@ai-sdk/gateway 4.0.0-beta.6 → 4.0.0-beta.61

package/src/gateway-language-model.ts

@@ -1,11 +1,10 @@
 import type {
-  LanguageModelV3,
-  LanguageModelV3CallOptions,
-  SharedV3Warning,
-  LanguageModelV3FilePart,
-  LanguageModelV3StreamPart,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3StreamResult,
+  LanguageModelV4,
+  LanguageModelV4CallOptions,
+  LanguageModelV4FilePart,
+  LanguageModelV4StreamPart,
+  LanguageModelV4GenerateResult,
+  LanguageModelV4StreamResult,
 } from '@ai-sdk/provider';
 import {
   combineHeaders,
@@ -14,6 +13,9 @@ import {
   createJsonResponseHandler,
   postJsonToApi,
   resolve,
+  serializeModelOptions,
+  WORKFLOW_SERIALIZE,
+  WORKFLOW_DESERIALIZE,
   type ParseResult,
   type Resolvable,
 } from '@ai-sdk/provider-utils';
@@ -28,10 +30,24 @@ type GatewayChatConfig = GatewayConfig & {
   o11yHeaders: Resolvable<Record<string, string>>;
 };
 
-export class GatewayLanguageModel implements LanguageModelV3 {
-  readonly specificationVersion = 'v3';
+export class GatewayLanguageModel implements LanguageModelV4 {
+  readonly specificationVersion = 'v4';
   readonly supportedUrls = { '*/*': [/.*/] };
 
+  static [WORKFLOW_SERIALIZE](model: GatewayLanguageModel) {
+    return serializeModelOptions({
+      modelId: model.modelId,
+      config: model.config,
+    });
+  }
+
+  static [WORKFLOW_DESERIALIZE](options: {
+    modelId: GatewayModelId;
+    config: GatewayChatConfig;
+  }) {
+    return new GatewayLanguageModel(options.modelId, options.config);
+  }
+
   constructor(
     readonly modelId: GatewayModelId,
     private readonly config: GatewayChatConfig,
@@ -41,7 +57,7 @@ export class GatewayLanguageModel implements LanguageModelV3 {
     return this.config.provider;
   }
 
-  private async getArgs(options: LanguageModelV3CallOptions) {
+  private async getArgs(options: LanguageModelV4CallOptions) {
     const { abortSignal: _abortSignal, ...optionsWithoutSignal } = options;
 
     return {
@@ -51,12 +67,14 @@ export class GatewayLanguageModel implements LanguageModelV3 {
   }
 
   async doGenerate(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3GenerateResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4GenerateResult> {
     const { args, warnings } = await this.getArgs(options);
     const { abortSignal } = options;
 
-    const resolvedHeaders = await resolve(this.config.headers());
+    const resolvedHeaders = this.config.headers
+      ? await resolve(this.config.headers)
+      : undefined;
 
     try {
       const {
@@ -88,17 +106,22 @@ export class GatewayLanguageModel implements LanguageModelV3 {
         warnings,
       };
     } catch (error) {
-      throw await asGatewayError(error, await parseAuthMethod(resolvedHeaders));
+      throw await asGatewayError(
+        error,
+        await parseAuthMethod(resolvedHeaders ?? {}),
+      );
     }
   }
 
   async doStream(
-    options: LanguageModelV3CallOptions,
-  ): Promise<LanguageModelV3StreamResult> {
+    options: LanguageModelV4CallOptions,
+  ): Promise<LanguageModelV4StreamResult> {
     const { args, warnings } = await this.getArgs(options);
     const { abortSignal } = options;
 
-    const resolvedHeaders = await resolve(this.config.headers());
+    const resolvedHeaders = this.config.headers
+      ? await resolve(this.config.headers)
+      : undefined;
 
     try {
       const { value: response, responseHeaders } = await postJsonToApi({
@@ -122,8 +145,8 @@ export class GatewayLanguageModel implements LanguageModelV3 {
       return {
         stream: response.pipeThrough(
           new TransformStream<
-            ParseResult<LanguageModelV3StreamPart>,
-            LanguageModelV3StreamPart
+            ParseResult<LanguageModelV4StreamPart>,
+            LanguageModelV4StreamPart
           >({
             start(controller) {
               if (warnings.length > 0) {
@@ -161,7 +184,10 @@ export class GatewayLanguageModel implements LanguageModelV3 {
         response: { headers: responseHeaders },
       };
     } catch (error) {
-      throw await asGatewayError(error, await parseAuthMethod(resolvedHeaders));
+      throw await asGatewayError(
+        error,
+        await parseAuthMethod(resolvedHeaders ?? {}),
+      );
     }
   }
 
