@livekit/agents-plugin-google 1.0.30 → 1.0.32

package/dist/llm.js

@@ -340,7 +340,9 @@ class LLMStream extends llm.LLMStream {
             llm.FunctionCall.create({
               callId: part.functionCall.id || shortuuid("function_call_"),
               name: part.functionCall.name,
-              args: JSON.stringify(part.functionCall.args)
+              args: JSON.stringify(part.functionCall.args),
+              // Preserve thought signature for Gemini 3+ thinking mode
+              thoughtSignature: part.thoughtSignature
             })
           ]
         }

package/dist/llm.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type * as types from '@google/genai';\nimport { FunctionCallingConfigMode, type GenerateContentConfig, GoogleGenAI } from '@google/genai';\nimport type { APIConnectOptions } from '@livekit/agents';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  llm,\n  shortuuid,\n} from '@livekit/agents';\nimport type { ChatModels } from './models.js';\nimport type { LLMTools } from './tools.js';\nimport { toFunctionDeclarations } from './utils.js';\n\ninterface GoogleFormatData {\n  systemMessages: string[] | null;\n}\n\nexport interface LLMOptions {\n  model: string | ChatModels;\n  apiKey?: string;\n  temperature?: number;\n  toolChoice?: llm.ToolChoice;\n  vertexai?: boolean;\n  project?: string;\n  location?: string;\n  maxOutputTokens?: number;\n  topP?: number;\n  topK?: number;\n  presencePenalty?: number;\n  frequencyPenalty?: number;\n  thinkingConfig?: types.ThinkingConfig;\n  automaticFunctionCallingConfig?: types.AutomaticFunctionCallingConfig;\n  geminiTools?: LLMTools;\n  httpOptions?: types.HttpOptions;\n  seed?: number;\n}\n\nexport class LLM extends llm.LLM {\n  #opts: LLMOptions;\n  #client: GoogleGenAI;\n\n  label(): string {\n    return 'google.LLM';\n  }\n\n  get model(): string {\n    return this.#opts.model;\n  }\n\n  /**\n   * Create a new instance of Google GenAI LLM.\n   *\n   * Environment Requirements:\n   * - For VertexAI: Set the `GOOGLE_APPLICATION_CREDENTIALS` environment variable to the path of the service account key file or use any of the other Google Cloud auth methods.\n   * The Google Cloud project and location can be set via `project` and `location` arguments or the environment variables\n   * `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION`. By default, the project is inferred from the service account key file,\n   * and the location defaults to \"us-central1\".\n   * - For Google Gemini API: Set the `apiKey` argument or the `GOOGLE_API_KEY` environment variable.\n   *\n   * @param model - The model name to use. Defaults to \"gemini-2.0-flash-001\".\n   * @param apiKey - The API key for Google Gemini. If not provided, it attempts to read from the `GOOGLE_API_KEY` environment variable.\n   * @param vertexai - Whether to use VertexAI. If not provided, it attempts to read from the `GOOGLE_GENAI_USE_VERTEXAI` environment variable. Defaults to false.\n   * @param project - The Google Cloud project to use (only for VertexAI). Defaults to undefined.\n   * @param location - The location to use for VertexAI API requests. Default value is \"us-central1\".\n   * @param temperature - Sampling temperature for response generation. Defaults to undefined.\n   * @param maxOutputTokens - Maximum number of tokens to generate in the output. Defaults to undefined.\n   * @param topP - The nucleus sampling probability for response generation. Defaults to undefined.\n   * @param topK - The top-k sampling value for response generation. Defaults to undefined.\n   * @param presencePenalty - Penalizes the model for generating previously mentioned concepts. Defaults to undefined.\n   * @param frequencyPenalty - Penalizes the model for repeating words. Defaults to undefined.\n   * @param toolChoice - Specifies whether to use tools during response generation. Defaults to \"auto\".\n   * @param thinkingConfig - The thinking configuration for response generation. Defaults to undefined.\n   * @param automaticFunctionCallingConfig - The automatic function calling configuration for response generation. Defaults to undefined.\n   * @param geminiTools - The Gemini-specific tools to use for the session.\n   * @param httpOptions - The HTTP options to use for the session.\n   * @param seed - Random seed for reproducible results. Defaults to undefined.\n   */\n  constructor(\n    {\n      model,\n      apiKey,\n      vertexai,\n      project,\n      location,\n      temperature,\n      maxOutputTokens,\n      topP,\n      topK,\n      presencePenalty,\n      frequencyPenalty,\n      toolChoice,\n      thinkingConfig,\n      automaticFunctionCallingConfig,\n      geminiTools,\n      httpOptions,\n      seed,\n    }: LLMOptions = {\n      model: 'gemini-2.0-flash-001',\n    },\n  ) {\n    super();\n\n    const useVertexAI =\n      vertexai ??