@tanstack/ai-gemini 0.0.3 → 0.1.0

package/dist/esm/{gemini-adapter.js → adapters/text.js}

@@ -1,16 +1,13 @@
-import { GoogleGenAI, FinishReason } from "@google/genai";
-import { BaseAdapter } from "@tanstack/ai";
-import { GEMINI_MODELS, GEMINI_EMBEDDING_MODELS } from "./model-meta.js";
-import { convertToolsToProviderFormat } from "./tools/tool-converter.js";
-class GeminiAdapter extends BaseAdapter {
-  constructor(config) {
-    super(config);
+import { FinishReason } from "@google/genai";
+import { BaseTextAdapter } from "@tanstack/ai/adapters";
+import { convertToolsToProviderFormat } from "../tools/tool-converter.js";
+import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "../utils/client.js";
+class GeminiTextAdapter extends BaseTextAdapter {
+  constructor(config, model) {
+    super({}, model);
+    this.kind = "text";
     this.name = "gemini";
-    this.models = GEMINI_MODELS;
-    this.embeddingModels = GEMINI_EMBEDDING_MODELS;
-    this.client = new GoogleGenAI({
-      apiKey: config.apiKey
-    });
+    this.client = createGeminiClient(config);
   }
   async *chatStream(options) {
     const mappedOptions = this.mapCommonOptionsToGemini(options);
@@ -21,7 +18,7 @@ class GeminiAdapter extends BaseAdapter {
       const timestamp = Date.now();
       yield {
         type: "error",
-        id: this.generateId(),
+        id: generateId(this.name),
         model: options.model,
         timestamp,
         error: {
@@ -30,99 +27,55 @@ class GeminiAdapter extends BaseAdapter {
       };
     }
   }
-  async summarize(options) {
-    const prompt = this.buildSummarizationPrompt(options, options.text);
-    const result = await this.client.models.generateContent({
-      model: options.model,
-      contents: [{ role: "user", parts: [{ text: prompt }] }],
-      config: {
-        temperature: 0.3,
-        maxOutputTokens: options.maxLength || 500
-      }
-    });
-    let summary = "";
-    if (result.candidates?.[0]?.content?.parts) {
-      const parts = result.candidates[0].content.parts;
-      for (const part of parts) {
-        if (part.text) {
-          summary += part.text;
+  /**
+   * Generate structured output using Gemini's native JSON response format.
+   * Uses responseMimeType: 'application/json' and responseSchema for structured output.
+   * The outputSchema is already JSON Schema (converted in the ai layer).
+   */
+  async structuredOutput(options) {
+    const { chatOptions, outputSchema } = options;
+    const mappedOptions = this.mapCommonOptionsToGemini(chatOptions);
+    try {
+      const result = await this.client.models.generateContent({
+        ...mappedOptions,
+        config: {
+          ...mappedOptions.config,
+          responseMimeType: "application/json",
+          responseSchema: outputSchema
         }
+      });
+      const rawText = this.extractTextFromResponse(result);
+      let parsed;
+      try {
+        parsed = JSON.parse(rawText);
+      } catch {
+        throw new Error(
+          `Failed to parse structured output as JSON. Content: ${rawText.slice(0, 200)}${rawText.length > 200 ? "..." : ""}`
+        );
       }
+      return {
+        data: parsed,
+        rawText
+      };
+    } catch (error) {
+      throw new Error(
+        error instanceof Error ? error.message : "An unknown error occurred during structured output generation."
+      );
     }
-    if (!summary && typeof result.text === "string") {
-      summary = result.text;
-    }
-    const promptTokens = this.estimateTokens(prompt);
-    const completionTokens = this.estimateTokens(summary);
-    return {
-      id: this.generateId(),
-      model: options.model,
-      summary,
-      usage: {
-        promptTokens,
-        completionTokens,
-        totalTokens: promptTokens + completionTokens
-      }
-    };
   }
-  async createEmbeddings(options) {
-    const inputs = Array.isArray(options.input) ? options.input : [options.input];
-    const result = await this.client.models.embedContent({
-      model: options.model,
-      contents: inputs
-    });
-    const embeddings = [];
-    if (result.embeddings && Array.isArray(result.embeddings)) {
-      for (const embedding of result.embeddings) {
-        if (embedding.values && Array.isArray(embedding.values)) {
-          embeddings.push(embedding.values);
-        } else if (Array.isArray(embedding)) {
-          embeddings.push(embedding);
+  /**
+   * Extract text content from a non-streaming response
+   */
+  extractTextFromResponse(response) {
+    let textContent = "";
+    if (response.candidates?.[0]?.content?.parts) {
+      for (const part of response.candidates[0].content.parts) {
+        if (part.text) {
+          textContent += part.text;
         }
       }
     }
-    const promptTokens = inputs.reduce(
-      (sum, input) => sum + this.estimateTokens(input),
-      0
-    );
-    return {
-      id: this.generateId(),
-      model: options.model || "gemini-embedding-001",
-      embeddings,
-      usage: {
-        promptTokens,
-        totalTokens: promptTokens
-      }
-    };
-  }
-  buildSummarizationPrompt(options, text) {
-    let prompt = "You are a professional summarizer. ";
-    switch (options.style) {
-      case "bullet-points":
-        prompt += "Provide a summary in bullet point format. ";
-        break;
-      case "paragraph":
-        prompt += "Provide a summary in paragraph format. ";
-        break;
-      case "concise":
-        prompt += "Provide a very concise summary in 1-2 sentences. ";
-        break;
-      default:
-        prompt += "Provide a clear and concise summary. ";
-    }
-    if (options.focus && options.focus.length > 0) {
-      prompt += `Focus on the following aspects: ${options.focus.join(", ")}. `;
-    }
-    prompt += `
-
-Text to summarize:
-${text}
-
-Summary:`;
-    return prompt;
-  }
-  estimateTokens(text) {
-    return Math.ceil(text.length / 4);
+    return textContent;
   }