@@ -177,11 +203,11 @@ export class GatewayLanguageModel implements LanguageModelV3 {
    * @param options - The options to encode.
    * @returns The options with the file parts encoded.
    */
-  private maybeEncodeFileParts(options: LanguageModelV3CallOptions) {
+  private maybeEncodeFileParts(options: LanguageModelV4CallOptions) {
     for (const message of options.prompt) {
       for (const part of message.content) {
         if (this.isFilePart(part)) {
-          const filePart = part as LanguageModelV3FilePart;
+          const filePart = part as LanguageModelV4FilePart;
           // If the file part is a URL it will get cleanly converted to a string.
           // If it's a binary file attachment we convert it to a data url.
           // In either case, server-side we should only ever see URLs as strings.
@@ -204,7 +230,7 @@ export class GatewayLanguageModel implements LanguageModelV3 {
 
   private getModelConfigHeaders(modelId: string, streaming: boolean) {
     return {
-      'ai-language-model-specification-version': '3',
+      'ai-language-model-specification-version': '4',
       'ai-language-model-id': modelId,
       'ai-language-model-streaming': String(streaming),
     };

package/src/gateway-model-entry.ts

@@ -1,4 +1,14 @@
-import type { LanguageModelV3 } from '@ai-sdk/provider';
+import type { LanguageModelV4 } from '@ai-sdk/provider';
+
+export const KNOWN_MODEL_TYPES = [
+  'embedding',
+  'image',
+  'language',
+  'reranking',
+  'video',
+] as const;
+
+export type KnownModelType = (typeof KNOWN_MODEL_TYPES)[number];
 
 export interface GatewayLanguageModelEntry {
   /**
@@ -49,10 +59,10 @@ export interface GatewayLanguageModelEntry {
   /**
    * Optional field to differentiate between model types.
    */
-  modelType?: 'language' | 'embedding' | 'image' | 'video' | null;
+  modelType?: KnownModelType | null;
 }
 
 export type GatewayLanguageModelSpecification = Pick<
-  LanguageModelV3,
+  LanguageModelV4,
   'specificationVersion' | 'provider' | 'modelId'
 >;

package/src/gateway-provider-options.ts

@@ -2,7 +2,7 @@ import { InferSchema, lazySchema, zodSchema } from '@ai-sdk/provider-utils';
 import { z } from 'zod/v4';
 
 // https://vercel.com/docs/ai-gateway/provider-options
-const gatewayLanguageModelOptions = lazySchema(() =>
+const gatewayProviderOptions = lazySchema(() =>
   zodSchema(
     z.object({
       /**
@@ -17,6 +17,14 @@ const gatewayLanguageModelOptions = lazySchema(() =>
        * Example: `['bedrock', 'anthropic']` will try Amazon Bedrock first, then Anthropic as fallback.
        */
       order: z.array(z.string()).optional(),
+      /**
+       * Sort providers by a performance or cost metric before routing.
+       *
+       * - `'cost'`: lowest cost first
+       * - `'ttft'`: lowest time-to-first-token first
+       * - `'tps'`: highest tokens-per-second first
+       */
+      sort: z.enum(['cost', 'ttft', 'tps']).optional(),
       /**
        * The unique identifier for the end user on behalf of whom the request was made.
        *
@@ -53,12 +61,33 @@ const gatewayLanguageModelOptions = lazySchema(() =>
         .record(z.string(), z.array(z.record(z.string(), z.unknown())))
         .optional(),
       /**
-       * Whether to filter by only providers that state they have zero data
-       * retention with Vercel AI Gateway. When enabled, only providers that
-       * have agreements with Vercel AI Gateway for zero data retention will be
-       * used.
+       * Whether to filter by only providers that have zero data retention
+       * agreements with Vercel for AI Gateway. When using BYOK credentials,
+       * this filter is not applied. If BYOK credentials fail and the request
+       * falls back to system credentials, only providers with zero data
+       * retention agreements will be used.
        */
       zeroDataRetention: z.boolean().optional(),
+      /**
+       * Whether to filter by only providers that do not train on prompt data.
+       * When using BYOK credentials, this filter is not applied. If BYOK
+       * credentials fail and the request falls back to system credentials,
+       * only providers that have agreements with Vercel for AI Gateway to not
+       * use prompts for model training will be used.
+       */
+      disallowPromptTraining: z.boolean().optional(),
+      /**
+       * Whether to filter by only providers that are HIPAA compliant with
+       * Vercel AI Gateway. When enabled, only providers that have agreements
+       * with Vercel AI Gateway for HIPAA compliance will be used.
+       */
+      hipaaCompliant: z.boolean().optional(),
+      /**
+       * The unique identifier for the entity against which quota is tracked.
+       *
+       * Used for quota management and enforcement purposes.
+       */
+      quotaEntityId: z.string().optional(),
       /**
        * Per-provider timeouts for BYOK credentials in milliseconds.
        * Controls how long to wait for a provider to start responding
@@ -75,6 +104,4 @@ const gatewayLanguageModelOptions = lazySchema(() =>
   ),
 );
 