\n      (process.env.GOOGLE_GENAI_USE_VERTEXAI === 'true' ||\n        process.env.GOOGLE_GENAI_USE_VERTEXAI === '1');\n\n    let gcpProject: string | undefined = project ?? process.env.GOOGLE_CLOUD_PROJECT;\n    let gcpLocation: string | undefined = location ?? process.env.GOOGLE_CLOUD_LOCATION;\n    let geminiApiKey: string | undefined = apiKey ?? process.env.GOOGLE_API_KEY;\n\n    if (useVertexAI) {\n      if (!gcpProject) {\n        // TODO(brian): use default_async to get the project ID\n        throw new Error(\n          'Project ID is required for Vertex AI. Set via project option or GOOGLE_CLOUD_PROJECT environment variable',\n        );\n      }\n      geminiApiKey = undefined;\n    } else {\n      gcpProject = undefined;\n      gcpLocation = undefined;\n      if (!geminiApiKey) {\n        throw new Error(\n          'API key is required for Google API either via apiKey or GOOGLE_API_KEY environment variable',\n        );\n      }\n    }\n\n    // Validate thinkingConfig\n    if (thinkingConfig?.thinkingBudget !== undefined) {\n      const budget = thinkingConfig.thinkingBudget;\n      if (budget < 0 || budget > 24576) {\n        throw new Error('thinkingBudget inside thinkingConfig must be between 0 and 24576');\n      }\n    }\n\n    const clientOptions: types.GoogleGenAIOptions = useVertexAI\n      ? {\n          vertexai: true,\n          project: gcpProject,\n          location: gcpLocation,\n        }\n      : {\n          apiKey: geminiApiKey,\n        };\n\n    this.#client = new GoogleGenAI(clientOptions);\n\n    this.#opts = {\n      model,\n      vertexai: useVertexAI,\n      project: gcpProject,\n      location: gcpLocation,\n      temperature,\n      maxOutputTokens,\n      topP,\n      topK,\n      presencePenalty,\n      frequencyPenalty,\n      toolChoice,\n      thinkingConfig,\n      automaticFunctionCallingConfig,\n      geminiTools,\n      httpOptions,\n      seed,\n      apiKey,\n    };\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    toolChoice,\n    extraKwargs,\n    geminiTools,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n    geminiTools?: LLMTools;\n  }): LLMStream {\n    const extras: GenerateContentConfig = { ...extraKwargs } as GenerateContentConfig;\n\n    toolChoice = toolChoice !== undefined ? toolChoice : this.#opts.toolChoice;\n\n    if (toolChoice) {\n      let geminiToolConfig: types.ToolConfig;\n\n      if (typeof toolChoice === 'object' && toolChoice.type === 'function') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.ANY,\n            allowedFunctionNames: [toolChoice.function.name],\n          },\n        };\n      } else if (toolChoice === 'required') {\n        const toolNames = Object.entries(toolCtx || {}).map(([name]) => name);\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.ANY,\n            allowedFunctionNames: toolNames.length > 0 ? toolNames : undefined,\n          },\n        };\n      } else if (toolChoice === 'auto') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.AUTO,\n          },\n        };\n      } else if (toolChoice === 'none') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.NONE,\n          },\n        };\n      } else {\n        throw new Error(`Invalid tool choice: ${toolChoice}`);\n      }\n\n      extras.toolConfig = geminiToolConfig;\n    }\n\n    if (this.#opts.temperature !== undefined) {\n      extras.temperature = this.#opts.temperature;\n    }\n    if (this.#opts.maxOutputTokens !== undefined) {\n      extras.maxOutputTokens = this.#opts.maxOutputTokens;\n    }\n    if (this.#opts.topP !== undefined) {\n      extras.topP = this.#opts.topP;\n    }\n    if (this.#opts.topK !== undefined) {\n      extras.topK = this.#opts.topK;\n    }\n    if (this.#opts.presencePenalty !== undefined) {\n      extras.presencePenalty = this.#opts.presencePenalty;\n    }\n    if (this.#opts.frequencyPenalty !== undefined) {\n      extras.frequencyPenalty = this.#opts.frequencyPenalty;\n    }\n    if (this.#opts.seed !== undefined) {\n      extras.seed = this.#opts.seed;\n    }\n\n    if (this.#opts.thinkingConfig !== undefined) {\n      extras.thinkingConfig = this.#opts.thinkingConfig;\n    }\n\n    if (this.#opts.automaticFunctionCallingConfig !== undefined) {\n      extras.automaticFunctionCalling = this.#opts.automaticFunctionCallingConfig;\n    }\n\n    geminiTools = geminiTools !== undefined ? geminiTools : this.#opts.geminiTools;\n\n    return new LLMStream(this, {\n      client: this.#client,\n      model: this.#opts.model,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      geminiTools,\n      extraKwargs: extras,\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  #client: GoogleGenAI;\n  #model: string;\n  #geminiTools?: LLMTools;\n  #extraKwargs: GenerateContentConfig;\n\n  constructor(\n    llm: LLM,\n    {\n      client,\n      model,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      geminiTools,\n      extraKwargs,\n    }: {\n      client: GoogleGenAI;\n      model: string;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      connOptions: APIConnectOptions;\n      geminiTools?: LLMTools;\n      extraKwargs: GenerateContentConfig;\n    },\n  ) {\n    // Call base constructor with dev 1.0 object parameter pattern\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.