workers-ai-provider 3.1.8 → 3.1.9

package/README.md

@@ -20,7 +20,7 @@ export default {
 		const workersai = createWorkersAI({ binding: env.AI });
 
 		const result = streamText({
-			model: workersai("@cf/meta/llama-4-scout-17b-16e-instruct"),
+			model: workersai("@cf/moonshotai/kimi-k2.5"),
 			messages: [{ role: "user", content: "Write a haiku about Cloudflare" }],
 		});
 
@@ -71,19 +71,20 @@ Browse the full catalog at [developers.cloudflare.com/workers-ai/models](https:/
 
 Some good defaults:
 
-| Task           | Model                                      | Notes                            |
-| -------------- | ------------------------------------------ | -------------------------------- |
-| Chat           | `@cf/meta/llama-4-scout-17b-16e-instruct`  | Fast, strong tool calling        |
-| Chat           | `@cf/meta/llama-3.3-70b-instruct-fp8-fast` | Largest Llama, best quality      |
-| Chat           | `@cf/openai/gpt-oss-120b`                  | OpenAI open-weights, high reason |
-| Reasoning      | `@cf/qwen/qwq-32b`                         | Emits `reasoning_content`        |
-| Embeddings     | `@cf/baai/bge-base-en-v1.5`                | 768-dim, English                 |
-| Embeddings     | `@cf/google/embeddinggemma-300m`           | 100+ languages, by Google        |
-| Images         | `@cf/black-forest-labs/flux-1-schnell`     | Fast image generation            |
-| Transcription  | `@cf/openai/whisper-large-v3-turbo`        | Best accuracy, multilingual      |
-| Transcription  | `@cf/deepgram/nova-3`                      | Fast, high accuracy              |
-| Text-to-Speech | `@cf/deepgram/aura-2-en`                   | Context-aware, natural pacing    |
-| Reranking      | `@cf/baai/bge-reranker-base`               | Fast document reranking          |
+| Task           | Model                                  | Notes                               |
+| -------------- | -------------------------------------- | ----------------------------------- |
+| Chat           | `@cf/moonshotai/kimi-k2.5`             | 256k ctx, tools, vision, reasoning  |
+| Chat           | `@cf/zai-org/glm-4.7-flash`            | Fast, multilingual, 131k ctx        |
+| Chat           | `@cf/openai/gpt-oss-120b`              | OpenAI open-weights, high reasoning |
+| Reasoning      | `@cf/moonshotai/kimi-k2.5`             | Configurable `reasoning_effort`     |
+| Reasoning      | `@cf/qwen/qwq-32b`                     | Emits `reasoning_content`           |
+| Embeddings     | `@cf/baai/bge-base-en-v1.5`            | 768-dim, English                    |
+| Embeddings     | `@cf/google/embeddinggemma-300m`       | 100+ languages, by Google           |
+| Images         | `@cf/black-forest-labs/flux-1-schnell` | Fast, free-tier image generation    |
+| Transcription  | `@cf/openai/whisper-large-v3-turbo`    | Best accuracy, multilingual         |
+| Transcription  | `@cf/deepgram/nova-3`                  | Fast, high accuracy                 |
+| Text-to-Speech | `@cf/deepgram/aura-2-en`               | Context-aware, natural pacing       |
+| Reranking      | `@cf/baai/bge-reranker-base`           | Fast document reranking             |
 
 ## Text Generation
 
@@ -91,7 +92,7 @@ Some good defaults:
 import { generateText } from "ai";
 
 const { text } = await generateText({
-	model: workersai("@cf/meta/llama-3.3-70b-instruct-fp8-fast"),
+	model: workersai("@cf/moonshotai/kimi-k2.5"),
 	prompt: "Explain Workers AI in one paragraph",
 });
 ```
@@ -102,7 +103,7 @@ Streaming:
 import { streamText } from "ai";
 
 const result = streamText({
-	model: workersai("@cf/meta/llama-4-scout-17b-16e-instruct"),
+	model: workersai("@cf/moonshotai/kimi-k2.5"),
 	messages: [{ role: "user", content: "Write a short story" }],
 });
 
@@ -113,13 +114,13 @@ for await (const chunk of result.textStream) {
 
 ## Vision (Image Inputs)
 
-Send images to vision-capable models like Llama 4 Scout and Kimi K2.5:
+Send images to vision-capable models like Kimi K2.5:
 