-export type GatewayLanguageModelOptions = InferSchema<
-  typeof gatewayLanguageModelOptions
->;
+export type GatewayProviderOptions = InferSchema<typeof gatewayProviderOptions>;

package/src/gateway-provider.ts

@@ -13,38 +13,51 @@ import {
   type GatewayFetchMetadataResponse,
   type GatewayCreditsResponse,
 } from './gateway-fetch-metadata';
+import {
+  GatewaySpendReport,
+  type GatewaySpendReportParams,
+  type GatewaySpendReportResponse,
+} from './gateway-spend-report';
+import {
+  GatewayGenerationInfoFetcher,
+  type GatewayGenerationInfoParams,
+  type GatewayGenerationInfo,
+} from './gateway-generation-info';
 import { GatewayLanguageModel } from './gateway-language-model';
 import { GatewayEmbeddingModel } from './gateway-embedding-model';
 import { GatewayImageModel } from './gateway-image-model';
 import { GatewayVideoModel } from './gateway-video-model';
+import { GatewayRerankingModel } from './gateway-reranking-model';
 import type { GatewayEmbeddingModelId } from './gateway-embedding-model-settings';
 import type { GatewayImageModelId } from './gateway-image-model-settings';
+import type { GatewayRerankingModelId } from './gateway-reranking-model-settings';
 import type { GatewayVideoModelId } from './gateway-video-model-settings';
 import { gatewayTools } from './gateway-tools';
 import { getVercelOidcToken, getVercelRequestId } from './vercel-environment';
 import type { GatewayModelId } from './gateway-language-model-settings';
 import type {
-  LanguageModelV3,
-  EmbeddingModelV3,
-  ImageModelV3,
-  Experimental_VideoModelV3,
-  ProviderV3,
+  LanguageModelV4,
+  EmbeddingModelV4,
+  ImageModelV4,
+  RerankingModelV4,
+  Experimental_VideoModelV4,
+  ProviderV4,
 } from '@ai-sdk/provider';
 import { withUserAgentSuffix } from '@ai-sdk/provider-utils';
 import { VERSION } from './version';
 
-export interface GatewayProvider extends ProviderV3 {
-  (modelId: GatewayModelId): LanguageModelV3;
+export interface GatewayProvider extends ProviderV4 {
+  (modelId: GatewayModelId): LanguageModelV4;
 
   /**
    * Creates a model for text generation.
    */
-  chat(modelId: GatewayModelId): LanguageModelV3;
+  chat(modelId: GatewayModelId): LanguageModelV4;
 
   /**
    * Creates a model for text generation.
    */
-  languageModel(modelId: GatewayModelId): LanguageModelV3;
+  languageModel(modelId: GatewayModelId): LanguageModelV4;
 
   /**
    * Returns available providers and models for use with the remote provider.
@@ -56,40 +69,66 @@ export interface GatewayProvider extends ProviderV3 {
    */
   getCredits(): Promise<GatewayCreditsResponse>;
 
