRequest for type consistency with the Ollama API.
 */
export interface OllamaEmbeddingSettings extends Pick<EmbedRequest, 'dimensions'> {
    /**
     * Additional embedding parameters (temperature, num_ctx, etc.)
     */
    options?: Partial<Options>;
}
/**
 * Schema for validating Ollama provider options
 */
export declare const ollamaProviderOptionsSchema: z.ZodObject<{
    headers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
}, z.core.$strip>;
/**
 * Options for configuring Ollama provider calls
 */
export type OllamaProviderOptions = z.infer<typeof ollamaProviderOptionsSchema>;
/**
 * Schema for validating Ollama chat provider options
 */
export declare const ollamaChatProviderOptionsSchema: z.ZodObject<{
    headers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
    structuredOutputs: z.ZodOptional<z.ZodBoolean>;
}, z.core.$strip>;
/**
 * Options for configuring Ollama chat model calls
 */
export type OllamaChatProviderOptions = z.infer<typeof ollamaChatProviderOptionsSchema>;
/**
 * Schema for validating Ollama embedding provider options
 */
export declare const ollamaEmbeddingProviderOptionsSchema: z.ZodObject<{
    headers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
    maxEmbeddingsPerCall: z.ZodOptional<z.ZodNumber>;
}, z.core.$strip>;
/**
 * Options for configuring Ollama embedding model calls
 */
export type OllamaEmbeddingProviderOptions = z.infer<typeof ollamaEmbeddingProviderOptionsSchema>;
/**
 * Create an Ollama provider instance
 */
export declare function createOllama(options?: OllamaProviderSettings): OllamaProvider;
/**
 * Default Ollama provider instance
 */
export declare const ollama: OllamaProvider;
//# sourceMappingURL=provider.d.ts.map