#client = client;\n    this.#model = model;\n    this.#geminiTools = geminiTools;\n    this.#extraKwargs = extraKwargs;\n  }\n\n  protected async run(): Promise<void> {\n    let retryable = true;\n    const requestId = `google_${Date.now()}`;\n\n    try {\n      const [turns, extraData] = (await this.chatCtx.toProviderFormat('google')) as [\n        Record<string, unknown>[],\n        GoogleFormatData,\n      ];\n\n      const contents: types.Content[] = turns.map((turn: Record<string, unknown>) => ({\n        role: turn.role as types.Content['role'],\n        parts: turn.parts as types.Part[],\n      }));\n\n      const functionDeclarations = this.toolCtx ? toFunctionDeclarations(this.toolCtx) : undefined;\n      const tools =\n        functionDeclarations && functionDeclarations.length > 0\n          ? [{ functionDeclarations }]\n          : undefined;\n\n      let systemInstruction: types.Content | undefined = undefined;\n      if (extraData.systemMessages && extraData.systemMessages.length > 0) {\n        systemInstruction = {\n          parts: extraData.systemMessages.map((content: string) => ({ text: content })),\n        };\n      }\n\n      const response = await this.#client.models.generateContentStream({\n        model: this.#model,\n        contents,\n        config: {\n          ...this.#extraKwargs,\n          systemInstruction,\n          httpOptions: this.#extraKwargs.httpOptions ?? {\n            timeout: Math.floor(this.connOptions.timeoutMs),\n          },\n          tools,\n        },\n      });\n\n      for await (const chunk of response) {\n        if (chunk.promptFeedback) {\n          throw new APIStatusError({\n            message: `Prompt feedback error: ${JSON.stringify(chunk.promptFeedback)}`,\n            options: {\n              retryable: false,\n              requestId,\n            },\n          });\n        }\n\n        if (!chunk.candidates || !chunk.candidates[0]?.content?.parts) {\n          this.logger.warn(`No candidates in the response: ${JSON.stringify(chunk)}`);\n          continue;\n        }\n\n        if (chunk.candidates.length > 1) {\n          this.logger.warn(\n            'Google LLM: there are multiple candidates in the response, returning response from the first one.',\n          );\n        }\n\n        for (const part of chunk.candidates[0].content.parts) {\n          const chatChunk = this.#parsePart(requestId, part);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usageMetadata) {\n          const usage = chunk.usageMetadata;\n          this.queue.put({\n            id: requestId,\n            usage: {\n              completionTokens: usage.candidatesTokenCount || 0,\n              promptTokens: usage.promptTokenCount || 0,\n              promptCachedTokens: usage.cachedContentTokenCount || 0,\n              totalTokens: usage.totalTokenCount || 0,\n            },\n          });\n        }\n      }\n    } catch (error: unknown) {\n      const err = error as {\n        code?: number;\n        message?: string;\n        status?: string;\n        type?: string;\n      };\n\n      if (err.code && err.code >= 400 && err.code < 500) {\n        if (err.code === 429) {\n          throw new APIStatusError({\n            message: `Google LLM: Rate limit error - ${err.message || 'Unknown error'}`,\n            options: {\n              statusCode: 429,\n              retryable: true,\n            },\n          });\n        } else {\n          throw new APIStatusError({\n            message: `Google LLM: Client error (${err.code}) - ${err.message || 'Unknown error'}`,\n            options: {\n              statusCode: err.code,\n              retryable: false,\n            },\n          });\n        }\n      }\n\n      if (err.code && err.code >= 500) {\n        throw new APIStatusError({\n          message: `Google LLM: Server error (${err.code}) - ${err.message || 'Unknown error'}`,\n          options: {\n            statusCode: err.code,\n            retryable,\n          },\n        });\n      }\n\n      throw new APIConnectionError({\n        message: `Google LLM: API error - ${err.message || 'Unknown error'}`,\n        options: {\n          retryable,\n        },\n      });\n    }\n  }\n\n  #parsePart(id: string, part: types.Part): llm.ChatChunk | null {\n    if (part.functionCall) {\n      return {\n        id,\n        delta: {\n          role: 'assistant',\n          toolCalls: [\n            llm.FunctionCall.create({\n              callId: part.functionCall.id || shortuuid('function_call_'),\n              name: part.functionCall.name!,\n              args: JSON.stringify(part.functionCall.args!),\n            }),\n          ],\n        },\n      };\n    }\n\n    return {\n      id,\n      delta: {\n        content: part.text,\n        role: 'assistant',\n      },\n    };\n  }\n}\n"],"mappings":"AAIA,SAAS,2BAAuD,mBAAmB;AAEnF;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAGP,SAAS,8BAA8B;AA0BhC,MAAM,YAAY,IAAI,IAAI;AAAA,EAC/B;AAAA,EACA;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,MAAM;AAAA,EACpB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA8BA,YACE;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,IAAgB;AAAA,IACd,OAAO;AAAA,EACT,GACA;AACA,UAAM;AAEN,UAAM,cACJ,aACC,QAAQ,IAAI,8BAA8B,UACzC,QAAQ,IAAI,8BAA8B;AAE9C,QAAI,aAAiC,WAAW,QAAQ,IAAI;AAC5D,QAAI,cAAkC,YAAY,QAAQ,IAAI;AAC9D,QAAI,eAAmC,UAAU,QAAQ,IAAI;AAE7D,QAAI,aAAa;AACf,UAAI,CAAC,YAAY;AAEf,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AACA,qBAAe;AAAA,IACjB,OAAO;AACL,mBAAa;AACb,oBAAc;AACd,UAAI,CAAC,cAAc;AACjB,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAGA,SAAI,iDAAgB,oBAAmB,QAAW;AAChD,YAAM,SAAS,eAAe;AAC9B,UAAI,SAAS,KAAK,SAAS,OAAO;AAChC,cAAM,IAAI,MAAM,kEAAkE;AAAA,MACpF;AAAA,IACF;AAEA,UAAM,gBAA0C,cAC5C;AAAA,MACE,UAAU;AAAA,MACV,SAAS;AAAA,MACT,UAAU;AAAA,IACZ,IACA;AAAA,MACE,QAAQ;AAAA,IACV;AAEJ,SAAK,UAAU,IAAI,YAAY,aAAa;AAE5C,SAAK,QAAQ;AAAA,MACX;AAAA,MACA,UAAU;AAAA,MACV,SAAS;AAAA,MACT,UAAU;AAAA,MACV;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAQc;AACZ,UAAM,SAAgC,EAAE,GAAG,YAAY;AAEvD,iBAAa,eAAe,SAAY,aAAa,KAAK,MAAM;AAEhE,QAAI,YAAY;AACd,UAAI;AAEJ,UAAI,OAAO,eAAe,YAAY,WAAW,SAAS,YAAY;AACpE,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,YAChC,sBAAsB,CAAC,WAAW,SAAS,IAAI;AAAA,UACjD;AAAA,QACF;AAAA,MACF,WAAW,eAAe,YAAY;AACpC,cAAM,YAAY,OAAO,QAAQ,WAAW,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,IAAI,MAAM,IAAI;AACpE,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,YAChC,sBAAsB,UAAU,SAAS,IAAI,YAAY;AAAA,UAC3D;AAAA,QACF;AAAA,MACF,WAAW,eAAe,QAAQ;AAChC,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,UAClC;AAAA,QACF;AAAA,MACF,WAAW,eAAe,QAAQ;AAChC,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,UAClC;AAAA,QACF;AAAA,MACF,OAAO;AACL,cAAM,IAAI,MAAM,wBAAwB,UAAU,EAAE;AAAA,MACtD;AAEA,aAAO,aAAa;AAAA,IACtB;AAEA,QAAI,KAAK,MAAM,gBAAgB,QAAW;AACxC,aAAO,cAAc,KAAK,MAAM;AAAA,IAClC;AACA,QAAI,KAAK,MAAM,oBAAoB,QAAW;AAC5C,aAAO,kBAAkB,KAAK,MAAM;AAAA,IACtC;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AACA,QAAI,KAAK,MAAM,oBAAoB,QAAW;AAC5C,aAAO,kBAAkB,KAAK,MAAM;AAAA,IACtC;AACA,QAAI,KAAK,MAAM,qBAAqB,QAAW;AAC7C,aAAO,mBAAmB,KAAK,MAAM;AAAA,IACvC;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AAEA,QAAI,KAAK,MAAM,mBAAmB,QAAW;AAC3C,aAAO,iBAAiB,KAAK,MAAM;AAAA,IACrC;AAEA,QAAI,KAAK,MAAM,mCAAmC,QAAW;AAC3D,aAAO,2BAA2B,KAAK,MAAM;AAAA,IAC/C;AAEA,kBAAc,gBAAgB,SAAY,cAAc,KAAK,MAAM;AAEnE,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,QAAQ,KAAK;AAAA,MACb,OAAO,KAAK,MAAM;AAAA,MAClB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,IACf,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EAC3C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GASA;AAEA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,UAAU;AACf,SAAK,SAAS;AACd,SAAK,eAAe;AACpB,SAAK,eAAe;AAAA,EACtB;AAAA,EAEA,MAAgB,MAAqB;AArTvC;AAsTI,QAAI,YAAY;AAChB,UAAM,YAAY,UAAU,KAAK,IAAI,CAAC;AAEtC,QAAI;AACF,YAAM,CAAC,OAAO,SAAS,IAAK,MAAM,KAAK,QAAQ,iBAAiB,QAAQ;AAKxE,YAAM,WAA4B,MAAM,IAAI,CAAC,UAAmC;AAAA,QAC9E,MAAM,KAAK;AAAA,QACX,OAAO,KAAK;AAAA,MACd,EAAE;AAEF,YAAM,uBAAuB,KAAK,UAAU,uBAAuB,KAAK,OAAO,IAAI;AACnF,YAAM,QACJ,wBAAwB,qBAAqB,SAAS,IAClD,CAAC,EAAE,qBAAqB,CAAC,IACzB;AAEN,UAAI,oBAA+C;AACnD,UAAI,UAAU,kBAAkB,UAAU,eAAe,SAAS,GAAG;AACnE,4BAAoB;AAAA,UAClB,OAAO,UAAU,eAAe,IAAI,CAAC,aAAqB,EAAE,MAAM,QAAQ,EAAE;AAAA,QAC9E;AAAA,MACF;AAEA,YAAM,WAAW,MAAM,KAAK,QAAQ,OAAO,sBAAsB;AAAA,QAC/D,OAAO,KAAK;AAAA,QACZ;AAAA,QACA,QAAQ;AAAA,UACN,GAAG,KAAK;AAAA,UACR;AAAA,UACA,aAAa,KAAK,aAAa,eAAe;AAAA,YAC5C,SAAS,KAAK,MAAM,KAAK,YAAY,SAAS;AAAA,UAChD;AAAA,UACA;AAAA,QACF;AAAA,MACF,CAAC;AAED,uBAAiB,SAAS,UAAU;AAClC,YAAI,MAAM,gBAAgB;AACxB,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,0BAA0B,KAAK,UAAU,MAAM,cAAc,CAAC;AAAA,YACvE,SAAS;AAAA,cACP,WAAW;AAAA,cACX;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAEA,YAAI,CAAC,MAAM,cAAc,GAAC,iBAAM,WAAW,CAAC,MAAlB,mBAAqB,YAArB,mBAA8B,QAAO;AAC7D,eAAK,OAAO,KAAK,kCAAkC,KAAK,UAAU,KAAK,CAAC,EAAE;AAC1E;AAAA,QACF;AAEA,YAAI,MAAM,WAAW,SAAS,GAAG;AAC/B,eAAK,OAAO;AAAA,YACV;AAAA,UACF;AAAA,QACF;AAEA,mBAAW,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ,OAAO;AACpD,gBAAM,YAAY,KAAK,WAAW,WAAW,IAAI;AACjD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,eAAe;AACvB,gBAAM,QAAQ,MAAM;AACpB,eAAK,MAAM,IAAI;AAAA,YACb,IAAI;AAAA,YACJ,OAAO;AAAA,cACL,kBAAkB,MAAM,wBAAwB;AAAA,cAChD,cAAc,MAAM,oBAAoB;AAAA,cACxC,oBAAoB,MAAM,2BAA2B;AAAA,cACrD,aAAa,MAAM,mBAAmB;AAAA,YACxC;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,YAAM,MAAM;AAOZ,UAAI,IAAI,QAAQ,IAAI,QAAQ,OAAO,IAAI,OAAO,KAAK;AACjD,YAAI,IAAI,SAAS,KAAK;AACpB,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,kCAAkC,IAAI,WAAW,eAAe;AAAA,YACzE,SAAS;AAAA,cACP,YAAY;AAAA,cACZ,WAAW;AAAA,YACb;AAAA,UACF,CAAC;AAAA,QACH,OAAO;AACL,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,6BAA6B,IAAI,IAAI,OAAO,IAAI,WAAW,eAAe;AAAA,YACnF,SAAS;AAAA,cACP,YAAY,IAAI;AAAA,cAChB,WAAW;AAAA,YACb;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAEA,UAAI,IAAI,QAAQ,IAAI,QAAQ,KAAK;AAC/B,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,6BAA6B,IAAI,IAAI,OAAO,IAAI,WAAW,eAAe;AAAA,UACnF,SAAS;AAAA,YACP,YAAY,IAAI;AAAA,YAChB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH;AAEA,YAAM,IAAI,mBAAmB;AAAA,QAC3B,SAAS,2BAA2B,IAAI,WAAW,eAAe;AAAA,QAClE,SAAS;AAAA,UACP;AAAA,QACF;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,WAAW,IAAY,MAAwC;AAC7D,QAAI,KAAK,cAAc;AACrB,aAAO;AAAA,QACL;AAAA,QACA,OAAO;AAAA,UACL,MAAM;AAAA,UACN,WAAW;AAAA,YACT,IAAI,aAAa,OAAO;AAAA,cACtB,QAAQ,KAAK,aAAa,MAAM,UAAU,gBAAgB;AAAA,cAC1D,MAAM,KAAK,aAAa;AAAA,cACxB,MAAM,KAAK,UAAU,KAAK,aAAa,IAAK;AAAA,YAC9C,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,SAAS,KAAK;AAAA,QACd,MAAM;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACF;","names":["llm"]}