 ```ts
 import { generateText } from "ai";
 
 const { text } = await generateText({
-	model: workersai("@cf/meta/llama-4-scout-17b-16e-instruct"),
+	model: workersai("@cf/moonshotai/kimi-k2.5"),
 	messages: [
 		{
 			role: "user",
@@ -141,7 +142,7 @@ import { generateText, stepCountIs } from "ai";
 import { z } from "zod";
 
 const { text } = await generateText({
-	model: workersai("@cf/meta/llama-4-scout-17b-16e-instruct"),
+	model: workersai("@cf/moonshotai/kimi-k2.5"),
 	prompt: "What's the weather in London?",
 	tools: {
 		getWeather: {
@@ -161,7 +162,7 @@ import { generateText, Output } from "ai";
 import { z } from "zod";
 
 const { output } = await generateText({
-	model: workersai("@cf/meta/llama-3.3-70b-instruct-fp8-fast"),
+	model: workersai("@cf/moonshotai/kimi-k2.5"),
 	prompt: "Recipe for spaghetti bolognese",
 	output: Output.object({
 		schema: z.object({
@@ -313,7 +314,7 @@ Streaming works the same way — use `streamText` instead of `generateText`.
 Returns a provider with model factories. Each factory accepts an optional second argument for per-model settings:
 
 ```ts
-workersai("@cf/meta/llama-3.3-70b-instruct-fp8-fast", {
+workersai("@cf/moonshotai/kimi-k2.5", {
 	sessionAffinity: "my-unique-session-id",
 });
 ```