   async *processStreamChunks(result, model) {
     const timestamp = Date.now();
@@ -137,7 +90,7 @@ Summary:`;
             accumulatedContent += part.text;
             yield {
               type: "content",
-              id: this.generateId(),
+              id: generateId(this.name),
               model,
               timestamp,
               delta: part.text,
@@ -169,7 +122,7 @@ Summary:`;
             }
             yield {
               type: "tool_call",
-              id: this.generateId(),
+              id: generateId(this.name),
               model,
               timestamp,
               toolCall: {
@@ -188,7 +141,7 @@ Summary:`;
         accumulatedContent += chunk.data;
         yield {
           type: "content",
-          id: this.generateId(),
+          id: generateId(this.name),
           model,
           timestamp,
           delta: chunk.data,
@@ -212,7 +165,7 @@ Summary:`;
                 });
                 yield {
                   type: "tool_call",
-                  id: this.generateId(),
+                  id: generateId(this.name),
                   model,
                   timestamp,
                   toolCall: {
@@ -232,7 +185,7 @@ Summary:`;
         if (finishReason === FinishReason.MAX_TOKENS) {
           yield {
             type: "error",
-            id: this.generateId(),
+            id: generateId(this.name),
             model,
             timestamp,
             error: {
@@ -242,13 +195,13 @@ Summary:`;
         }
         yield {
           type: "done",
-          id: this.generateId(),
+          id: generateId(this.name),
           model,
           timestamp,
           finishReason: toolCallMap.size > 0 ? "tool_calls" : "stop",
           usage: chunk.usageMetadata ? {
             promptTokens: chunk.usageMetadata.promptTokenCount ?? 0,
-            completionTokens: chunk.usageMetadata.thoughtsTokenCount ?? 0,
+            completionTokens: chunk.usageMetadata.candidatesTokenCount ?? 0,
             totalTokens: chunk.usageMetadata.totalTokenCount ?? 0
           } : void 0
         };
@@ -256,6 +209,18 @@ Summary:`;
     }
   }
   convertContentPartToGemini(part) {
+    const getDefaultFileType = (part2) => {
+      switch (part2) {
+        case "image":
+          return "image/jpeg";
+        case "audio":
+          return "audio/mp3";
+        case "video":
+          return "video/mp4";
+        case "document":
+          return "application/pdf";
+      }
+    };
     switch (part.type) {
       case "text":
         return { text: part.content };
@@ -268,14 +233,14 @@ Summary:`;
           return {
             inlineData: {
               data: part.source.value,
-              mimeType: metadata?.mimeType ?? "image/jpeg"
+              mimeType: metadata?.mimeType ?? getDefaultFileType(part.type)
             }
           };
         } else {
           return {
             fileData: {
               fileUri: part.source.value,
-              mimeType: metadata?.mimeType ?? "image/jpeg"
+              mimeType: metadata?.mimeType ?? getDefaultFileType(part.type)
             }
           };
         }
@@ -319,7 +284,6 @@ Summary:`;
         parts.push({
           functionResponse: {
             name: msg.toolCallId,
-            // Gemini uses function name here
             response: {
               content: msg.content || ""
             }
@@ -332,20 +296,16 @@ Summary:`;
       };
     });
   }
-  /**
-   * Maps common options to Gemini-specific format
-   * Handles translation of normalized options to Gemini's API format
-   */
   mapCommonOptionsToGemini(options) {
-    const providerOpts = options.providerOptions;
+    const providerOpts = options.modelOptions;
     const requestOptions = {
       model: options.model,
       contents: this.formatMessages(options.messages),
       config: {
         ...providerOpts,
-        temperature: options.options?.temperature,
-        topP: options.options?.topP,
-        maxOutputTokens: options.options?.maxTokens,
+        temperature: options.temperature,
+        topP: options.topP,
+        maxOutputTokens: options.maxTokens,
         systemInstruction: options.systemPrompts?.join("\n"),
         ...providerOpts?.generationConfig,
         tools: convertToolsToProviderFormat(options.tools)
@@ -354,22 +314,16 @@ Summary:`;
     return requestOptions;
   }
 }
-function createGemini(apiKey, config) {
-  return new GeminiAdapter({ apiKey, ...config });
+function createGeminiChat(model, apiKey, config) {
+  return new GeminiTextAdapter({ apiKey, ...config }, model);
 }
-function gemini(config) {
-  const env = typeof globalThis !== "undefined" && globalThis.window?.env ? globalThis.window.env : typeof process !== "undefined" ? process.env : void 0;
-  const key = env?.GOOGLE_API_KEY || env?.GEMINI_API_KEY;
-  if (!key) {
-    throw new Error(
-      "GOOGLE_API_KEY or GEMINI_API_KEY is required. Please set it in your environment variables or use createGemini(apiKey, config) instead."
-    );
-  }
-  return createGemini(key, config);
+function geminiText(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiChat(model, apiKey, config);
 }
 export {
-  GeminiAdapter,
-  createGemini,
-  gemini
+  GeminiTextAdapter,
+  createGeminiChat,
+  geminiText
 };
-//# sourceMappingURL=gemini-adapter.js.map
+//# sourceMappingURL=text.js.map