+{"version":3,"sources":["../src/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type * as types from '@google/genai';\nimport { FunctionCallingConfigMode, type GenerateContentConfig, GoogleGenAI } from '@google/genai';\nimport type { APIConnectOptions } from '@livekit/agents';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  llm,\n  shortuuid,\n} from '@livekit/agents';\nimport type { ChatModels } from './models.js';\nimport type { LLMTools } from './tools.js';\nimport { toFunctionDeclarations } from './utils.js';\n\ninterface GoogleFormatData {\n  systemMessages: string[] | null;\n}\n\nexport interface LLMOptions {\n  model: string | ChatModels;\n  apiKey?: string;\n  temperature?: number;\n  toolChoice?: llm.ToolChoice;\n  vertexai?: boolean;\n  project?: string;\n  location?: string;\n  maxOutputTokens?: number;\n  topP?: number;\n  topK?: number;\n  presencePenalty?: number;\n  frequencyPenalty?: number;\n  thinkingConfig?: types.ThinkingConfig;\n  automaticFunctionCallingConfig?: types.AutomaticFunctionCallingConfig;\n  geminiTools?: LLMTools;\n  httpOptions?: types.HttpOptions;\n  seed?: number;\n}\n\nexport class LLM extends llm.LLM {\n  #opts: LLMOptions;\n  #client: GoogleGenAI;\n\n  label(): string {\n    return 'google.LLM';\n  }\n\n  get model(): string {\n    return this.#opts.model;\n  }\n\n  /**\n   * Create a new instance of Google GenAI LLM.\n   *\n   * Environment Requirements:\n   * - For VertexAI: Set the `GOOGLE_APPLICATION_CREDENTIALS` environment variable to the path of the service account key file or use any of the other Google Cloud auth methods.\n   * The Google Cloud project and location can be set via `project` and `location` arguments or the environment variables\n   * `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION`. By default, the project is inferred from the service account key file,\n   * and the location defaults to \"us-central1\".\n   * - For Google Gemini API: Set the `apiKey` argument or the `GOOGLE_API_KEY` environment variable.\n   *\n   * @param model - The model name to use. Defaults to \"gemini-2.0-flash-001\".\n   * @param apiKey - The API key for Google Gemini. If not provided, it attempts to read from the `GOOGLE_API_KEY` environment variable.\n   * @param vertexai - Whether to use VertexAI. If not provided, it attempts to read from the `GOOGLE_GENAI_USE_VERTEXAI` environment variable. Defaults to false.\n   * @param project - The Google Cloud project to use (only for VertexAI). Defaults to undefined.\n   * @param location - The location to use for VertexAI API requests. Default value is \"us-central1\".\n   * @param temperature - Sampling temperature for response generation. Defaults to undefined.\n   * @param maxOutputTokens - Maximum number of tokens to generate in the output. Defaults to undefined.\n   * @param topP - The nucleus sampling probability for response generation. Defaults to undefined.\n   * @param topK - The top-k sampling value for response generation. Defaults to undefined.\n   * @param presencePenalty - Penalizes the model for generating previously mentioned concepts. Defaults to undefined.\n   * @param frequencyPenalty - Penalizes the model for repeating words. Defaults to undefined.\n   * @param toolChoice - Specifies whether to use tools during response generation. Defaults to \"auto\".\n   * @param thinkingConfig - The thinking configuration for response generation. Defaults to undefined.\n   * @param automaticFunctionCallingConfig - The automatic function calling configuration for response generation. Defaults to undefined.\n   * @param geminiTools - The Gemini-specific tools to use for the session.\n   * @param httpOptions - The HTTP options to use for the session.\n   * @param seed - Random seed for reproducible results. Defaults to undefined.\n   */\n  constructor(\n    {\n      model,\n      apiKey,\n      vertexai,\n      project,\n      location,\n      temperature,\n      maxOutputTokens,\n      topP,\n      topK,\n      presencePenalty,\n      frequencyPenalty,\n      toolChoice,\n      thinkingConfig,\n      automaticFunctionCallingConfig,\n      geminiTools,\n      httpOptions,\n      seed,\n    }: LLMOptions = {\n      model: 'gemini-2.0-flash-001',\n    },\n  ) {\n    super();\n\n    const useVertexAI =\n      vertexai ??