package/dist/index.d.mts

@@ -0,0 +1,422 @@
+import { EmbeddingModelV3, EmbeddingModelV3CallOptions, EmbeddingModelV3Result, ImageModelV3, LanguageModelV3, RerankingModelV3, SpeechModelV3, TranscriptionModelV3 } from "@ai-sdk/provider";
+
+//#region src/aisearch-chat-settings.d.ts
+type AISearchChatSettings = {
+  /**
+   * Whether to inject a safety prompt before all conversations.
+   * Defaults to `false`.
+   */
+  safePrompt?: boolean;
+  /**
+   * Passthrough settings that are provided directly to the run function.
+   */
+  [key: string]: unknown;
+};
+//#endregion
+//#region src/workersai-models.d.ts
+/**
+ * The names of the BaseAiTextGeneration models.
+ *
+ * Accepts any string at runtime, but provides autocomplete for known models.
+ */
+type TextGenerationModels = Exclude<value2key<AiModels, BaseAiTextGeneration>, value2key<AiModels, BaseAiTextToImage>> | (string & {});
+type ImageGenerationModels = value2key<AiModels, BaseAiTextToImage> | (string & {});
+/**
+ * The names of the BaseAiTextToEmbeddings models.
+ *
+ * Accepts any string at runtime, but provides autocomplete for known models.
+ */
+type EmbeddingModels = value2key<AiModels, BaseAiTextEmbeddings> | (string & {});
+/**
+ * Workers AI models that support speech-to-text transcription.
+ *
+ * Includes Whisper variants from `@cloudflare/workers-types` plus
+ * Deepgram partner models that may not be in the typed interface yet.
+ * Accepts any string at runtime, but provides autocomplete for known models.
+ */
+type TranscriptionModels = value2key<AiModels, BaseAiAutomaticSpeechRecognition> | "@cf/deepgram/nova-3" | (string & {});
+/**
+ * Workers AI models that support text-to-speech.
+ *
+ * Includes models from `@cloudflare/workers-types` plus Deepgram partner
+ * models that may not be in the typed interface yet.
+ * Accepts any string at runtime, but provides autocomplete for known models.
+ */
+type SpeechModels = value2key<AiModels, BaseAiTextToSpeech> | "@cf/deepgram/aura-1" | "@cf/deepgram/aura-2-en" | "@cf/deepgram/aura-2-es" | (string & {});
+/**
+ * Workers AI models that support reranking.
+ *
+ * Accepts any string at runtime, but provides autocomplete for known models.
+ */
+type RerankingModels = "@cf/baai/bge-reranker-base" | "@cf/baai/bge-reranker-v2-m3" | (string & {});
+type value2key<T, V> = { [K in keyof T]: T[K] extends V ? K : never }[keyof T];
+//#endregion
+//#region src/aisearch-chat-language-model.d.ts
+type AISearchChatConfig = {
+  provider: string;
+  binding: AutoRAG;
+  gateway?: GatewayOptions;
+};
+declare class AISearchChatLanguageModel implements LanguageModelV3 {
+  readonly specificationVersion = "v3";
+  readonly defaultObjectGenerationMode = "json";
+  readonly supportedUrls: Record<string, RegExp[]> | PromiseLike<Record<string, RegExp[]>>;
+  readonly modelId: TextGenerationModels;
+  readonly settings: AISearchChatSettings;
+  private readonly config;
+  constructor(modelId: TextGenerationModels, settings: AISearchChatSettings, config: AISearchChatConfig);
+  get provider(): string;
+  private getWarnings;
+  /**
+   * Build the search query from messages.
+   * Flattens the conversation into a single string for aiSearch.
+   */
+  private buildQuery;
+  doGenerate(options: Parameters<LanguageModelV3["doGenerate"]>[0]): Promise<Awaited<ReturnType<LanguageModelV3["doGenerate"]>>>;
+  doStream(options: Parameters<LanguageModelV3["doStream"]>[0]): Promise<Awaited<ReturnType<LanguageModelV3["doStream"]>>>;
+}
+//#endregion
+//#region src/workersai-embedding-model.d.ts
+type WorkersAIEmbeddingConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions;
+};
+type WorkersAIEmbeddingSettings = {
+  gateway?: GatewayOptions;
+  maxEmbeddingsPerCall?: number;
+  supportsParallelCalls?: boolean;
+  /**
+   * Passthrough settings that are provided directly to the run function.
+   */
+  [key: string]: unknown;
+};
+declare class WorkersAIEmbeddingModel implements EmbeddingModelV3 {
+  readonly specificationVersion = "v3";
+  readonly modelId: EmbeddingModels;
+  private readonly config;
+  private readonly settings;
+  get provider(): string;
