@ai-sdk/baseten 2.1.12 → 2.1.14

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # @ai-sdk/baseten
 
+## 2.1.14
+
+### Patch Changes
+
+- c8cc95f: fix(baseten): enable structured outputs so response_format json_schema is forwarded
+- Updated dependencies [ece5bdb]
+  - @ai-sdk/openai-compatible@3.0.36
+
+## 2.1.13
+
+### Patch Changes
+
+- Updated dependencies [b74971f]
+- Updated dependencies [7dd9ec3]
+  - @ai-sdk/provider-utils@5.0.29
+  - @ai-sdk/openai-compatible@3.0.35
+
 ## 2.1.12
 
 ### Patch Changes

package/dist/index.js

@@ -14,7 +14,7 @@ import {
 import { z } from "zod/v4";
 
 // src/version.ts
-var VERSION = true ? "2.1.12" : "0.0.0-test";
+var VERSION = true ? "2.1.14" : "0.0.0-test";
 
 // src/baseten-provider.ts
 var MAX_EMBEDDINGS_PER_CALL = 128;
@@ -68,7 +68,8 @@ function createBaseten(options = {}) {
           ...getCommonModelConfig("chat", customURL),
           errorStructure: basetenErrorStructure,
           // Or stream_options.include_usage is omitted and streams report no usage.
-          includeUsage: true
+          includeUsage: true,
+          supportsStructuredOutputs: true
         });
       } else if (customURL.includes("/predict")) {
         throw new Error(
@@ -79,7 +80,8 @@ function createBaseten(options = {}) {
     return new OpenAICompatibleChatLanguageModel(modelId != null ? modelId : "chat", {
       ...getCommonModelConfig("chat"),
       errorStructure: basetenErrorStructure,
-      includeUsage: true
+      includeUsage: true,
+      supportsStructuredOutputs: true
     });
   };
   const createEmbeddingModel = (modelId) => {

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/baseten-provider.ts","../src/version.ts"],"sourcesContent":["import {\n  OpenAICompatibleChatLanguageModel,\n  OpenAICompatibleEmbeddingModel,\n  type ProviderErrorStructure,\n} from '@ai-sdk/openai-compatible';\nimport {\n  NoSuchModelError,\n  type EmbeddingModelV4,\n  type LanguageModelV4,\n  type ProviderV4,\n} from '@ai-sdk/provider';\nimport {\n  loadApiKey,\n  withoutTrailingSlash,\n  withUserAgentSuffix,\n  type FetchFunction,\n} from '@ai-sdk/provider-utils';\nimport { z } from 'zod/v4';\nimport type { BasetenChatModelId } from './baseten-chat-options';\nimport type { BasetenEmbeddingModelId } from './baseten-embedding-options';\nimport { VERSION } from './version';\n\n/**\n * Baseten's per-request embedding input limit: larger batches are rejected with\n * `413 batch size N > maximum allowed batch size 128`. It is also the native\n * performance client's own default `batchSize`. `embedMany` splits larger\n * inputs into chunks of this size and runs them in parallel.\n */\nconst MAX_EMBEDDINGS_PER_CALL = 128;\n\n/**\n * The part of `@basetenlabs/performance-client` we use, declared structurally to\n * keep that native addon out of our dependency and type graph.\n */\nexport type BasetenPerformanceClient = {\n  embed(\n    input: string[],\n    model: string,\n  ): Promise<{\n    data: { embedding: number[] }[];\n    usage?: { total_tokens?: number };\n  }>;\n};\n\nexport type BasetenPerformanceClientConstructor = new (\n  baseUrl: string,\n  apiKey?: string,\n) => BasetenPerformanceClient;\n\nexport type BasetenErrorData = z.infer<typeof basetenErrorSchema>;\n\n// Baseten returns two different envelopes. The Model APIs send a bare string\n// (`{\"error\":\"please check the model you provided\"}`), while dedicated\n// deployments pass through their server's OpenAI-shaped object. Parsing only\n// the string form left dedicated-deployment errors falling back to the HTTP\n// reason phrase — \"Not Found\", or nothing at all over HTTP/2.