+  /**
+   * Returns a spend report with cost, token, and request count data,
+   * aggregated by the specified dimension.
+   */
+  getSpendReport(
+    params: GatewaySpendReportParams,
+  ): Promise<GatewaySpendReportResponse>;
+
+  /**
+   * Returns detailed information about a specific generation by its ID,
+   * including cost, token usage, latency, and provider details.
+   */
+  getGenerationInfo(
+    params: GatewayGenerationInfoParams,
+  ): Promise<GatewayGenerationInfo>;
+
   /**
    * Creates a model for generating text embeddings.
    */
-  embedding(modelId: GatewayEmbeddingModelId): EmbeddingModelV3;
+  embedding(modelId: GatewayEmbeddingModelId): EmbeddingModelV4;
 
   /**
    * Creates a model for generating text embeddings.
    */
-  embeddingModel(modelId: GatewayEmbeddingModelId): EmbeddingModelV3;
+  embeddingModel(modelId: GatewayEmbeddingModelId): EmbeddingModelV4;
 
   /**
    * @deprecated Use `embeddingModel` instead.
    */
-  textEmbeddingModel(modelId: GatewayEmbeddingModelId): EmbeddingModelV3;
+  textEmbeddingModel(modelId: GatewayEmbeddingModelId): EmbeddingModelV4;
 
   /**
    * Creates a model for generating images.
    */
-  image(modelId: GatewayImageModelId): ImageModelV3;
+  image(modelId: GatewayImageModelId): ImageModelV4;
 
   /**
    * Creates a model for generating images.
    */
-  imageModel(modelId: GatewayImageModelId): ImageModelV3;
+  imageModel(modelId: GatewayImageModelId): ImageModelV4;
 
   /**
    * Creates a model for generating videos.
    */
-  video(modelId: GatewayVideoModelId): Experimental_VideoModelV3;
+  video(modelId: GatewayVideoModelId): Experimental_VideoModelV4;
 
   /**
    * Creates a model for generating videos.
    */
-  videoModel(modelId: GatewayVideoModelId): Experimental_VideoModelV3;
+  videoModel(modelId: GatewayVideoModelId): Experimental_VideoModelV4;
+
+  /**
+   * Creates a model for reranking documents.
+   */
+  reranking(modelId: GatewayRerankingModelId): RerankingModelV4;
+
+  /**
+   * Creates a model for reranking documents.
+   */
+  rerankingModel(modelId: GatewayRerankingModelId): RerankingModelV4;
 
   /**
    * Gateway-specific tools executed server-side.
@@ -148,7 +187,7 @@ export function createGatewayProvider(
 
   const baseURL =
     withoutTrailingSlash(options.baseURL) ??
-    'https://ai-gateway.vercel.sh/v3/ai';
+    'https://ai-gateway.vercel.sh/v4/ai';
 
   const getHeaders = async () => {
     try {
@@ -253,6 +292,36 @@ export function createGatewayProvider(
       });
   };
 
+  const getSpendReport = async (params: GatewaySpendReportParams) => {
+    return new GatewaySpendReport({
+      baseURL,
+      headers: getHeaders,
+      fetch: options.fetch,
+    })
+      .getSpendReport(params)
+      .catch(async (error: unknown) => {
+        throw await asGatewayError(
+          error,
+          await parseAuthMethod(await getHeaders()),
+        );
+      });
+  };
+
+  const getGenerationInfo = async (params: GatewayGenerationInfoParams) => {
+    return new GatewayGenerationInfoFetcher({
+      baseURL,
+      headers: getHeaders,
+      fetch: options.fetch,
+    })
+      .getGenerationInfo(params)
+      .catch(async (error: unknown) => {
+        throw await asGatewayError(
+          error,
+          await parseAuthMethod(await getHeaders()),
+        );
+      });
+  };
+
   const provider = function (modelId: GatewayModelId) {
     if (new.target) {
       throw new Error(
@@ -263,9 +332,11 @@ export function createGatewayProvider(
     return createLanguageModel(modelId);
   };
 