+  get maxEmbeddingsPerCall(): number;
+  get supportsParallelCalls(): boolean;
+  constructor(modelId: EmbeddingModels, settings: WorkersAIEmbeddingSettings, config: WorkersAIEmbeddingConfig);
+  doEmbed({
+    values,
+    abortSignal
+  }: EmbeddingModelV3CallOptions): Promise<EmbeddingModelV3Result>;
+}
+//#endregion
+//#region src/workersai-chat-settings.d.ts
+type WorkersAIChatSettings = {
+  /**
+   * Whether to inject a safety prompt before all conversations.
+   * Defaults to `false`.
+   */
+  safePrompt?: boolean;
+  /**
+   * Optionally set Cloudflare AI Gateway options.
+   */
+  gateway?: GatewayOptions;
+  /**
+   * Session affinity key for prefix-cache optimization.
+   * Routes requests with the same key to the same backend replica.
+   */
+  sessionAffinity?: string;
+  /**
+   * Passthrough settings that are provided directly to the run function.
+   * Use this for any provider-specific options not covered by the typed fields.
+   */
+  [key: string]: unknown;
+};
+//#endregion
+//#region src/workersai-chat-language-model.d.ts
+type WorkersAIChatConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions; /** True when using a real Workers AI binding (not the REST shim). */
+  isBinding: boolean;
+};
+declare class WorkersAIChatLanguageModel implements LanguageModelV3 {
+  readonly specificationVersion = "v3";
+  readonly defaultObjectGenerationMode = "json";
+  readonly supportedUrls: Record<string, RegExp[]> | PromiseLike<Record<string, RegExp[]>>;
+  readonly modelId: TextGenerationModels;
+  readonly settings: WorkersAIChatSettings;
+  private readonly config;
+  constructor(modelId: TextGenerationModels, settings: WorkersAIChatSettings, config: WorkersAIChatConfig);
+  get provider(): string;
+  private getArgs;
+  /**
+   * Build the inputs object for `binding.run()`, shared by doGenerate and doStream.
+   *
+   * Images are embedded inline in messages as OpenAI-compatible content
+   * arrays with `image_url` parts. Both the REST API and the binding
+   * accept this format at runtime.
+   *
+   * The binding path additionally normalises null content to empty strings.
+   */
+  private buildRunInputs;
+  /**
+   * Get passthrough options for binding.run() from settings.
+   */
+  private getRunOptions;
+  doGenerate(options: Parameters<LanguageModelV3["doGenerate"]>[0]): Promise<Awaited<ReturnType<LanguageModelV3["doGenerate"]>>>;
+  doStream(options: Parameters<LanguageModelV3["doStream"]>[0]): Promise<Awaited<ReturnType<LanguageModelV3["doStream"]>>>;
+}
+//#endregion
+//#region src/workersai-image-settings.d.ts
+type WorkersAIImageSettings = {
+  maxImagesPerCall?: number;
+};
+//#endregion
+//#region src/workersai-image-model.d.ts
+type WorkersAIImageConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions;
+};
+declare class WorkersAIImageModel implements ImageModelV3 {
+  readonly modelId: ImageGenerationModels;
+  readonly settings: WorkersAIImageSettings;
+  readonly config: WorkersAIImageConfig;
+  readonly specificationVersion = "v3";
+  get maxImagesPerCall(): number;
+  get provider(): string;
+  constructor(modelId: ImageGenerationModels, settings: WorkersAIImageSettings, config: WorkersAIImageConfig);
+  doGenerate({
+    prompt,
+    n,
+    size,
+    aspectRatio,
+    seed,
+    abortSignal
+  }: Parameters<ImageModelV3["doGenerate"]>[0]): Promise<Awaited<ReturnType<ImageModelV3["doGenerate"]>>>;
+}
+//#endregion
+//#region src/workersai-transcription-settings.d.ts
+type WorkersAITranscriptionSettings = {
+  /**
+   * Language of the audio, as an ISO-639-1 code (e.g. "en", "fr").
+   * Only supported by Whisper models. Nova-3 detects language automatically.
+   */
+  language?: string;
+  /**
+   * Initial prompt / context to guide the transcription.
+   * Mapped to `initial_prompt` for Whisper models.
+   */
+  prompt?: string;
+};
+//#endregion
+//#region src/utils.d.ts
+/**
+ * Parameters for configuring the Cloudflare-based AI runner.
+ */
+interface CreateRunConfig {
+  /** Your Cloudflare account identifier. */
+  accountId: string;
+  /** Cloudflare API token/key with appropriate permissions. */
+  apiKey: string;