package/dist/esm/adapters/text.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"text.js","sources":["../../../src/adapters/text.ts"],"sourcesContent":["import { FinishReason } from '@google/genai'\nimport { BaseTextAdapter } from '@tanstack/ai/adapters'\nimport { convertToolsToProviderFormat } from '../tools/tool-converter'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport type {\n  GEMINI_MODELS,\n  GeminiChatModelProviderOptionsByName,\n  GeminiModelInputModalitiesByName,\n} from '../model-meta'\nimport type {\n  StructuredOutputOptions,\n  StructuredOutputResult,\n} from '@tanstack/ai/adapters'\nimport type {\n  GenerateContentParameters,\n  GenerateContentResponse,\n  GoogleGenAI,\n  Part,\n} from '@google/genai'\nimport type {\n  ContentPart,\n  Modality,\n  ModelMessage,\n  StreamChunk,\n  TextOptions,\n} from '@tanstack/ai'\nimport type { ExternalTextProviderOptions } from '../text/text-provider-options'\nimport type {\n  GeminiAudioMetadata,\n  GeminiDocumentMetadata,\n  GeminiImageMetadata,\n  GeminiMessageMetadataByModality,\n  GeminiVideoMetadata,\n} from '../message-types'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini text adapter\n */\nexport interface GeminiTextConfig extends GeminiClientConfig {}\n\n/**\n * Gemini-specific provider options for text/chat\n */\nexport type GeminiTextProviderOptions = ExternalTextProviderOptions\n\n// ===========================\n// Type Resolution Helpers\n// ===========================\n\n/**\n * Resolve provider options for a specific model.\n * If the model has explicit options in the map, use those; otherwise use base options.\n */\ntype ResolveProviderOptions<TModel extends string> =\n  TModel extends keyof GeminiChatModelProviderOptionsByName\n    ? GeminiChatModelProviderOptionsByName[TModel]\n    : GeminiTextProviderOptions\n\n/**\n * Resolve input modalities for a specific model.\n * If the model has explicit modalities in the map, use those; otherwise use all modalities.\n */\ntype ResolveInputModalities<TModel extends string> =\n  TModel extends keyof GeminiModelInputModalitiesByName\n    ? GeminiModelInputModalitiesByName[TModel]\n    : readonly ['text', 'image', 'audio', 'video', 'document']\n\n// ===========================\n// Adapter Implementation\n// ===========================\n\n/**\n * Gemini Text (Chat) Adapter\n *\n * Tree-shakeable adapter for Gemini chat/text completion functionality.\n * Import only what you need for smaller bundle sizes.\n */\nexport class GeminiTextAdapter<\n  TModel extends (typeof GEMINI_MODELS)[number],\n  TProviderOptions extends object = ResolveProviderOptions<TModel>,\n  TInputModalities extends ReadonlyArray<Modality> =\n    ResolveInputModalities<TModel>,\n> extends BaseTextAdapter<\n  TModel,\n  TProviderOptions,\n  TInputModalities,\n  GeminiMessageMetadataByModality\n> {\n  readonly kind = 'text' as const\n  readonly name = 'gemini' as const\n\n  private client: GoogleGenAI\n\n  constructor(config: GeminiTextConfig, model: TModel) {\n    super({}, model)\n    this.client = createGeminiClient(config)\n  }\n\n  async *chatStream(\n    options: TextOptions<GeminiTextProviderOptions>,\n  ): AsyncIterable<StreamChunk> {\n    const mappedOptions = this.mapCommonOptionsToGemini(options)\n\n    try {\n      const result =\n        await this.client.models.generateContentStream(mappedOptions)\n\n      yield* this.processStreamChunks(result, options.model)\n    } catch (error) {\n      const timestamp = Date.now()\n      yield {\n        type: 'error',\n        id: generateId(this.name),\n        model: options.model,\n        timestamp,\n        error: {\n          message:\n            error instanceof Error\n              ? error.message\n              : 'An unknown error occurred during the chat stream.',\n        },\n      }\n    }\n  }\n\n  /**\n   * Generate structured output using Gemini's native JSON response format.\n   * Uses responseMimeType: 'application/json' and responseSchema for structured output.\n   * The outputSchema is already JSON Schema (converted in the ai layer).\n   */\n  async structuredOutput(\n    options: StructuredOutputOptions<GeminiTextProviderOptions>,\n  ): Promise<StructuredOutputResult<unknown>> {\n    const { chatOptions, outputSchema } = options\n\n    const mappedOptions = this.mapCommonOptionsToGemini(chatOptions)\n\n    try {\n      // Add structured output configuration\n      const result = await this.client.models.generateContent({\n        ...mappedOptions,\n        config: {\n          ...mappedOptions.config,\n          responseMimeType: 'application/json',\n          responseSchema: outputSchema,\n        },\n      })\n\n      // Extract text content from the response\n      const rawText = this.extractTextFromResponse(result)\n\n      // Parse the JSON response\n      let parsed: unknown\n      try {\n        parsed = JSON.parse(rawText)\n      } catch {\n        throw new Error(\n          `Failed to parse structured output as JSON. Content: ${rawText.slice(0, 200)}${rawText.length > 200 ? '...' : ''}`,\n        )\n      }\n\n      return {\n        data: parsed,\n        rawText,\n      }\n    } catch (error) {\n      throw new Error(\n        error instanceof Error\n          ? error.message\n          : 'An unknown error occurred during structured output generation.',\n      )\n    }\n  }\n\n  /**\n   * Extract text content from a non-streaming response\n   */\n  private extractTextFromResponse(response: GenerateContentResponse): string {\n    let textContent = ''\n\n    if (response.candidates?.[0]?.content?.parts) {\n      for (const part of response.candidates[0].content.parts) {\n        if (part.text) {\n          textContent += part.text\n        }\n      }\n    }\n\n    return textContent\n  }\n\n  private async *processStreamChunks(\n    result: AsyncGenerator<GenerateContentResponse, unknown, unknown>,\n    model: string,\n  ): AsyncIterable<StreamChunk> {\n    const timestamp = Date.now()\n    let accumulatedContent = ''\n    const toolCallMap = new Map<\n      string,\n      { name: string; args: string; index: number }\n    >()\n    let nextToolIndex = 0\n\n    for await (const chunk of result) {\n      if (chunk.candidates?.