\n      (process.env.GOOGLE_GENAI_USE_VERTEXAI === 'true' ||\n        process.env.GOOGLE_GENAI_USE_VERTEXAI === '1');\n\n    let gcpProject: string | undefined = project ?? process.env.GOOGLE_CLOUD_PROJECT;\n    let gcpLocation: string | undefined = location ?? process.env.GOOGLE_CLOUD_LOCATION;\n    let geminiApiKey: string | undefined = apiKey ?? process.env.GOOGLE_API_KEY;\n\n    if (useVertexAI) {\n      if (!gcpProject) {\n        // TODO(brian): use default_async to get the project ID\n        throw new Error(\n          'Project ID is required for Vertex AI. Set via project option or GOOGLE_CLOUD_PROJECT environment variable',\n        );\n      }\n      geminiApiKey = undefined;\n    } else {\n      gcpProject = undefined;\n      gcpLocation = undefined;\n      if (!geminiApiKey) {\n        throw new Error(\n          'API key is required for Google API either via apiKey or GOOGLE_API_KEY environment variable',\n        );\n      }\n    }\n\n    // Validate thinkingConfig\n    if (thinkingConfig?.thinkingBudget !== undefined) {\n      const budget = thinkingConfig.thinkingBudget;\n      if (budget < 0 || budget > 24576) {\n        throw new Error('thinkingBudget inside thinkingConfig must be between 0 and 24576');\n      }\n    }\n\n    const clientOptions: types.GoogleGenAIOptions = useVertexAI\n      ? {\n          vertexai: true,\n          project: gcpProject,\n          location: gcpLocation,\n        }\n      : {\n          apiKey: geminiApiKey,\n        };\n\n    this.#client = new GoogleGenAI(clientOptions);\n\n    this.#opts = {\n      model,\n      vertexai: useVertexAI,\n      project: gcpProject,\n      location: gcpLocation,\n      temperature,\n      maxOutputTokens,\n      topP,\n      topK,\n      presencePenalty,\n      frequencyPenalty,\n      toolChoice,\n      thinkingConfig,\n      automaticFunctionCallingConfig,\n      geminiTools,\n      httpOptions,\n      seed,\n      apiKey,\n    };\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    toolChoice,\n    extraKwargs,\n    geminiTools,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n    geminiTools?: LLMTools;\n  }): LLMStream {\n    const extras: GenerateContentConfig = { ...extraKwargs } as GenerateContentConfig;\n\n    toolChoice = toolChoice !== undefined ? toolChoice : this.#opts.toolChoice;\n\n    if (toolChoice) {\n      let geminiToolConfig: types.ToolConfig;\n\n      if (typeof toolChoice === 'object' && toolChoice.type === 'function') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.ANY,\n            allowedFunctionNames: [toolChoice.function.name],\n          },\n        };\n      } else if (toolChoice === 'required') {\n        const toolNames = Object.entries(toolCtx || {}).map(([name]) => name);\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.ANY,\n            allowedFunctionNames: toolNames.length > 0 ? toolNames : undefined,\n          },\n        };\n      } else if (toolChoice === 'auto') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.AUTO,\n          },\n        };\n      } else if (toolChoice === 'none') {\n        geminiToolConfig = {\n          functionCallingConfig: {\n            mode: FunctionCallingConfigMode.NONE,\n          },\n        };\n      } else {\n        throw new Error(`Invalid tool choice: ${toolChoice}`);\n      }\n\n      extras.toolConfig = geminiToolConfig;\n    }\n\n    if (this.#opts.temperature !== undefined) {\n      extras.temperature = this.#opts.temperature;\n    }\n    if (this.#opts.maxOutputTokens !== undefined) {\n      extras.maxOutputTokens = this.#opts.maxOutputTokens;\n    }\n    if (this.#opts.topP !== undefined) {\n      extras.topP = this.#opts.topP;\n    }\n    if (this.#opts.topK !== undefined) {\n      extras.topK = this.#opts.topK;\n    }\n    if (this.#opts.presencePenalty !== undefined) {\n      extras.presencePenalty = this.#opts.presencePenalty;\n    }\n    if (this.#opts.frequencyPenalty !== undefined) {\n      extras.frequencyPenalty = this.#opts.frequencyPenalty;\n    }\n    if (this.#opts.seed !== undefined) {\n      extras.seed = this.#opts.seed;\n    }\n\n    if (this.#opts.thinkingConfig !== undefined) {\n      extras.thinkingConfig = this.#opts.thinkingConfig;\n    }\n\n    if (this.#opts.automaticFunctionCallingConfig !== undefined) {\n      extras.automaticFunctionCalling = this.#opts.automaticFunctionCallingConfig;\n    }\n\n    geminiTools = geminiTools !== undefined ? geminiTools : this.#opts.geminiTools;\n\n    return new LLMStream(this, {\n      client: this.#client,\n      model: this.#opts.model,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      geminiTools,\n      extraKwargs: extras,\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  #client: GoogleGenAI;\n  #model: string;\n  #geminiTools?: LLMTools;\n  #extraKwargs: GenerateContentConfig;\n\n  constructor(\n    llm: LLM,\n    {\n      client,\n      model,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      geminiTools,\n      extraKwargs,\n    }: {\n      client: GoogleGenAI;\n      model: string;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      connOptions: APIConnectOptions;\n      geminiTools?: LLMTools;\n      extraKwargs: GenerateContentConfig;\n    },\n  ) {\n    // Call base constructor with dev 1.0 object parameter pattern\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.