+}
+//#endregion
+//#region src/workersai-transcription-model.d.ts
+type WorkersAITranscriptionConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions;
+  /**
+   * Whether the binding is a real `env.AI` binding (true) or a REST shim (false).
+   * Nova-3 uses different upload paths depending on this.
+   */
+  isBinding: boolean;
+  /**
+   * REST credentials, only set when `isBinding` is false.
+   * Needed for Nova-3 which requires binary upload, bypassing the JSON-based REST shim.
+   */
+  credentials?: CreateRunConfig;
+};
+/**
+ * Workers AI transcription model implementing the AI SDK's `TranscriptionModelV3` interface.
+ *
+ * Supports:
+ * - Whisper models (`@cf/openai/whisper`, `whisper-tiny-en`, `whisper-large-v3-turbo`)
+ * - Deepgram Nova-3 (`@cf/deepgram/nova-3`) — uses a different input/output format
+ */
+declare class WorkersAITranscriptionModel implements TranscriptionModelV3 {
+  readonly modelId: TranscriptionModels;
+  readonly settings: WorkersAITranscriptionSettings;
+  readonly config: WorkersAITranscriptionConfig;
+  readonly specificationVersion = "v3";
+  get provider(): string;
+  constructor(modelId: TranscriptionModels, settings: WorkersAITranscriptionSettings, config: WorkersAITranscriptionConfig);
+  doGenerate(options: Parameters<TranscriptionModelV3["doGenerate"]>[0]): Promise<Awaited<ReturnType<TranscriptionModelV3["doGenerate"]>>>;
+  private runWhisper;
+  private normalizeWhisperResponse;
+  private runNova3;
+  private normalizeNova3Response;
+}
+//#endregion
+//#region src/workersai-speech-settings.d.ts
+type WorkersAISpeechSettings = {};
+//#endregion
+//#region src/workersai-speech-model.d.ts
+type WorkersAISpeechConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions;
+};
+/**
+ * Workers AI speech (text-to-speech) model implementing the AI SDK's `SpeechModelV3` interface.
+ *
+ * Currently supports Deepgram Aura-1 (`@cf/deepgram/aura-1`).
+ * The model accepts `{ text, voice?, speed? }` and returns raw audio bytes.
+ */
+declare class WorkersAISpeechModel implements SpeechModelV3 {
+  readonly modelId: SpeechModels;
+  readonly settings: WorkersAISpeechSettings;
+  readonly config: WorkersAISpeechConfig;
+  readonly specificationVersion = "v3";
+  get provider(): string;
+  constructor(modelId: SpeechModels, settings: WorkersAISpeechSettings, config: WorkersAISpeechConfig);
+  doGenerate(options: Parameters<SpeechModelV3["doGenerate"]>[0]): Promise<Awaited<ReturnType<SpeechModelV3["doGenerate"]>>>;
+}
+//#endregion
+//#region src/workersai-reranking-settings.d.ts
+type WorkersAIRerankingSettings = {};
+//#endregion
+//#region src/workersai-reranking-model.d.ts
+type WorkersAIRerankingConfig = {
+  provider: string;
+  binding: Ai;
+  gateway?: GatewayOptions;
+};
+/**
+ * Workers AI reranking model implementing the AI SDK's `RerankingModelV3` interface.
+ *
+ * Supports BGE reranker models (`@cf/baai/bge-reranker-base`, `bge-reranker-v2-m3`).
+ *
+ * Workers AI reranking API:
+ * - Input: `{ query, contexts: [{ text }], top_k? }`
+ * - Output: `{ response: [{ id, score }] }`
+ */
+declare class WorkersAIRerankingModel implements RerankingModelV3 {
+  readonly modelId: RerankingModels;
+  readonly settings: WorkersAIRerankingSettings;
+  readonly config: WorkersAIRerankingConfig;
+  readonly specificationVersion = "v3";
+  get provider(): string;
+  constructor(modelId: RerankingModels, settings: WorkersAIRerankingSettings, config: WorkersAIRerankingConfig);
+  doRerank(options: Parameters<RerankingModelV3["doRerank"]>[0]): Promise<Awaited<ReturnType<RerankingModelV3["doRerank"]>>>;
+}
+//#endregion
+//#region src/autorag-chat-language-model.d.ts
+/**
+ * @deprecated Use `AISearchChatLanguageModel` instead. AutoRAG has been renamed to AI Search.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+declare class AutoRAGChatLanguageModel extends AISearchChatLanguageModel {}
+//#endregion
+//#region src/autorag-chat-settings.d.ts
+/**
+ * @deprecated Use `AISearchChatSettings` instead. AutoRAG has been renamed to AI Search.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+type AutoRAGChatSettings = AISearchChatSettings;