[0]?.content?.parts) {\n        const parts = chunk.candidates[0].content.parts\n\n        for (const part of parts) {\n          if (part.text) {\n            accumulatedContent += part.text\n            yield {\n              type: 'content',\n              id: generateId(this.name),\n              model,\n              timestamp,\n              delta: part.text,\n              content: accumulatedContent,\n              role: 'assistant',\n            }\n          }\n\n          const functionCall = part.functionCall\n          if (functionCall) {\n            const toolCallId =\n              functionCall.name || `call_${Date.now()}_${nextToolIndex}`\n            const functionArgs = functionCall.args || {}\n\n            let toolCallData = toolCallMap.get(toolCallId)\n            if (!toolCallData) {\n              toolCallData = {\n                name: functionCall.name || '',\n                args:\n                  typeof functionArgs === 'string'\n                    ? functionArgs\n                    : JSON.stringify(functionArgs),\n                index: nextToolIndex++,\n              }\n              toolCallMap.set(toolCallId, toolCallData)\n            } else {\n              try {\n                const existingArgs = JSON.parse(toolCallData.args)\n                const newArgs =\n                  typeof functionArgs === 'string'\n                    ? JSON.parse(functionArgs)\n                    : functionArgs\n                const mergedArgs = { ...existingArgs, ...newArgs }\n                toolCallData.args = JSON.stringify(mergedArgs)\n              } catch {\n                toolCallData.args =\n                  typeof functionArgs === 'string'\n                    ? functionArgs\n                    : JSON.stringify(functionArgs)\n              }\n            }\n\n            yield {\n              type: 'tool_call',\n              id: generateId(this.name),\n              model,\n              timestamp,\n              toolCall: {\n                id: toolCallId,\n                type: 'function',\n                function: {\n                  name: toolCallData.name,\n                  arguments: toolCallData.args,\n                },\n              },\n              index: toolCallData.index,\n            }\n          }\n        }\n      } else if (chunk.data) {\n        accumulatedContent += chunk.data\n        yield {\n          type: 'content',\n          id: generateId(this.name),\n          model,\n          timestamp,\n          delta: chunk.data,\n          content: accumulatedContent,\n          role: 'assistant',\n        }\n      }\n\n      if (chunk.candidates?.[0]?.finishReason) {\n        const finishReason = chunk.candidates[0].finishReason\n\n        if (finishReason === FinishReason.UNEXPECTED_TOOL_CALL) {\n          if (chunk.candidates[0].content?.parts) {\n            for (const part of chunk.candidates[0].content.parts) {\n              const functionCall = part.functionCall\n              if (functionCall) {\n                const toolCallId =\n                  functionCall.name || `call_${Date.now()}_${nextToolIndex}`\n                const functionArgs = functionCall.args || {}\n\n                toolCallMap.set(toolCallId, {\n                  name: functionCall.name || '',\n                  args:\n                    typeof functionArgs === 'string'\n                      ? functionArgs\n                      : JSON.stringify(functionArgs),\n                  index: nextToolIndex++,\n                })\n\n                yield {\n                  type: 'tool_call',\n                  id: generateId(this.name),\n                  model,\n                  timestamp,\n                  toolCall: {\n                    id: toolCallId,\n                    type: 'function',\n                    function: {\n                      name: functionCall.name || '',\n                      arguments:\n                        typeof functionArgs === 'string'\n                          ? functionArgs\n                          : JSON.stringify(functionArgs),\n                    },\n                  },\n                  index: nextToolIndex - 1,\n                }\n              }\n            }\n          }\n        }\n        if (finishReason === FinishReason.MAX_TOKENS) {\n          yield {\n            type: 'error',\n            id: generateId(this.name),\n            model,\n            timestamp,\n            error: {\n              message:\n                'The response was cut off because the maximum token limit was reached.',\n            },\n          }\n        }\n\n        yield {\n          type: 'done',\n          id: generateId(this.name),\n          model,\n          timestamp,\n          finishReason: toolCallMap.size > 0 ? 'tool_calls' : 'stop',\n          usage: chunk.usageMetadata\n            ? {\n                promptTokens: chunk.usageMetadata.promptTokenCount ?? 0,\n                completionTokens: chunk.usageMetadata.candidatesTokenCount ?? 0,\n                totalTokens: chunk.usageMetadata.totalTokenCount ?? 0,\n              }\n            : undefined,\n        }\n      }\n    }\n  }\n\n  private convertContentPartToGemini(part: ContentPart): Part {\n    const getDefaultFileType = (\n      part: 'image' | 'audio' | 'video' | 'document',\n    ) => {\n      switch (part) {\n        case 'image':\n          return 'image/jpeg'\n        case 'audio':\n          return 'audio/mp3'\n        case 'video':\n          return 'video/mp4'\n        case 'document':\n          return 'application/pdf'\n      }\n    }\n    switch (part.type) {\n      case 'text':\n        return { text: part.content }\n      case 'image':\n      case 'audio':\n      case 'video':\n      case 'document': {\n        const metadata = part.metadata as\n          | GeminiDocumentMetadata\n          | GeminiImageMetadata\n          | GeminiVideoMetadata\n          | GeminiAudioMetadata\n          | undefined\n        if (part.source.type === 'data') {\n          return {\n            inlineData: {\n              data: part.source.value,\n              mimeType: metadata?.mimeType ?? getDefaultFileType(part.type),\n            },\n          }\n        } else {\n          return {\n            fileData: {\n              fileUri: part.source.value,\n              mimeType: metadata?.mimeType ?? getDefaultFileType(part.type),\n            },\n          }\n        }\n      }\n      default: {\n        const _exhaustiveCheck: never = part\n        throw new Error(\n          `Unsupported content part type: ${(_exhaustiveCheck as ContentPart).type}`,\n        )\n      }\n    }\n  }\n\n  private formatMessages(\n    messages: Array<ModelMessage>,\n  ): GenerateContentParameters['contents'] {\n    return messages.map((msg) => {\n      const role: 'user' | 'model' = msg.role === 'assistant' ? 