-  provider.specificationVersion = 'v3' as const;
+  provider.specificationVersion = 'v4' as const;
   provider.getAvailableModels = getAvailableModels;
   provider.getCredits = getCredits;
+  provider.getSpendReport = getSpendReport;
+  provider.getGenerationInfo = getGenerationInfo;
   provider.imageModel = (modelId: GatewayImageModelId) => {
     return new GatewayImageModel(modelId, {
       provider: 'gateway',
@@ -296,6 +367,17 @@ export function createGatewayProvider(
       o11yHeaders: createO11yHeaders(),
     });
   };
+  const createRerankingModel = (modelId: GatewayRerankingModelId) => {
+    return new GatewayRerankingModel(modelId, {
+      provider: 'gateway',
+      baseURL,
+      headers: getHeaders,
+      fetch: options.fetch,
+      o11yHeaders: createO11yHeaders(),
+    });
+  };
+  provider.rerankingModel = createRerankingModel;
+  provider.reranking = createRerankingModel;
   provider.chat = provider.languageModel;
   provider.embedding = provider.embeddingModel;
   provider.image = provider.imageModel;

package/src/gateway-reranking-model-settings.ts

@@ -0,0 +1,7 @@
+export type GatewayRerankingModelId =
+  | 'cohere/rerank-v3.5'
+  | 'cohere/rerank-v4-fast'
+  | 'cohere/rerank-v4-pro'
+  | 'voyage/rerank-2.5'
+  | 'voyage/rerank-2.5-lite'
+  | (string & {});

package/src/gateway-reranking-model.ts

@@ -0,0 +1,119 @@
+import type {
+  RerankingModelV4,
+  SharedV4ProviderMetadata,
+} from '@ai-sdk/provider';
+import {
+  combineHeaders,
+  createJsonErrorResponseHandler,
+  createJsonResponseHandler,
+  lazySchema,
+  postJsonToApi,
+  resolve,
+  zodSchema,
+  type Resolvable,
+} from '@ai-sdk/provider-utils';
+import { z } from 'zod/v4';
+import { asGatewayError } from './errors';
+import { parseAuthMethod } from './errors/parse-auth-method';
+import type { GatewayConfig } from './gateway-config';
+
+export class GatewayRerankingModel implements RerankingModelV4 {
+  readonly specificationVersion = 'v4';
+
+  constructor(
+    readonly modelId: string,
+    private readonly config: GatewayConfig & {
+      provider: string;
+      o11yHeaders: Resolvable<Record<string, string>>;
+    },
+  ) {}
+
+  get provider(): string {
+    return this.config.provider;
+  }
+
+  async doRerank({
+    documents,
+    query,
+    topN,
+    headers,
+    abortSignal,
+    providerOptions,
+  }: Parameters<RerankingModelV4['doRerank']>[0]): Promise<
+    Awaited<ReturnType<RerankingModelV4['doRerank']>>
+  > {
+    const resolvedHeaders = this.config.headers
+      ? await resolve(this.config.headers)
+      : undefined;
+    try {
+      const {
+        responseHeaders,
+        value: responseBody,
+        rawValue,
+      } = await postJsonToApi({
+        url: this.getUrl(),
+        headers: combineHeaders(
+          resolvedHeaders,
+          headers ?? {},
+          this.getModelConfigHeaders(),
+          await resolve(this.config.o11yHeaders),
+        ),
+        body: {
+          documents,
+          query,
+          ...(topN != null ? { topN } : {}),
+          ...(providerOptions ? { providerOptions } : {}),
+        },
+        successfulResponseHandler: createJsonResponseHandler(
+          gatewayRerankingResponseSchema,
+        ),
+        failedResponseHandler: createJsonErrorResponseHandler({
+          errorSchema: z.any(),
+          errorToMessage: data => data,
+        }),
+        ...(abortSignal && { abortSignal }),
+        fetch: this.config.fetch,
+      });
+
+      return {
+        ranking: responseBody.ranking,
+        providerMetadata:
+          responseBody.providerMetadata as unknown as SharedV4ProviderMetadata,
+        response: { headers: responseHeaders, body: rawValue },
+        warnings: [],
+      };
+    } catch (error) {
+      throw await asGatewayError(
+        error,
+        await parseAuthMethod(resolvedHeaders ?? {}),
+      );
+    }
+  }
+
+  private getUrl() {
+    return `${this.config.baseURL}/reranking-model`;
+  }
+
+  private getModelConfigHeaders() {
+    return {
+      'ai-reranking-model-specification-version': '4',
+      'ai-model-id': this.modelId,
+    };
+  }
+}
+
+const gatewayRerankingResponseSchema = lazySchema(() =>
+  zodSchema(
+    z.object({
+      ranking: z.array(
+        z.object({
+          index: z.number(),
+          relevanceScore: z.number(),
+        }),
+      ),
+      providerMetadata: z
+        .record(z.string(), z.record(z.string(), z.unknown()))
+        .optional(),
+    }),
+  ),
+);