+//#endregion
+//#region src/index.d.ts
+type WorkersAISettings = ({
+  /**
+   * Provide a Cloudflare AI binding.
+   */
+  binding: Ai;
+  /**
+   * Credentials must be absent when a binding is given.
+   */
+  accountId?: never;
+  apiKey?: never;
+} | {
+  /**
+   * Provide Cloudflare API credentials directly. Must be used if a binding is not specified.
+   */
+  accountId: string;
+  apiKey: string;
+  /**
+   * Both binding must be absent if credentials are used directly.
+   */
+  binding?: never;
+}) & {
+  /**
+   * Optionally specify a gateway.
+   */
+  gateway?: GatewayOptions;
+};
+interface WorkersAI {
+  (modelId: TextGenerationModels, settings?: WorkersAIChatSettings): WorkersAIChatLanguageModel;
+  /**
+   * Creates a model for text generation.
+   **/
+  chat(modelId: TextGenerationModels, settings?: WorkersAIChatSettings): WorkersAIChatLanguageModel;
+  embedding(modelId: EmbeddingModels, settings?: WorkersAIEmbeddingSettings): WorkersAIEmbeddingModel;
+  textEmbedding(modelId: EmbeddingModels, settings?: WorkersAIEmbeddingSettings): WorkersAIEmbeddingModel;
+  textEmbeddingModel(modelId: EmbeddingModels, settings?: WorkersAIEmbeddingSettings): WorkersAIEmbeddingModel;
+  /**
+   * Creates a model for image generation.
+   **/
+  image(modelId: ImageGenerationModels, settings?: WorkersAIImageSettings): WorkersAIImageModel;
+  imageModel(modelId: ImageGenerationModels, settings?: WorkersAIImageSettings): WorkersAIImageModel;
+  /**
+   * Creates a model for speech-to-text transcription.
+   **/
+  transcription(modelId: TranscriptionModels, settings?: WorkersAITranscriptionSettings): WorkersAITranscriptionModel;
+  transcriptionModel(modelId: TranscriptionModels, settings?: WorkersAITranscriptionSettings): WorkersAITranscriptionModel;
+  /**
+   * Creates a model for text-to-speech synthesis.
+   **/
+  speech(modelId: SpeechModels, settings?: WorkersAISpeechSettings): WorkersAISpeechModel;
+  speechModel(modelId: SpeechModels, settings?: WorkersAISpeechSettings): WorkersAISpeechModel;
+  /**
+   * Creates a model for document reranking.
+   **/
+  reranking(modelId: RerankingModels, settings?: WorkersAIRerankingSettings): WorkersAIRerankingModel;
+  rerankingModel(modelId: RerankingModels, settings?: WorkersAIRerankingSettings): WorkersAIRerankingModel;
+}
+/**
+ * Create a Workers AI provider instance.
+ */
+declare function createWorkersAI(options: WorkersAISettings): WorkersAI;
+type AISearchSettings = {
+  binding: AutoRAG;
+};
+interface AISearchProvider {
+  (settings?: AISearchChatSettings): AISearchChatLanguageModel;
+  /**
+   * Creates a model for text generation.
+   **/
+  chat(settings?: AISearchChatSettings): AISearchChatLanguageModel;
+}
+/**
+ * Create an AI Search provider instance.
+ *
+ * AI Search (formerly AutoRAG) is Cloudflare's managed search service.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+declare function createAISearch(options: AISearchSettings, /** @internal */
+
+providerName?: string): AISearchProvider;
+/**
+ * @deprecated Use `AISearchSettings` instead. AutoRAG has been renamed to AI Search.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+type AutoRAGSettings = AISearchSettings;
+/**
+ * @deprecated Use `AISearchProvider` instead. AutoRAG has been renamed to AI Search.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+type AutoRAGProvider = AISearchProvider;
+/**
+ * @deprecated Use `createAISearch` instead. AutoRAG has been renamed to AI Search.
+ * @see https://developers.cloudflare.com/ai-search/
+ */
+declare function createAutoRAG(options: AISearchSettings): AISearchProvider;
+//#endregion
+export { AISearchChatLanguageModel, type AISearchChatSettings, AISearchProvider, AISearchSettings, AutoRAGChatLanguageModel, type AutoRAGChatSettings, AutoRAGProvider, AutoRAGSettings, WorkersAI, WorkersAIRerankingModel, type WorkersAIRerankingSettings, WorkersAISettings, WorkersAISpeechModel, type WorkersAISpeechSettings, WorkersAITranscriptionModel, type WorkersAITranscriptionSettings, createAISearch, createAutoRAG, createWorkersAI };
+//# sourceMappingURL=index.d.mts.map