'model' : 'user'\n      const parts: Array<Part> = []\n\n      if (Array.isArray(msg.content)) {\n        for (const contentPart of msg.content) {\n          parts.push(this.convertContentPartToGemini(contentPart))\n        }\n      } else if (msg.content) {\n        parts.push({ text: msg.content })\n      }\n\n      if (msg.role === 'assistant' && msg.toolCalls?.length) {\n        for (const toolCall of msg.toolCalls) {\n          let parsedArgs: Record<string, unknown> = {}\n          try {\n            parsedArgs = toolCall.function.arguments\n              ? (JSON.parse(toolCall.function.arguments) as Record<\n                  string,\n                  unknown\n                >)\n              : {}\n          } catch {\n            parsedArgs = toolCall.function.arguments as unknown as Record<\n              string,\n              unknown\n            >\n          }\n\n          parts.push({\n            functionCall: {\n              name: toolCall.function.name,\n              args: parsedArgs,\n            },\n          })\n        }\n      }\n\n      if (msg.role === 'tool' && msg.toolCallId) {\n        parts.push({\n          functionResponse: {\n            name: msg.toolCallId,\n            response: {\n              content: msg.content || '',\n            },\n          },\n        })\n      }\n\n      return {\n        role,\n        parts: parts.length > 0 ? parts : [{ text: '' }],\n      }\n    })\n  }\n\n  private mapCommonOptionsToGemini(options: TextOptions) {\n    const providerOpts = options.modelOptions\n    const requestOptions: GenerateContentParameters = {\n      model: options.model,\n      contents: this.formatMessages(options.messages),\n      config: {\n        ...providerOpts,\n        temperature: options.temperature,\n        topP: options.topP,\n        maxOutputTokens: options.maxTokens,\n        systemInstruction: options.systemPrompts?.join('\\n'),\n        ...((providerOpts as Record<string, unknown> | undefined)\n          ?.generationConfig as Record<string, unknown> | undefined),\n        tools: convertToolsToProviderFormat(options.tools),\n      },\n    }\n\n    return requestOptions\n  }\n}\n\n/**\n * Creates a Gemini text adapter with explicit API key.\n * Type resolution happens here at the call site.\n */\nexport function createGeminiChat<TModel extends (typeof GEMINI_MODELS)[number]>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiTextConfig, 'apiKey'>,\n): GeminiTextAdapter<\n  TModel,\n  ResolveProviderOptions<TModel>,\n  ResolveInputModalities<TModel>\n> {\n  return new GeminiTextAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini text adapter with automatic API key detection.\n * Type resolution happens here at the call site.\n */\nexport function geminiText<TModel extends (typeof GEMINI_MODELS)[number]>(\n  model: TModel,\n  config?: Omit<GeminiTextConfig, 'apiKey'>,\n): GeminiTextAdapter<\n  TModel,\n  ResolveProviderOptions<TModel>,\n  ResolveInputModalities<TModel>\n> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiChat(model, apiKey, config)\n}\n"],"names":["part"],"mappings":";;;;AAkFO,MAAM,0BAKH,gBAKR;AAAA,EAMA,YAAY,QAA0B,OAAe;AACnD,UAAM,CAAA,GAAI,KAAK;AANjB,SAAS,OAAO;AAChB,SAAS,OAAO;AAMd,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,OAAO,WACL,SAC4B;AAC5B,UAAM,gBAAgB,KAAK,yBAAyB,OAAO;AAE3D,QAAI;AACF,YAAM,SACJ,MAAM,KAAK,OAAO,OAAO,sBAAsB,aAAa;AAE9D,aAAO,KAAK,oBAAoB,QAAQ,QAAQ,KAAK;AAAA,IACvD,SAAS,OAAO;AACd,YAAM,YAAY,KAAK,IAAA;AACvB,YAAM;AAAA,QACJ,MAAM;AAAA,QACN,IAAI,WAAW,KAAK,IAAI;AAAA,QACxB,OAAO,QAAQ;AAAA,QACf;AAAA,QACA,OAAO;AAAA,UACL,SACE,iBAAiB,QACb,MAAM,UACN;AAAA,QAAA;AAAA,MACR;AAAA,IAEJ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBACJ,SAC0C;AAC1C,UAAM,EAAE,aAAa,aAAA,IAAiB;AAEtC,UAAM,gBAAgB,KAAK,yBAAyB,WAAW;AAE/D,QAAI;AAEF,YAAM,SAAS,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,QACtD,GAAG;AAAA,QACH,QAAQ;AAAA,UACN,GAAG,cAAc;AAAA,UACjB,kBAAkB;AAAA,UAClB,gBAAgB;AAAA,QAAA;AAAA,MAClB,CACD;AAGD,YAAM,UAAU,KAAK,wBAAwB,MAAM;AAGnD,UAAI;AACJ,UAAI;AACF,iBAAS,KAAK,MAAM,OAAO;AAAA,MAC7B,QAAQ;AACN,cAAM,IAAI;AAAA,UACR,uDAAuD,QAAQ,MAAM,GAAG,GAAG,CAAC,GAAG,QAAQ,SAAS,MAAM,QAAQ,EAAE;AAAA,QAAA;AAAA,MAEpH;AAEA,aAAO;AAAA,QACL,MAAM;AAAA,QACN;AAAA,MAAA;AAAA,IAEJ,SAAS,OAAO;AACd,YAAM,IAAI;AAAA,QACR,iBAAiB,QACb,MAAM,UACN;AAAA,MAAA;AAAA,IAER;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKQ,wBAAwB,UAA2C;AACzE,QAAI,cAAc;AAElB,QAAI,SAAS,aAAa,CAAC,GAAG,SAAS,OAAO;AAC5C,iBAAW,QAAQ,SAAS,WAAW,CAAC,EAAE,QAAQ,OAAO;AACvD,YAAI,KAAK,MAAM;AACb,yBAAe,KAAK;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAAA,EAEA,OAAe,oBACb,QACA,OAC4B;AAC5B,UAAM,YAAY,KAAK,IAAA;AACvB,QAAI,qBAAqB;AACzB,UAAM,kCAAkB,IAAA;AAIxB,QAAI,gBAAgB;AAEpB,qBAAiB,SAAS,QAAQ;AAChC,UAAI,MAAM,aAAa,CAAC,GAAG,SAAS,OAAO;AACzC,cAAM,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ;AAE1C,mBAAW,QAAQ,OAAO;AACxB,cAAI,KAAK,MAAM;AACb,kCAAsB,KAAK;AAC3B,kBAAM;AAAA,cACJ,MAAM;AAAA,cACN,IAAI,WAAW,KAAK,IAAI;AAAA,cACxB;AAAA,cACA;AAAA,cACA,OAAO,KAAK;AAAA,cACZ,SAAS;AAAA,cACT,MAAM;AAAA,YAAA;AAAA,UAEV;AAEA,gBAAM,eAAe,KAAK;AAC1B,cAAI,cAAc;AAChB,kBAAM,aACJ,aAAa,QAAQ,QAAQ,KAAK,IAAA,CAAK,IAAI,aAAa;AAC1D,kBAAM,eAAe,aAAa,QAAQ,CAAA;AAE1C,gBAAI,eAAe,YAAY,IAAI,UAAU;AAC7C,gBAAI,CAAC,cAAc;AACjB,6BAAe;AAAA,gBACb,MAAM,aAAa,QAAQ;AAAA,gBAC3B,MACE,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,gBACjC,OAAO;AAAA,cAAA;AAET,0BAAY,IAAI,YAAY,YAAY;AAAA,YAC1C,OAAO;AACL,kBAAI;AACF,sBAAM,eAAe,KAAK,MAAM,aAAa,IAAI;AACjD,sBAAM,UACJ,OAAO,iBAAiB,WACpB,KAAK,MAAM,YAAY,IACvB;AACN,sBAAM,aAAa,EAAE,GAAG,cAAc,GAAG,QAAA;AACzC,6BAAa,OAAO,KAAK,UAAU,UAAU;AAAA,cAC/C,QAAQ;AACN,6BAAa,OACX,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,cACnC;AAAA,YACF;AAEA,kBAAM;AAAA,cACJ,MAAM;AAAA,cACN,IAAI,WAAW,KAAK,IAAI;AAAA,cACxB;AAAA,cACA;AAAA,cACA,UAAU;AAAA,gBACR,IAAI;AAAA,gBACJ,MAAM;AAAA,gBACN,UAAU;AAAA,kBACR,MAAM,aAAa;AAAA,kBACnB,WAAW,aAAa;AAAA,gBAAA;AAAA,cAC1B;AAAA,cAEF,OAAO,aAAa;AAAA,YAAA;AAAA,UAExB;AAAA,QACF;AAAA,MACF,WAAW,MAAM,MAAM;AACrB,8BAAsB,MAAM;AAC5B,cAAM;AAAA,UACJ,MAAM;AAAA,UACN,IAAI,WAAW,KAAK,IAAI;AAAA,UACxB;AAAA,UACA;AAAA,UACA,OAAO,MAAM;AAAA,UACb,SAAS;AAAA,UACT,MAAM;AAAA,QAAA;AAAA,MAEV;AAEA,UAAI,MAAM,aAAa,CAAC,GAAG,cAAc;AACvC,cAAM,eAAe,MAAM,WAAW,CAAC,EAAE;AAEzC,YAAI,iBAAiB,aAAa,sBAAsB;AACtD,cAAI,MAAM,WAAW,CAAC,EAAE,SAAS,OAAO;AACtC,uBAAW,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ,OAAO;AACpD,oBAAM,eAAe,KAAK;AAC1B,kBAAI,cAAc;AAChB,sBAAM,aACJ,aAAa,QAAQ,QAAQ,KAAK,IAAA,CAAK,IAAI,aAAa;AAC1D,sBAAM,eAAe,aAAa,QAAQ,CAAA;AAE1C,4BAAY,IAAI,YAAY;AAAA,kBAC1B,MAAM,aAAa,QAAQ;AAAA,kBAC3B,MACE,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,kBACjC,OAAO;AAAA,gBAAA,CACR;AAED,sBAAM;AAAA,kBACJ,MAAM;AAAA,kBACN,IAAI,WAAW,KAAK,IAAI;AAAA,kBACxB;AAAA,kBACA;AAAA,kBACA,UAAU;AAAA,oBACR,IAAI;AAAA,oBACJ,MAAM;AAAA,oBACN,UAAU;AAAA,sBACR,MAAM,aAAa,QAAQ;AAAA,sBAC3B,WACE,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,oBAAA;AAAA,kBACnC;AAAA,kBAEF,OAAO,gBAAgB;AAAA,gBAAA;AAAA,cAE3B;AAAA,YACF;AAAA,UACF;AAAA,QACF;AACA,YAAI,iBAAiB,aAAa,YAAY;AAC5C,gBAAM;AAAA,YACJ,MAAM;AAAA,YACN,IAAI,WAAW,KAAK,IAAI;AAAA,YACxB;AAAA,YACA;AAAA,YACA,OAAO;AAAA,cACL,SACE;AAAA,YAAA;AAAA,UACJ;AAAA,QAEJ;AAEA,cAAM;AAAA,UACJ,MAAM;AAAA,UACN,IAAI,WAAW,KAAK,IAAI;AAAA,UACxB;AAAA,UACA;AAAA,UACA,cAAc,YAAY,OAAO,IAAI,eAAe;AAAA,UACpD,OAAO,MAAM,gBACT;AAAA,YACE,cAAc,MAAM,cAAc,oBAAoB;AAAA,YACtD,kBAAkB,MAAM,cAAc,wBAAwB;AAAA,YAC9D,aAAa,MAAM,cAAc,mBAAmB;AAAA,UAAA,IAEtD;AAAA,QAAA;AAAA,MAER;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BAA2B,MAAyB;AAC1D,UAAM,qBAAqB,CACzBA,UACG;AACH,cAAQA,OAAAA;AAAAA,QACN,KAAK;AACH,iBAAO;AAAA,QACT,KAAK;AACH,iBAAO;AAAA,QACT,KAAK;AACH,iBAAO;AAAA,QACT,KAAK;AACH,iBAAO;AAAA,MAAA;AAAA,IAEb;AACA,YAAQ,KAAK,MAAA;AAAA,MACX,KAAK;AACH,eAAO,EAAE,MAAM,KAAK,QAAA;AAAA,MACtB,KAAK;AAAA,MACL,KAAK;AAAA,MACL,KAAK;AAAA,MACL,KAAK,YAAY;AACf,cAAM,WAAW,KAAK;AAMtB,YAAI,KAAK,OAAO,SAAS,QAAQ;AAC/B,iBAAO;AAAA,YACL,YAAY;AAAA,cACV,MAAM,KAAK,OAAO;AAAA,cAClB,UAAU,UAAU,YAAY,mBAAmB,KAAK,IAAI;AAAA,YAAA;AAAA,UAC9D;AAAA,QAEJ,OAAO;AACL,iBAAO;AAAA,YACL,UAAU;AAAA,cACR,SAAS,KAAK,OAAO;AAAA,cACrB,UAAU,UAAU,YAAY,mBAAmB,KAAK,IAAI;AAAA,YAAA;AAAA,UAC9D;AAAA,QAEJ;AAAA,MACF;AAAA,MACA,SAAS;AACP,cAAM,mBAA0B;AAChC,cAAM,IAAI;AAAA,UACR,kCAAmC,iBAAiC,IAAI;AAAA,QAAA;AAAA,MAE5E;AAAA,IAAA;AAAA,EAEJ;AAAA,EAEQ,eACN,UACuC;AACvC,WAAO,SAAS,IAAI,CAAC,QAAQ;AAC3B,YAAM,OAAyB,IAAI,SAAS,cAAc,UAAU;AACpE,YAAM,QAAqB,CAAA;AAE3B,UAAI,MAAM,QAAQ,IAAI,OAAO,GAAG;AAC9B,mBAAW,eAAe,IAAI,SAAS;AACrC,gBAAM,KAAK,KAAK,2BAA2B,WAAW,CAAC;AAAA,QACzD;AAAA,MACF,WAAW,IAAI,SAAS;AACtB,cAAM,KAAK,EAAE,MAAM,IAAI,SAAS;AAAA,MAClC;AAEA,UAAI,IAAI,SAAS,eAAe,IAAI,WAAW,QAAQ;AACrD,mBAAW,YAAY,IAAI,WAAW;AACpC,cAAI,aAAsC,CAAA;AAC1C,cAAI;AACF,yBAAa,SAAS,SAAS,YAC1B,KAAK,MAAM,SAAS,SAAS,SAAS,IAIvC,CAAA;AAAA,UACN,QAAQ;AACN,yBAAa,SAAS,SAAS;AAAA,UAIjC;AAEA,gBAAM,KAAK;AAAA,YACT,cAAc;AAAA,cACZ,MAAM,SAAS,SAAS;AAAA,cACxB,MAAM;AAAA,YAAA;AAAA,UACR,CACD;AAAA,QACH;AAAA,MACF;AAEA,UAAI,IAAI,SAAS,UAAU,IAAI,YAAY;AACzC,cAAM,KAAK;AAAA,UACT,kBAAkB;AAAA,YAChB,MAAM,IAAI;AAAA,YACV,UAAU;AAAA,cACR,SAAS,IAAI,WAAW;AAAA,YAAA;AAAA,UAC1B;AAAA,QACF,CACD;AAAA,MACH;AAEA,aAAO;AAAA,QACL;AAAA,QACA,OAAO,MAAM,SAAS,IAAI,QAAQ,CAAC,EAAE,MAAM,GAAA,CAAI;AAAA,MAAA;AAAA,IAEnD,CAAC;AAAA,EACH;AAAA,EAEQ,yBAAyB,SAAsB;AACrD,UAAM,eAAe,QAAQ;AAC7B,UAAM,iBAA4C;AAAA,MAChD,OAAO,QAAQ;AAAA,MACf,UAAU,KAAK,eAAe,QAAQ,QAAQ;AAAA,MAC9C,QAAQ;AAAA,QACN,GAAG;AAAA,QACH,aAAa,QAAQ;AAAA,QACrB,MAAM,QAAQ;AAAA,QACd,iBAAiB,QAAQ;AAAA,QACzB,mBAAmB,QAAQ,eAAe,KAAK,IAAI;AAAA,QACnD,GAAK,cACD;AAAA,QACJ,OAAO,6BAA6B,QAAQ,KAAK;AAAA,MAAA;AAAA,IACnD;AAGF,WAAO;AAAA,EACT;AACF;AAMO,SAAS,iBACd,OACA,QACA,QAKA;AACA,SAAO,IAAI,kBAAkB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC3D;AAMO,SAAS,WACd,OACA,QAKA;AACA,QAAM,SAAS,uBAAA;AACf,SAAO,iBAAiB,OAAO,QAAQ,MAAM;AAC/C;"}

package/dist/esm/adapters/tts.d.ts

@@ -0,0 +1,129 @@
+import { BaseTTSAdapter } from '@tanstack/ai/adapters';
+import { GEMINI_TTS_MODELS, GeminiTTSVoice } from '../model-meta.js';
+import { TTSOptions, TTSResult } from '@tanstack/ai';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Provider-specific options for Gemini TTS
+ *
+ * @experimental Gemini TTS is an experimental feature.
+ * @see https://ai.google.dev/gemini-api/docs/speech-generation
+ */
+export interface GeminiTTSProviderOptions {
+    /**
+     * Voice configuration for TTS.
+     * Choose from 30 available voices with different characteristics.
+     */
+    voiceConfig?: {
+        prebuiltVoiceConfig?: {
+            /**
+             * The voice name to use for speech synthesis.
+             * @see https://ai.google.dev/gemini-api/docs/speech-generation#voices
+             */
+            voiceName?: GeminiTTSVoice;
+        };
+    };
+    /**
+     * System instruction for controlling speech style.
+     * Use natural language to describe the desired speaking style,
+     * pace, tone, accent, or other characteristics.
+     *
+     * @example "Speak slowly and calmly, as if telling a bedtime story"
+     * @example "Use an upbeat, enthusiastic tone with moderate pace"
+     * @example "Speak with a British accent"
+     */
+    systemInstruction?: string;
+    /**
+     * Language code hint for the speech synthesis.
+     * Gemini TTS supports 24 languages and can auto-detect,
+     * but you can provide a hint for better results.
+     *
+     * @example "en-US" for American English
+     * @example "es-ES" for Spanish (Spain)
+     * @example "ja-JP" for Japanese
+     */
+    languageCode?: string;
+}
+/**
+ * Configuration for Gemini TTS adapter
+ *
+ * @experimental Gemini TTS is an experimental feature.