\nconst basetenErrorSchema = z.object({\n  error: z.union([\n    z.string(),\n    z.object({\n      message: z.string(),\n      object: z.string().nullish(),\n      type: z.string().nullish(),\n      param: z.any().nullish(),\n      code: z.union([z.string(), z.number()]).nullish(),\n    }),\n  ]),\n});\n\nconst basetenErrorStructure: ProviderErrorStructure<BasetenErrorData> = {\n  errorSchema: basetenErrorSchema,\n  errorToMessage: data =>\n    typeof data.error === 'string' ? data.error : data.error.message,\n};\n\nexport interface BasetenProviderSettings {\n  /**\n   * Baseten API key. Default value is taken from the `BASETEN_API_KEY`\n   * environment variable.\n   */\n  apiKey?: string;\n\n  /**\n   * Base URL for the Model APIs. Default: 'https://inference.baseten.co/v1'\n   */\n  baseURL?: string;\n\n  /**\n   * Model URL for custom models (chat or embeddings).\n   * If not supplied, the default Model APIs will be used.\n   */\n  modelURL?: string;\n  /**\n   * Custom headers to include in the requests.\n   */\n  headers?: Record<string, string>;\n\n  /**\n   * Custom fetch implementation. You can use it as a middleware to intercept requests,\n   * or to provide a custom fetch implementation for e.g. testing.\n   */\n  fetch?: FetchFunction;\n\n  /**\n   * Opt in to Baseten's native performance client for embeddings, for\n   * client-side batching and request hedging. Pass the `PerformanceClient`\n   * constructor from `@basetenlabs/performance-client`, which you install\n   * yourself:\n   *\n   * ```ts\n   * import { PerformanceClient } from '@basetenlabs/performance-client';\n   *\n   * const baseten = createBaseten({ modelURL, performanceClient: PerformanceClient });\n   * ```\n   *\n   * When omitted, embeddings go over plain HTTP to Baseten's OpenAI-compatible\n   * endpoint — the default, since this NAPI addon cannot load in edge runtimes\n   * and bundlers cannot resolve its platform binaries.\n   */\n  performanceClient?: BasetenPerformanceClientConstructor;\n}\n\nexport interface BasetenProvider extends ProviderV4 {\n  /**\n   * Creates a chat model for text generation.\n   */\n  (modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a chat model for text generation.\n   */\n  chatModel(modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a language model for text generation. Alias for chatModel.\n   */\n  languageModel(modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a embedding model for text generation.\n   */\n  embeddingModel(modelId?: BasetenEmbeddingModelId): EmbeddingModelV4;\n\n  /**\n   * @deprecated Use `embeddingModel` instead.\n   */\n  textEmbeddingModel(modelId?: BasetenEmbeddingModelId): EmbeddingModelV4;\n}\n\n// by default, we use the Model APIs\nconst defaultBaseURL = 'https://inference.baseten.co/v1';\n\nexport function createBaseten(\n  options: BasetenProviderSettings = {},\n): BasetenProvider {\n  const baseURL = withoutTrailingSlash(options.baseURL ?? defaultBaseURL);\n  const getHeaders = () =>\n    withUserAgentSuffix(\n      {\n        Authorization: `Bearer ${loadApiKey({\n          apiKey: options.apiKey,\n          environmentVariableName: 'BASETEN_API_KEY',\n          description: 'Baseten API key',\n        })}`,\n        ...options.headers,\n      },\n      `ai-sdk/baseten/${VERSION}`,\n    );\n\n  interface CommonModelConfig {\n    provider: string;\n    url: ({ path }: { path: string }) => string;\n    headers: () => Record<string, string>;\n    fetch?: FetchFunction;\n  }\n\n  const getCommonModelConfig = (\n    modelType: string,\n    customURL?: string,\n  ): CommonModelConfig => ({\n    provider: `baseten.