+ */
+export interface GeminiTTSConfig extends GeminiClientConfig {
+}
+/** Model type for Gemini TTS */
+export type GeminiTTSModel = (typeof GEMINI_TTS_MODELS)[number];
+/**
+ * Gemini Text-to-Speech Adapter
+ *
+ * Tree-shakeable adapter for Gemini TTS functionality.
+ *
+ * **IMPORTANT**: Gemini TTS uses the Live API (WebSocket-based) which requires
+ * different handling than traditional REST APIs. This adapter provides a
+ * simplified interface but may have limitations.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * Models:
+ * - gemini-2.5-flash-preview-tts
+ */
+export declare class GeminiTTSAdapter<TModel extends GeminiTTSModel> extends BaseTTSAdapter<TModel, GeminiTTSProviderOptions> {
+    readonly name: "gemini";
+    private client;
+    constructor(config: GeminiTTSConfig, model: TModel);
+    /**
+     * Generate speech from text using Gemini's TTS model.
+     *
+     * @experimental This implementation is experimental and may change.
+     * @see https://ai.google.dev/gemini-api/docs/speech-generation
+     */
+    generateSpeech(options: TTSOptions<GeminiTTSProviderOptions>): Promise<TTSResult>;
+}
+/**
+ * Creates a Gemini TTS adapter with explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')
+ * @param apiKey - Your Google API key
+ * @param config - Optional additional configuration
+ * @returns Configured Gemini TTS adapter instance with resolved types
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiSpeech('gemini-2.5-flash-preview-tts', "your-api-key");
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Hello, world!'
+ * });
+ * ```
+ */
+export declare function createGeminiSpeech<TModel extends GeminiTTSModel>(model: TModel, apiKey: string, config?: Omit<GeminiTTSConfig, 'apiKey'>): GeminiTTSAdapter<TModel>;
+/**
+ * Creates a Gemini speech adapter with automatic API key detection from environment variables.
+ * Type resolution happens here at the call site.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:
+ * - `process.env` (Node.js)
+ * - `window.env` (Browser with injected env)
+ *
+ * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Gemini speech adapter instance with resolved types
+ * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment
+ *
+ * @example
+ * ```typescript
+ * // Automatically uses GOOGLE_API_KEY from environment
+ * const adapter = geminiSpeech('gemini-2.5-flash-preview-tts');
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Welcome to TanStack AI!'
+ * });
+ * ```
+ */
+export declare function geminiSpeech<TModel extends GeminiTTSModel>(model: TModel, config?: Omit<GeminiTTSConfig, 'apiKey'>): GeminiTTSAdapter<TModel>;

package/dist/esm/adapters/tts.js

@@ -0,0 +1,78 @@
+import { BaseTTSAdapter } from "@tanstack/ai/adapters";
+import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "../utils/client.js";
+class GeminiTTSAdapter extends BaseTTSAdapter {
+  constructor(config, model) {
+    super(config, model);
+    this.name = "gemini";
+    this.client = createGeminiClient(config);
+  }
+  /**
+   * Generate speech from text using Gemini's TTS model.
+   *
+   * @experimental This implementation is experimental and may change.
+   * @see https://ai.google.dev/gemini-api/docs/speech-generation
+   */
+  async generateSpeech(options) {
+    const { model, text, modelOptions } = options;
+    const voiceConfig = modelOptions?.voiceConfig || {
+      prebuiltVoiceConfig: {
+        voiceName: "Kore"
+      }
+    };
+    const response = await this.client.models.generateContent({
+      model,
+      contents: [
+        {
+          role: "user",
+          parts: [{ text }]
+        }
+      ],
+      config: {
+        responseModalities: ["AUDIO"],
+        speechConfig: {
+          voiceConfig,
+          ...modelOptions?.languageCode && {
+            languageCode: modelOptions.languageCode
+          }
+        }
+      },
+      ...modelOptions?.systemInstruction && {
+        systemInstruction: modelOptions.systemInstruction
+      }
+    });
+    const candidate = response.candidates?.[0];
+    const parts = candidate?.content?.parts;
+    if (!parts || parts.length === 0) {
+      throw new Error("No audio output received from Gemini TTS");
+    }
+    const audioPart = parts.find(
+      (part) => part.inlineData?.mimeType?.startsWith("audio/")
+    );
+    if (!audioPart || !audioPart.inlineData || !audioPart.inlineData.data) {
+      throw new Error("No audio data in Gemini TTS response");
+    }
+    const audioBase64 = audioPart.inlineData.data;
+    const mimeType = audioPart.inlineData.mimeType || "audio/wav";
+    const format = mimeType.split("/")[1] || "wav";
+    return {
+      id: generateId(this.name),
+      model,
+      audio: audioBase64,
+      format,
+      contentType: mimeType
+    };
+  }
+}
+function createGeminiSpeech(model, apiKey, config) {
+  return new GeminiTTSAdapter({ apiKey, ...config }, model);
+}
+function geminiSpeech(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiSpeech(model, apiKey, config);
+}
+export {
+  GeminiTTSAdapter,
+  createGeminiSpeech,
+  geminiSpeech
+};
+//# sourceMappingURL=tts.js.map

package/dist/esm/adapters/tts.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tts.js","sources":["../../../src/adapters/tts.ts"],"sourcesContent":["import { BaseTTSAdapter } from '@tanstack/ai/adapters'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport type { GEMINI_TTS_MODELS, GeminiTTSVoice } from '../model-meta'\nimport type { TTSOptions, TTSResult } from '@tanstack/ai'\nimport type { GoogleGenAI } from '@google/genai'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Provider-specific options for Gemini TTS\n *\n * @experimental Gemini TTS is an experimental feature.\n * @see https://ai.google.dev/gemini-api/docs/speech-generation\n */\nexport interface GeminiTTSProviderOptions {\n  /**\n   * Voice configuration for TTS.\n   * Choose from 30 available voices with different characteristics.\n   */\n  voiceConfig?: {\n    prebuiltVoiceConfig?: {\n      /**\n       * The voice name to use for speech synthesis.\n       * @see https://ai.google.dev/gemini-api/docs/speech-generation#voices\n       */\n      voiceName?: GeminiTTSVoice\n    }\n  }\n\n  /**\n   * System instruction for controlling speech style.\n   * Use natural language to describe the desired speaking style,\n   * pace, tone, accent, or other characteristics.