${modelType}`,\n    url: ({ path }) => {\n      // For embeddings with /sync URLs (but not /sync/v1), we need to add /v1\n      if (\n        modelType === 'embedding' &&\n        customURL?.includes('/sync') &&\n        !customURL?.includes('/sync/v1')\n      ) {\n        return `${customURL}/v1${path}`;\n      }\n      return `${customURL || baseURL}${path}`;\n    },\n    headers: getHeaders,\n    fetch: options.fetch,\n  });\n\n  const createChatModel = (modelId?: BasetenChatModelId) => {\n    const customURL = options.modelURL;\n    if (customURL) {\n      if (customURL.includes('/sync/v1')) {\n        return new OpenAICompatibleChatLanguageModel(modelId ?? 'placeholder', {\n          ...getCommonModelConfig('chat', customURL),\n          errorStructure: basetenErrorStructure,\n          // Or stream_options.include_usage is omitted and streams report no usage.\n          includeUsage: true,\n        });\n      } else if (customURL.includes('/predict')) {\n        throw new Error(\n          'Not supported. You must use a /sync/v1 endpoint for chat models.',\n        );\n      }\n    }\n\n    return new OpenAICompatibleChatLanguageModel(modelId ?? 'chat', {\n      ...getCommonModelConfig('chat'),\n      errorStructure: basetenErrorStructure,\n      includeUsage: true,\n    });\n  };\n\n  const createEmbeddingModel = (modelId?: BasetenEmbeddingModelId) => {\n    const customURL = options.modelURL;\n    if (!customURL) {\n      throw new Error(\n        'No model URL provided for embeddings. Please set modelURL option for embeddings.',\n      );\n    }\n\n    if (!customURL.includes('/sync')) {\n      throw new Error(\n        'Not supported. You must use a /sync or /sync/v1 endpoint for embeddings.',\n      );\n    }\n\n    // BEI embedding deployments are OpenAI-compatible with no extra settings, so\n    // plain HTTP is the default and needs no override.\n    const model = new OpenAICompatibleEmbeddingModel(modelId ?? 'embeddings', {\n      ...getCommonModelConfig('embedding', customURL),\n      errorStructure: basetenErrorStructure,\n      // Over HTTP, cap each request and let `embedMany` split and parallelise.\n      // The native client does its own batching, so let it take everything at\n      // once — `embedMany` treats Infinity as \"one call\".\n      maxEmbeddingsPerCall: options.performanceClient\n        ? Number.POSITIVE_INFINITY\n        : MAX_EMBEDDINGS_PER_CALL,\n    });\n\n    if (!options.performanceClient) {\n      return model;\n    }\n\n    // Opted in to the native client. It appends /v1 itself, so hand it the bare\n    // /sync form.\n    const performanceClient = new options.performanceClient(\n      customURL.replace('/sync/v1', '/sync'),\n      loadApiKey({\n        apiKey: options.apiKey,\n        environmentVariableName: 'BASETEN_API_KEY',\n        description: 'Baseten API key',\n      }),\n    );\n\n    model.doEmbed = async params => {\n      if (!params.values || !Array.isArray(params.values)) {\n        throw new Error('params.values must be an array of strings');\n      }\n\n      const response = await performanceClient.embed(\n        params.values,\n        // model_id is for Model APIs; dedicated deployments ignore it.\n        modelId ?? 'embeddings',\n      );\n\n      return {\n        embeddings: response.data.map(item => item.embedding),\n        // The native client types its response as `any`; only report usage when\n        // a token count is actually present rather than `{ tokens: undefined }`.\n        usage:\n          typeof response.usage?.total_tokens === 'number'\n            ? { tokens: response.usage.total_tokens }\n            : undefined,\n        response: { headers: {}, body: response },\n        warnings: [],\n      };\n    };\n\n    return model;\n  };\n\n  const provider = (modelId?: BasetenChatModelId) => createChatModel(modelId);\n\n  provider.specificationVersion = 'v4' as const;\n  provider.chatModel = createChatModel;\n  provider.languageModel = createChatModel;\n  provider.imageModel = (modelId: string) => {\n    throw new NoSuchModelError({ modelId, modelType: 'imageModel' });\n  };\n  provider.embeddingModel = createEmbeddingModel;\n  provider.textEmbeddingModel = createEmbeddingModel;\n  return provider;\n}\n\nexport const baseten = createBaseten();\n","// Version string of this package injected at build time.\ndeclare const __PACKAGE_VERSION__: string | undefined;\nexport const VERSION: string =\n  typeof __PACKAGE_VERSION__ !== 'undefined'\n    ? __PACKAGE_VERSION__\n    : '0.0.0-test';\n"],"mappings":";AAAA;AAAA,EACE;AAAA,EACA;AAAA,OAEK;AACP;AAAA,EACE;AAAA,OAIK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OAEK;AACP,SAAS,SAAS;;;ACfX,IAAM,UACX,OACI,WACA;;;ADuBN,IAAM,0BAA0B;AA4BhC,IAAM,qBAAqB,EAAE,OAAO;AAAA,EAClC,OAAO,EAAE,MAAM;AAAA,IACb,EAAE,OAAO;AAAA,IACT,EAAE,OAAO;AAAA,MACP,SAAS,EAAE,OAAO;AAAA,MAClB,QAAQ,EAAE,OAAO,EAAE,QAAQ;AAAA,MAC3B,MAAM,EAAE,OAAO,EAAE,QAAQ;AAAA,MACzB,OAAO,EAAE,IAAI,EAAE,QAAQ;AAAA,MACvB,MAAM,EAAE,MAAM,CAAC,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,CAAC,EAAE,QAAQ;AAAA,IAClD,CAAC;AAAA,EACH,CAAC;AACH,CAAC;AAED,IAAM,wBAAkE;AAAA,EACtE,aAAa;AAAA,EACb,gBAAgB,UACd,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ,KAAK,MAAM;AAC7D;AA6EA,IAAM,iBAAiB;AAEhB,SAAS,cACd,UAAmC,CAAC,GACnB;AA1JnB;AA2JE,QAAM,UAAU,sBAAqB,aAAQ,YAAR,YAAmB,cAAc;AACtE,QAAM,aAAa,MACjB;AAAA,IACE;AAAA,MACE,eAAe,UAAU,WAAW;AAAA,QAClC,QAAQ,QAAQ;AAAA,QAChB,yBAAyB;AAAA,QACzB,aAAa;AAAA,MACf,CAAC,CAAC;AAAA,MACF,GAAG,QAAQ;AAAA,IACb;AAAA,IACA,kBAAkB,OAAO;AAAA,EAC3B;AASF,QAAM,uBAAuB,CAC3B,WACA,eACuB;AAAA,IACvB,UAAU,WAAW,SAAS;AAAA,IAC9B,KAAK,CAAC,EAAE,KAAK,MAAM;AAEjB,UACE,cAAc,gBACd,uCAAW,SAAS,aACpB,EAAC,uCAAW,SAAS,cACrB;AACA,eAAO,GAAG,SAAS,MAAM,IAAI;AAAA,MAC/B;AACA,aAAO,GAAG,aAAa,OAAO,GAAG,IAAI;AAAA,IACvC;AAAA,IACA,SAAS;AAAA,IACT,OAAO,QAAQ;AAAA,EACjB;AAEA,QAAM,kBAAkB,CAAC,YAAiC;AACxD,UAAM,YAAY,QAAQ;AAC1B,QAAI,WAAW;AACb,UAAI,UAAU,SAAS,UAAU,GAAG;AAClC,eAAO,IAAI,kCAAkC,4BAAW,eAAe;AAAA,UACrE,GAAG,qBAAqB,QAAQ,SAAS;AAAA,UACzC,gBAAgB;AAAA;AAAA,UAEhB,cAAc;AAAA,QAChB,CAAC;AAAA,MACH,WAAW,UAAU,SAAS,UAAU,GAAG;AACzC,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAEA,WAAO,IAAI,kCAAkC,4BAAW,QAAQ;AAAA,MAC9D,GAAG,qBAAqB,MAAM;AAAA,MAC9B,gBAAgB;AAAA,MAChB,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,QAAM,uBAAuB,CAAC,YAAsC;AAClE,UAAM,YAAY,QAAQ;AAC1B,QAAI,CAAC,WAAW;AACd,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,QAAI,CAAC,UAAU,SAAS,OAAO,GAAG;AAChC,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAIA,UAAM,QAAQ,IAAI,+BAA+B,4BAAW,cAAc;AAAA,MACxE,GAAG,qBAAqB,aAAa,SAAS;AAAA,MAC9C,gBAAgB;AAAA;AAAA;AAAA;AAAA,MAIhB,sBAAsB,QAAQ,oBAC1B,OAAO,oBACP;AAAA,IACN,CAAC;AAED,QAAI,CAAC,QAAQ,mBAAmB;AAC9B,aAAO;AAAA,IACT;AAIA,UAAM,oBAAoB,IAAI,QAAQ;AAAA,MACpC,UAAU,QAAQ,YAAY,OAAO;AAAA,MACrC,WAAW;AAAA,QACT,QAAQ,QAAQ;AAAA,QAChB,yBAAyB;AAAA,QACzB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAEA,UAAM,UAAU,OAAM,WAAU;AAtQpC,UAAAA;AAuQM,UAAI,CAAC,OAAO,UAAU,CAAC,MAAM,QAAQ,OAAO,MAAM,GAAG;AACnD,cAAM,IAAI,MAAM,2CAA2C;AAAA,MAC7D;AAEA,YAAM,WAAW,MAAM,kBAAkB;AAAA,QACvC,OAAO;AAAA;AAAA,QAEP,4BAAW;AAAA,MACb;AAEA,aAAO;AAAA,QACL,YAAY,SAAS,KAAK,IAAI,UAAQ,KAAK,SAAS;AAAA;AAAA;AAAA,QAGpD,OACE,SAAOA,MAAA,SAAS,UAAT,gBAAAA,IAAgB,kBAAiB,WACpC,EAAE,QAAQ,SAAS,MAAM,aAAa,IACtC;AAAA,QACN,UAAU,EAAE,SAAS,CAAC,GAAG,MAAM,SAAS;AAAA,QACxC,UAAU,CAAC;AAAA,MACb;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,QAAM,WAAW,CAAC,YAAiC,gBAAgB,OAAO;AAE1E,WAAS,uBAAuB;AAChC,WAAS,YAAY;AACrB,WAAS,gBAAgB;AACzB,WAAS,aAAa,CAAC,YAAoB;AACzC,UAAM,IAAI,iBAAiB,EAAE,SAAS,WAAW,aAAa,CAAC;AAAA,EACjE;AACA,WAAS,iBAAiB;AAC1B,WAAS,qBAAqB;AAC9B,SAAO;AACT;AAEO,IAAM,UAAU,cAAc;","names":["_a"]}