\n   *\n   * @example \"Speak slowly and calmly, as if telling a bedtime story\"\n   * @example \"Use an upbeat, enthusiastic tone with moderate pace\"\n   * @example \"Speak with a British accent\"\n   */\n  systemInstruction?: string\n\n  /**\n   * Language code hint for the speech synthesis.\n   * Gemini TTS supports 24 languages and can auto-detect,\n   * but you can provide a hint for better results.\n   *\n   * @example \"en-US\" for American English\n   * @example \"es-ES\" for Spanish (Spain)\n   * @example \"ja-JP\" for Japanese\n   */\n  languageCode?: string\n}\n\n/**\n * Configuration for Gemini TTS adapter\n *\n * @experimental Gemini TTS is an experimental feature.\n */\nexport interface GeminiTTSConfig extends GeminiClientConfig {}\n\n/** Model type for Gemini TTS */\nexport type GeminiTTSModel = (typeof GEMINI_TTS_MODELS)[number]\n\n/**\n * Gemini Text-to-Speech Adapter\n *\n * Tree-shakeable adapter for Gemini TTS functionality.\n *\n * **IMPORTANT**: Gemini TTS uses the Live API (WebSocket-based) which requires\n * different handling than traditional REST APIs. This adapter provides a\n * simplified interface but may have limitations.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\n *\n * Models:\n * - gemini-2.5-flash-preview-tts\n */\nexport class GeminiTTSAdapter<\n  TModel extends GeminiTTSModel,\n> extends BaseTTSAdapter<TModel, GeminiTTSProviderOptions> {\n  readonly name = 'gemini' as const\n\n  private client: GoogleGenAI\n\n  constructor(config: GeminiTTSConfig, model: TModel) {\n    super(config, model)\n    this.client = createGeminiClient(config)\n  }\n\n  /**\n   * Generate speech from text using Gemini's TTS model.\n   *\n   * @experimental This implementation is experimental and may change.\n   * @see https://ai.google.dev/gemini-api/docs/speech-generation\n   */\n  async generateSpeech(\n    options: TTSOptions<GeminiTTSProviderOptions>,\n  ): Promise<TTSResult> {\n    const { model, text, modelOptions } = options\n\n    const voiceConfig = modelOptions?.voiceConfig || {\n      prebuiltVoiceConfig: {\n        voiceName: 'Kore',\n      },\n    }\n\n    const response = await this.client.models.generateContent({\n      model,\n      contents: [\n        {\n          role: 'user',\n          parts: [{ text }],\n        },\n      ],\n      config: {\n        responseModalities: ['AUDIO'],\n        speechConfig: {\n          voiceConfig,\n          ...(modelOptions?.languageCode && {\n            languageCode: modelOptions.languageCode,\n          }),\n        },\n      },\n      ...(modelOptions?.systemInstruction && {\n        systemInstruction: modelOptions.systemInstruction,\n      }),\n    })\n\n    // Extract audio data from response\n    const candidate = response.candidates?.[0]\n    const parts = candidate?.content?.parts\n\n    if (!parts || parts.length === 0) {\n      throw new Error('No audio output received from Gemini TTS')\n    }\n\n    // Look for inline data (audio)\n    const audioPart = parts.find((part: any) =>\n      part.inlineData?.mimeType?.startsWith('audio/'),\n    )\n\n    if (!audioPart || !audioPart.inlineData || !audioPart.inlineData.data) {\n      throw new Error('No audio data in Gemini TTS response')\n    }\n\n    const audioBase64 = audioPart.inlineData.data\n    const mimeType = audioPart.inlineData.mimeType || 'audio/wav'\n    const format = mimeType.split('/')[1] || 'wav'\n\n    return {\n      id: generateId(this.name),\n      model,\n      audio: audioBase64,\n      format,\n      contentType: mimeType,\n    }\n  }\n}\n\n/**\n * Creates a Gemini TTS adapter with explicit API key.\n * Type resolution happens here at the call site.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n * @returns Configured Gemini TTS adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGeminiSpeech('gemini-2.5-flash-preview-tts', \"your-api-key\");\n *\n * const result = await generateSpeech({\n *   adapter,\n *   text: 'Hello, world!'\n * });\n * ```\n */\nexport function createGeminiSpeech<TModel extends GeminiTTSModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiTTSConfig, 'apiKey'>,\n): GeminiTTSAdapter<TModel> {\n  return new GeminiTTSAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini speech adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\n *\n * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Gemini speech adapter instance with resolved types\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses GOOGLE_API_KEY from environment\n * const adapter = geminiSpeech('gemini-2.5-flash-preview-tts');\n *\n * const result = await generateSpeech({\n *   adapter,\n *   text: 'Welcome to TanStack AI!'\n * });\n * ```\n */\nexport function geminiSpeech<TModel extends GeminiTTSModel>(\n  model: TModel,\n  config?: Omit<GeminiTTSConfig, 'apiKey'>,\n): GeminiTTSAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiSpeech(model, apiKey, config)\n}\n"],"names":[],"mappings":";;AA+EO,MAAM,yBAEH,eAAiD;AAAA,EAKzD,YAAY,QAAyB,OAAe;AAClD,UAAM,QAAQ,KAAK;AALrB,SAAS,OAAO;AAMd,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eACJ,SACoB;AACpB,UAAM,EAAE,OAAO,MAAM,aAAA,IAAiB;AAEtC,UAAM,cAAc,cAAc,eAAe;AAAA,MAC/C,qBAAqB;AAAA,QACnB,WAAW;AAAA,MAAA;AAAA,IACb;AAGF,UAAM,WAAW,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,MACxD;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,OAAO,CAAC,EAAE,KAAA,CAAM;AAAA,QAAA;AAAA,MAClB;AAAA,MAEF,QAAQ;AAAA,QACN,oBAAoB,CAAC,OAAO;AAAA,QAC5B,cAAc;AAAA,UACZ;AAAA,UACA,GAAI,cAAc,gBAAgB;AAAA,YAChC,cAAc,aAAa;AAAA,UAAA;AAAA,QAC7B;AAAA,MACF;AAAA,MAEF,GAAI,cAAc,qBAAqB;AAAA,QACrC,mBAAmB,aAAa;AAAA,MAAA;AAAA,IAClC,CACD;AAGD,UAAM,YAAY,SAAS,aAAa,CAAC;AACzC,UAAM,QAAQ,WAAW,SAAS;AAElC,QAAI,CAAC,SAAS,MAAM,WAAW,GAAG;AAChC,YAAM,IAAI,MAAM,0CAA0C;AAAA,IAC5D;AAGA,UAAM,YAAY,MAAM;AAAA,MAAK,CAAC,SAC5B,KAAK,YAAY,UAAU,WAAW,QAAQ;AAAA,IAAA;AAGhD,QAAI,CAAC,aAAa,CAAC,UAAU,cAAc,CAAC,UAAU,WAAW,MAAM;AACrE,YAAM,IAAI,MAAM,sCAAsC;AAAA,IACxD;AAEA,UAAM,cAAc,UAAU,WAAW;AACzC,UAAM,WAAW,UAAU,WAAW,YAAY;AAClD,UAAM,SAAS,SAAS,MAAM,GAAG,EAAE,CAAC,KAAK;AAEzC,WAAO;AAAA,MACL,IAAI,WAAW,KAAK,IAAI;AAAA,MACxB;AAAA,MACA,OAAO;AAAA,MACP;AAAA,MACA,aAAa;AAAA,IAAA;AAAA,EAEjB;AACF;AAuBO,SAAS,mBACd,OACA,QACA,QAC0B;AAC1B,SAAO,IAAI,iBAAiB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC1D;AA4BO,SAAS,aACd,OACA,QAC0B;AAC1B,QAAM,SAAS,uBAAA;AACf,SAAO,mBAAmB,OAAO,QAAQ,MAAM;AACjD;"}