+{"version":3,"sources":["../src/baseten-provider.ts","../src/version.ts"],"sourcesContent":["import {\n  OpenAICompatibleChatLanguageModel,\n  OpenAICompatibleEmbeddingModel,\n  type ProviderErrorStructure,\n} from '@ai-sdk/openai-compatible';\nimport {\n  NoSuchModelError,\n  type EmbeddingModelV4,\n  type LanguageModelV4,\n  type ProviderV4,\n} from '@ai-sdk/provider';\nimport {\n  loadApiKey,\n  withoutTrailingSlash,\n  withUserAgentSuffix,\n  type FetchFunction,\n} from '@ai-sdk/provider-utils';\nimport { z } from 'zod/v4';\nimport type { BasetenChatModelId } from './baseten-chat-options';\nimport type { BasetenEmbeddingModelId } from './baseten-embedding-options';\nimport { VERSION } from './version';\n\n/**\n * Baseten's per-request embedding input limit: larger batches are rejected with\n * `413 batch size N > maximum allowed batch size 128`. It is also the native\n * performance client's own default `batchSize`. `embedMany` splits larger\n * inputs into chunks of this size and runs them in parallel.\n */\nconst MAX_EMBEDDINGS_PER_CALL = 128;\n\n/**\n * The part of `@basetenlabs/performance-client` we use, declared structurally to\n * keep that native addon out of our dependency and type graph.\n */\nexport type BasetenPerformanceClient = {\n  embed(\n    input: string[],\n    model: string,\n  ): Promise<{\n    data: { embedding: number[] }[];\n    usage?: { total_tokens?: number };\n  }>;\n};\n\nexport type BasetenPerformanceClientConstructor = new (\n  baseUrl: string,\n  apiKey?: string,\n) => BasetenPerformanceClient;\n\nexport type BasetenErrorData = z.infer<typeof basetenErrorSchema>;\n\n// Baseten returns two different envelopes. The Model APIs send a bare string\n// (`{\"error\":\"please check the model you provided\"}`), while dedicated\n// deployments pass through their server's OpenAI-shaped object. Parsing only\n// the string form left dedicated-deployment errors falling back to the HTTP\n// reason phrase — \"Not Found\", or nothing at all over HTTP/2.\nconst basetenErrorSchema = z.object({\n  error: z.union([\n    z.string(),\n    z.object({\n      message: z.string(),\n      object: z.string().nullish(),\n      type: z.string().nullish(),\n      param: z.any().nullish(),\n      code: z.union([z.string(), z.number()]).nullish(),\n    }),\n  ]),\n});\n\nconst basetenErrorStructure: ProviderErrorStructure<BasetenErrorData> = {\n  errorSchema: basetenErrorSchema,\n  errorToMessage: data =>\n    typeof data.error === 'string' ? data.error : data.error.message,\n};\n\nexport interface BasetenProviderSettings {\n  /**\n   * Baseten API key. Default value is taken from the `BASETEN_API_KEY`\n   * environment variable.\n   */\n  apiKey?: string;\n\n  /**\n   * Base URL for the Model APIs. Default: 'https://inference.baseten.co/v1'\n   */\n  baseURL?: string;\n\n  /**\n   * Model URL for custom models (chat or embeddings).\n   * If not supplied, the default Model APIs will be used.\n   */\n  modelURL?: string;\n  /**\n   * Custom headers to include in the requests.\n   */\n  headers?: Record<string, string>;\n\n  /**\n   * Custom fetch implementation. You can use it as a middleware to intercept requests,\n   * or to provide a custom fetch implementation for e.g. testing.\n   */\n  fetch?: FetchFunction;\n\n  /**\n   * Opt in to Baseten's native performance client for embeddings, for\n   * client-side batching and request hedging. Pass the `PerformanceClient`\n   * constructor from `@basetenlabs/performance-client`, which you install\n   * yourself:\n   *\n   * ```ts\n   * import { PerformanceClient } from '@basetenlabs/performance-client';\n   *\n   * const baseten = createBaseten({ modelURL, performanceClient: PerformanceClient });\n   * ```\n   *\n   * When omitted, embeddings go over plain HTTP to Baseten's OpenAI-compatible\n   * endpoint — the default, since this NAPI addon cannot load in edge runtimes\n   * and bundlers cannot resolve its platform binaries.\n   */\n  performanceClient?: BasetenPerformanceClientConstructor;\n}\n\nexport interface BasetenProvider extends ProviderV4 {\n  /**\n   * Creates a chat model for text generation.\n   */\n  (modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a chat model for text generation.\n   */\n  chatModel(modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a language model for text generation. Alias for chatModel.\n   */\n  languageModel(modelId?: BasetenChatModelId): LanguageModelV4;\n\n  /**\n   * Creates a embedding model for text generation.\n   */\n  embeddingModel(modelId?: BasetenEmbeddingModelId): EmbeddingModelV4;\n\n  /**\n   * @deprecated Use `embeddingModel` instead.\n   */\n  textEmbeddingModel(modelId?: BasetenEmbeddingModelId): EmbeddingModelV4;\n}\n\n// by default, we use the Model APIs\nconst defaultBaseURL = 'https://inference.baseten.co/v1';\n\nexport function createBaseten(\n  options: BasetenProviderSettings = {},\n): BasetenProvider {\n  const baseURL = withoutTrailingSlash(options.baseURL ?? defaultBaseURL);\n  const getHeaders = () =>\n    withUserAgentSuffix(\n      {\n        Authorization: `Bearer ${loadApiKey({\n          apiKey: options.apiKey,\n          environmentVariableName: 'BASETEN_API_KEY',\n          description: 'Baseten API key',\n        })}`,\n        ...options.headers,\n      },\n      `ai-sdk/baseten/${VERSION}`,\n    );\n\n  interface CommonModelConfig {\n    provider: string;\n    url: ({ path }: { path: string }) => string;\n    headers: () => Record<string, string>;\n    fetch?: FetchFunction;\n  }\n\n  const getCommonModelConfig = (\n    modelType: string,\n    customURL?: string,\n  ): CommonModelConfig => ({\n    provider: `baseten.${modelType}`,\n    url: ({ path }) => {\n      // For embeddings with /sync URLs (but not /sync/v1), we need to add /v1\n      if (\n        modelType === 'embedding' &&\n        customURL?.includes('/sync') &&\n        !customURL?.includes('/sync/v1')\n      ) {\n        return `${customURL}/v1${path}`;\n      }\n      return `${customURL || baseURL}${path}`;\n    },\n    headers: getHeaders,\n    fetch: options.fetch,\n  });\n\n  const createChatModel = (modelId?: BasetenChatModelId) => {\n    const customURL = options.modelURL;\n    if (customURL) {\n      if (customURL.includes('/sync/v1')) {\n        return new OpenAICompatibleChatLanguageModel(modelId ?? 'placeholder', {\n          ...getCommonModelConfig('chat', customURL),\n          errorStructure: basetenErrorStructure,\n          // Or stream_options.include_usage is omitted and streams report no usage.\n          includeUsage: true,\n          supportsStructuredOutputs: true,\n        });\n      } else if (customURL.includes('/predict')) {\n        throw new Error(\n          'Not supported. You must use a /sync/v1 endpoint for chat models.',\n        );\n      }\n    }\n\n    return new OpenAICompatibleChatLanguageModel(modelId ?? 'chat', {\n      ...getCommonModelConfig('chat'),\n      errorStructure: basetenErrorStructure,\n      includeUsage: true,\n      supportsStructuredOutputs: true,\n    });\n  };\n\n  const createEmbeddingModel = (modelId?: BasetenEmbeddingModelId) => {\n    const customURL = options.modelURL;\n    if (!customURL) {\n      throw new Error(\n        'No model URL provided for embeddings. Please set modelURL option for embeddings.',\n      );\n    }\n\n    if (!customURL.includes('/sync')) {\n      throw new Error(\n        'Not supported. You must use a /sync or /sync/v1 endpoint for embeddings.',\n      );\n    }\n\n    // BEI embedding deployments are OpenAI-compatible with no extra settings, so\n    // plain HTTP is the default and needs no override.\n    const model = new OpenAICompatibleEmbeddingModel(modelId ?? 'embeddings', {\n      ...getCommonModelConfig('embedding', customURL),\n      errorStructure: basetenErrorStructure,\n      // Over HTTP, cap each request and let `embedMany` split and parallelise.\n      // The native client does its own batching, so let it take everything at\n      // once — `embedMany` treats Infinity as \"one call\".\n      maxEmbeddingsPerCall: options.performanceClient\n        ? Number.POSITIVE_INFINITY\n        : MAX_EMBEDDINGS_PER_CALL,\n    });\n\n    if (!options.performanceClient) {\n      return model;\n    }\n\n    // Opted in to the native client. It appends /v1 itself, so hand it the bare\n    // /sync form.\n    const performanceClient = new options.performanceClient(\n      customURL.replace('/sync/v1', '/sync'),\n      loadApiKey({\n        apiKey: options.apiKey,\n        environmentVariableName: 'BASETEN_API_KEY',\n        description: 'Baseten API key',\n      }),\n    );\n\n    model.doEmbed = async params => {\n      if (!params.values || !Array.isArray(params.values)) {\n        throw new Error('params.values must be an array of strings');\n      }\n\n      const response = await performanceClient.embed(\n        params.values,\n        // model_id is for Model APIs; dedicated deployments ignore it.\n        modelId ?? 'embeddings',\n      );\n\n      return {\n        embeddings: response.data.map(item => item.embedding),\n        // The native client types its response as `any`; only report usage when\n        // a token count is actually present rather than `{ tokens: undefined }`.\n        usage:\n          typeof response.usage?.total_tokens === 'number'\n            ? { tokens: response.usage.total_tokens }\n            : undefined,\n        response: { headers: {}, body: response },\n        warnings: [],\n      };\n    };\n\n    return model;\n  };\n\n  const provider = (modelId?: BasetenChatModelId) => createChatModel(modelId);\n\n  provider.specificationVersion = 'v4' as const;\n  provider.chatModel = createChatModel;\n  provider.languageModel = createChatModel;\n  provider.imageModel = (modelId: string) => {\n    throw new NoSuchModelError({ modelId, modelType: 'imageModel' });\n  };\n  provider.embeddingModel = createEmbeddingModel;\n  provider.textEmbeddingModel = createEmbeddingModel;\n  return provider;\n}\n\nexport const baseten = createBaseten();\n","// Version string of this package injected at build time.\ndeclare const __PACKAGE_VERSION__: string | undefined;\nexport const VERSION: string =\n  typeof __PACKAGE_VERSION__ !== 'undefined'\n    ? __PACKAGE_VERSION__\n    : '0.0.0-test';\n"],"mappings":";AAAA;AAAA,EACE;AAAA,EACA;AAAA,OAEK;AACP;AAAA,EACE;AAAA,OAIK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OAEK;AACP,SAAS,SAAS;;;ACfX,IAAM,UACX,OACI,WACA;;;ADuBN,IAAM,0BAA0B;AA4BhC,IAAM,qBAAqB,EAAE,OAAO;AAAA,EAClC,OAAO,EAAE,MAAM;AAAA,IACb,EAAE,OAAO;AAAA,IACT,EAAE,OAAO;AAAA,MACP,SAAS,EAAE,OAAO;AAAA,MAClB,QAAQ,EAAE,OAAO,EAAE,QAAQ;AAAA,MAC3B,MAAM,EAAE,OAAO,EAAE,QAAQ;AAAA,MACzB,OAAO,EAAE,IAAI,EAAE,QAAQ;AAAA,MACvB,MAAM,EAAE,MAAM,CAAC,EAAE,OAAO,GAAG,EAAE,OAAO,CAAC,CAAC,EAAE,QAAQ;AAAA,IAClD,CAAC;AAAA,EACH,CAAC;AACH,CAAC;AAED,IAAM,wBAAkE;AAAA,EACtE,aAAa;AAAA,EACb,gBAAgB,UACd,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ,KAAK,MAAM;AAC7D;AA6EA,IAAM,iBAAiB;AAEhB,SAAS,cACd,UAAmC,CAAC,GACnB;AA1JnB;AA2JE,QAAM,UAAU,sBAAqB,aAAQ,YAAR,YAAmB,cAAc;AACtE,QAAM,aAAa,MACjB;AAAA,IACE;AAAA,MACE,eAAe,UAAU,WAAW;AAAA,QAClC,QAAQ,QAAQ;AAAA,QAChB,yBAAyB;AAAA,QACzB,aAAa;AAAA,MACf,CAAC,CAAC;AAAA,MACF,GAAG,QAAQ;AAAA,IACb;AAAA,IACA,kBAAkB,OAAO;AAAA,EAC3B;AASF,QAAM,uBAAuB,CAC3B,WACA,eACuB;AAAA,IACvB,UAAU,WAAW,SAAS;AAAA,IAC9B,KAAK,CAAC,EAAE,KAAK,MAAM;AAEjB,UACE,cAAc,gBACd,uCAAW,SAAS,aACpB,EAAC,uCAAW,SAAS,cACrB;AACA,eAAO,GAAG,SAAS,MAAM,IAAI;AAAA,MAC/B;AACA,aAAO,GAAG,aAAa,OAAO,GAAG,IAAI;AAAA,IACvC;AAAA,IACA,SAAS;AAAA,IACT,OAAO,QAAQ;AAAA,EACjB;AAEA,QAAM,kBAAkB,CAAC,YAAiC;AACxD,UAAM,YAAY,QAAQ;AAC1B,QAAI,WAAW;AACb,UAAI,UAAU,SAAS,UAAU,GAAG;AAClC,eAAO,IAAI,kCAAkC,4BAAW,eAAe;AAAA,UACrE,GAAG,qBAAqB,QAAQ,SAAS;AAAA,UACzC,gBAAgB;AAAA;AAAA,UAEhB,cAAc;AAAA,UACd,2BAA2B;AAAA,QAC7B,CAAC;AAAA,MACH,WAAW,UAAU,SAAS,UAAU,GAAG;AACzC,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAEA,WAAO,IAAI,kCAAkC,4BAAW,QAAQ;AAAA,MAC9D,GAAG,qBAAqB,MAAM;AAAA,MAC9B,gBAAgB;AAAA,MAChB,cAAc;AAAA,MACd,2BAA2B;AAAA,IAC7B,CAAC;AAAA,EACH;AAEA,QAAM,uBAAuB,CAAC,YAAsC;AAClE,UAAM,YAAY,QAAQ;AAC1B,QAAI,CAAC,WAAW;AACd,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAEA,QAAI,CAAC,UAAU,SAAS,OAAO,GAAG;AAChC,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAIA,UAAM,QAAQ,IAAI,+BAA+B,4BAAW,cAAc;AAAA,MACxE,GAAG,qBAAqB,aAAa,SAAS;AAAA,MAC9C,gBAAgB;AAAA;AAAA;AAAA;AAAA,MAIhB,sBAAsB,QAAQ,oBAC1B,OAAO,oBACP;AAAA,IACN,CAAC;AAED,QAAI,CAAC,QAAQ,mBAAmB;AAC9B,aAAO;AAAA,IACT;AAIA,UAAM,oBAAoB,IAAI,QAAQ;AAAA,MACpC,UAAU,QAAQ,YAAY,OAAO;AAAA,MACrC,WAAW;AAAA,QACT,QAAQ,QAAQ;AAAA,QAChB,yBAAyB;AAAA,QACzB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAEA,UAAM,UAAU,OAAM,WAAU;AAxQpC,UAAAA;AAyQM,UAAI,CAAC,OAAO,UAAU,CAAC,MAAM,QAAQ,OAAO,MAAM,GAAG;AACnD,cAAM,IAAI,MAAM,2CAA2C;AAAA,MAC7D;AAEA,YAAM,WAAW,MAAM,kBAAkB;AAAA,QACvC,OAAO;AAAA;AAAA,QAEP,4BAAW;AAAA,MACb;AAEA,aAAO;AAAA,QACL,YAAY,SAAS,KAAK,IAAI,UAAQ,KAAK,SAAS;AAAA;AAAA;AAAA,QAGpD,OACE,SAAOA,MAAA,SAAS,UAAT,gBAAAA,IAAgB,kBAAiB,WACpC,EAAE,QAAQ,SAAS,MAAM,aAAa,IACtC;AAAA,QACN,UAAU,EAAE,SAAS,CAAC,GAAG,MAAM,SAAS;AAAA,QACxC,UAAU,CAAC;AAAA,MACb;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,QAAM,WAAW,CAAC,YAAiC,gBAAgB,OAAO;AAE1E,WAAS,uBAAuB;AAChC,WAAS,YAAY;AACrB,WAAS,gBAAgB;AACzB,WAAS,aAAa,CAAC,YAAoB;AACzC,UAAM,IAAI,iBAAiB,EAAE,SAAS,WAAW,aAAa,CAAC;AAAA,EACjE;AACA,WAAS,iBAAiB;AAC1B,WAAS,qBAAqB;AAC9B,SAAO;AACT;AAEO,IAAM,UAAU,cAAc;","names":["_a"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/baseten",
-  "version": "2.1.12",
+  "version": "2.1.14",
   "type": "module",
   "license": "Apache-2.0",
   "sideEffects": false,
@@ -29,9 +29,9 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/openai-compatible": "3.0.34",
+    "@ai-sdk/openai-compatible": "3.0.36",
     "@ai-sdk/provider": "4.0.7",
-    "@ai-sdk/provider-utils": "5.0.28"
+    "@ai-sdk/provider-utils": "5.0.29"
   },
   "devDependencies": {
     "@types/node": "22.19.19",

package/src/baseten-provider.ts

@@ -203,6 +203,7 @@ export function createBaseten(
           errorStructure: basetenErrorStructure,
           // Or stream_options.include_usage is omitted and streams report no usage.
           includeUsage: true,
+          supportsStructuredOutputs: true,
         });
       } else if (customURL.includes('/predict')) {
         throw new Error(
@@ -215,6 +216,7 @@ export function createBaseten(
       ...getCommonModelConfig('chat'),
       errorStructure: basetenErrorStructure,
       includeUsage: true,
+      supportsStructuredOutputs: true,
     });
   };