#client = client;\n    this.#model = model;\n    this.#geminiTools = geminiTools;\n    this.#extraKwargs = extraKwargs;\n  }\n\n  protected async run(): Promise<void> {\n    let retryable = true;\n    const requestId = `google_${Date.now()}`;\n\n    try {\n      const [turns, extraData] = (await this.chatCtx.toProviderFormat('google')) as [\n        Record<string, unknown>[],\n        GoogleFormatData,\n      ];\n\n      const contents: types.Content[] = turns.map((turn: Record<string, unknown>) => ({\n        role: turn.role as types.Content['role'],\n        parts: turn.parts as types.Part[],\n      }));\n\n      const functionDeclarations = this.toolCtx ? toFunctionDeclarations(this.toolCtx) : undefined;\n      const tools =\n        functionDeclarations && functionDeclarations.length > 0\n          ? [{ functionDeclarations }]\n          : undefined;\n\n      let systemInstruction: types.Content | undefined = undefined;\n      if (extraData.systemMessages && extraData.systemMessages.length > 0) {\n        systemInstruction = {\n          parts: extraData.systemMessages.map((content: string) => ({ text: content })),\n        };\n      }\n\n      const response = await this.#client.models.generateContentStream({\n        model: this.#model,\n        contents,\n        config: {\n          ...this.#extraKwargs,\n          systemInstruction,\n          httpOptions: this.#extraKwargs.httpOptions ?? {\n            timeout: Math.floor(this.connOptions.timeoutMs),\n          },\n          tools,\n        },\n      });\n\n      for await (const chunk of response) {\n        if (chunk.promptFeedback) {\n          throw new APIStatusError({\n            message: `Prompt feedback error: ${JSON.stringify(chunk.promptFeedback)}`,\n            options: {\n              retryable: false,\n              requestId,\n            },\n          });\n        }\n\n        if (!chunk.candidates || !chunk.candidates[0]?.content?.parts) {\n          this.logger.warn(`No candidates in the response: ${JSON.stringify(chunk)}`);\n          continue;\n        }\n\n        if (chunk.candidates.length > 1) {\n          this.logger.warn(\n            'Google LLM: there are multiple candidates in the response, returning response from the first one.',\n          );\n        }\n\n        for (const part of chunk.candidates[0].content.parts) {\n          const chatChunk = this.#parsePart(requestId, part);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usageMetadata) {\n          const usage = chunk.usageMetadata;\n          this.queue.put({\n            id: requestId,\n            usage: {\n              completionTokens: usage.candidatesTokenCount || 0,\n              promptTokens: usage.promptTokenCount || 0,\n              promptCachedTokens: usage.cachedContentTokenCount || 0,\n              totalTokens: usage.totalTokenCount || 0,\n            },\n          });\n        }\n      }\n    } catch (error: unknown) {\n      const err = error as {\n        code?: number;\n        message?: string;\n        status?: string;\n        type?: string;\n      };\n\n      if (err.code && err.code >= 400 && err.code < 500) {\n        if (err.code === 429) {\n          throw new APIStatusError({\n            message: `Google LLM: Rate limit error - ${err.message || 'Unknown error'}`,\n            options: {\n              statusCode: 429,\n              retryable: true,\n            },\n          });\n        } else {\n          throw new APIStatusError({\n            message: `Google LLM: Client error (${err.code}) - ${err.message || 'Unknown error'}`,\n            options: {\n              statusCode: err.code,\n              retryable: false,\n            },\n          });\n        }\n      }\n\n      if (err.code && err.code >= 500) {\n        throw new APIStatusError({\n          message: `Google LLM: Server error (${err.code}) - ${err.message || 'Unknown error'}`,\n          options: {\n            statusCode: err.code,\n            retryable,\n          },\n        });\n      }\n\n      throw new APIConnectionError({\n        message: `Google LLM: API error - ${err.message || 'Unknown error'}`,\n        options: {\n          retryable,\n        },\n      });\n    }\n  }\n\n  #parsePart(id: string, part: types.Part): llm.ChatChunk | null {\n    if (part.functionCall) {\n      return {\n        id,\n        delta: {\n          role: 'assistant',\n          toolCalls: [\n            llm.FunctionCall.create({\n              callId: part.functionCall.id || shortuuid('function_call_'),\n              name: part.functionCall.name!,\n              args: JSON.stringify(part.functionCall.args!),\n              // Preserve thought signature for Gemini 3+ thinking mode\n              thoughtSignature: part.thoughtSignature,\n            }),\n          ],\n        },\n      };\n    }\n\n    return {\n      id,\n      delta: {\n        content: part.text,\n        role: 'assistant',\n      },\n    };\n  }\n}\n"],"mappings":"AAIA,SAAS,2BAAuD,mBAAmB;AAEnF;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAGP,SAAS,8BAA8B;AA0BhC,MAAM,YAAY,IAAI,IAAI;AAAA,EAC/B;AAAA,EACA;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,MAAM;AAAA,EACpB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA8BA,YACE;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,IAAgB;AAAA,IACd,OAAO;AAAA,EACT,GACA;AACA,UAAM;AAEN,UAAM,cACJ,aACC,QAAQ,IAAI,8BAA8B,UACzC,QAAQ,IAAI,8BAA8B;AAE9C,QAAI,aAAiC,WAAW,QAAQ,IAAI;AAC5D,QAAI,cAAkC,YAAY,QAAQ,IAAI;AAC9D,QAAI,eAAmC,UAAU,QAAQ,IAAI;AAE7D,QAAI,aAAa;AACf,UAAI,CAAC,YAAY;AAEf,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AACA,qBAAe;AAAA,IACjB,OAAO;AACL,mBAAa;AACb,oBAAc;AACd,UAAI,CAAC,cAAc;AACjB,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAGA,SAAI,iDAAgB,oBAAmB,QAAW;AAChD,YAAM,SAAS,eAAe;AAC9B,UAAI,SAAS,KAAK,SAAS,OAAO;AAChC,cAAM,IAAI,MAAM,kEAAkE;AAAA,MACpF;AAAA,IACF;AAEA,UAAM,gBAA0C,cAC5C;AAAA,MACE,UAAU;AAAA,MACV,SAAS;AAAA,MACT,UAAU;AAAA,IACZ,IACA;AAAA,MACE,QAAQ;AAAA,IACV;AAEJ,SAAK,UAAU,IAAI,YAAY,aAAa;AAE5C,SAAK,QAAQ;AAAA,MACX;AAAA,MACA,UAAU;AAAA,MACV,SAAS;AAAA,MACT,UAAU;AAAA,MACV;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAQc;AACZ,UAAM,SAAgC,EAAE,GAAG,YAAY;AAEvD,iBAAa,eAAe,SAAY,aAAa,KAAK,MAAM;AAEhE,QAAI,YAAY;AACd,UAAI;AAEJ,UAAI,OAAO,eAAe,YAAY,WAAW,SAAS,YAAY;AACpE,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,YAChC,sBAAsB,CAAC,WAAW,SAAS,IAAI;AAAA,UACjD;AAAA,QACF;AAAA,MACF,WAAW,eAAe,YAAY;AACpC,cAAM,YAAY,OAAO,QAAQ,WAAW,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,IAAI,MAAM,IAAI;AACpE,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,YAChC,sBAAsB,UAAU,SAAS,IAAI,YAAY;AAAA,UAC3D;AAAA,QACF;AAAA,MACF,WAAW,eAAe,QAAQ;AAChC,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,UAClC;AAAA,QACF;AAAA,MACF,WAAW,eAAe,QAAQ;AAChC,2BAAmB;AAAA,UACjB,uBAAuB;AAAA,YACrB,MAAM,0BAA0B;AAAA,UAClC;AAAA,QACF;AAAA,MACF,OAAO;AACL,cAAM,IAAI,MAAM,wBAAwB,UAAU,EAAE;AAAA,MACtD;AAEA,aAAO,aAAa;AAAA,IACtB;AAEA,QAAI,KAAK,MAAM,gBAAgB,QAAW;AACxC,aAAO,cAAc,KAAK,MAAM;AAAA,IAClC;AACA,QAAI,KAAK,MAAM,oBAAoB,QAAW;AAC5C,aAAO,kBAAkB,KAAK,MAAM;AAAA,IACtC;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AACA,QAAI,KAAK,MAAM,oBAAoB,QAAW;AAC5C,aAAO,kBAAkB,KAAK,MAAM;AAAA,IACtC;AACA,QAAI,KAAK,MAAM,qBAAqB,QAAW;AAC7C,aAAO,mBAAmB,KAAK,MAAM;AAAA,IACvC;AACA,QAAI,KAAK,MAAM,SAAS,QAAW;AACjC,aAAO,OAAO,KAAK,MAAM;AAAA,IAC3B;AAEA,QAAI,KAAK,MAAM,mBAAmB,QAAW;AAC3C,aAAO,iBAAiB,KAAK,MAAM;AAAA,IACrC;AAEA,QAAI,KAAK,MAAM,mCAAmC,QAAW;AAC3D,aAAO,2BAA2B,KAAK,MAAM;AAAA,IAC/C;AAEA,kBAAc,gBAAgB,SAAY,cAAc,KAAK,MAAM;AAEnE,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,QAAQ,KAAK;AAAA,MACb,OAAO,KAAK,MAAM;AAAA,MAClB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,IACf,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EAC3C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GASA;AAEA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,UAAU;AACf,SAAK,SAAS;AACd,SAAK,eAAe;AACpB,SAAK,eAAe;AAAA,EACtB;AAAA,EAEA,MAAgB,MAAqB;AArTvC;AAsTI,QAAI,YAAY;AAChB,UAAM,YAAY,UAAU,KAAK,IAAI,CAAC;AAEtC,QAAI;AACF,YAAM,CAAC,OAAO,SAAS,IAAK,MAAM,KAAK,QAAQ,iBAAiB,QAAQ;AAKxE,YAAM,WAA4B,MAAM,IAAI,CAAC,UAAmC;AAAA,QAC9E,MAAM,KAAK;AAAA,QACX,OAAO,KAAK;AAAA,MACd,EAAE;AAEF,YAAM,uBAAuB,KAAK,UAAU,uBAAuB,KAAK,OAAO,IAAI;AACnF,YAAM,QACJ,wBAAwB,qBAAqB,SAAS,IAClD,CAAC,EAAE,qBAAqB,CAAC,IACzB;AAEN,UAAI,oBAA+C;AACnD,UAAI,UAAU,kBAAkB,UAAU,eAAe,SAAS,GAAG;AACnE,4BAAoB;AAAA,UAClB,OAAO,UAAU,eAAe,IAAI,CAAC,aAAqB,EAAE,MAAM,QAAQ,EAAE;AAAA,QAC9E;AAAA,MACF;AAEA,YAAM,WAAW,MAAM,KAAK,QAAQ,OAAO,sBAAsB;AAAA,QAC/D,OAAO,KAAK;AAAA,QACZ;AAAA,QACA,QAAQ;AAAA,UACN,GAAG,KAAK;AAAA,UACR;AAAA,UACA,aAAa,KAAK,aAAa,eAAe;AAAA,YAC5C,SAAS,KAAK,MAAM,KAAK,YAAY,SAAS;AAAA,UAChD;AAAA,UACA;AAAA,QACF;AAAA,MACF,CAAC;AAED,uBAAiB,SAAS,UAAU;AAClC,YAAI,MAAM,gBAAgB;AACxB,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,0BAA0B,KAAK,UAAU,MAAM,cAAc,CAAC;AAAA,YACvE,SAAS;AAAA,cACP,WAAW;AAAA,cACX;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAEA,YAAI,CAAC,MAAM,cAAc,GAAC,iBAAM,WAAW,CAAC,MAAlB,mBAAqB,YAArB,mBAA8B,QAAO;AAC7D,eAAK,OAAO,KAAK,kCAAkC,KAAK,UAAU,KAAK,CAAC,EAAE;AAC1E;AAAA,QACF;AAEA,YAAI,MAAM,WAAW,SAAS,GAAG;AAC/B,eAAK,OAAO;AAAA,YACV;AAAA,UACF;AAAA,QACF;AAEA,mBAAW,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ,OAAO;AACpD,gBAAM,YAAY,KAAK,WAAW,WAAW,IAAI;AACjD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,eAAe;AACvB,gBAAM,QAAQ,MAAM;AACpB,eAAK,MAAM,IAAI;AAAA,YACb,IAAI;AAAA,YACJ,OAAO;AAAA,cACL,kBAAkB,MAAM,wBAAwB;AAAA,cAChD,cAAc,MAAM,oBAAoB;AAAA,cACxC,oBAAoB,MAAM,2BAA2B;AAAA,cACrD,aAAa,MAAM,mBAAmB;AAAA,YACxC;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AACvB,YAAM,MAAM;AAOZ,UAAI,IAAI,QAAQ,IAAI,QAAQ,OAAO,IAAI,OAAO,KAAK;AACjD,YAAI,IAAI,SAAS,KAAK;AACpB,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,kCAAkC,IAAI,WAAW,eAAe;AAAA,YACzE,SAAS;AAAA,cACP,YAAY;AAAA,cACZ,WAAW;AAAA,YACb;AAAA,UACF,CAAC;AAAA,QACH,OAAO;AACL,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,6BAA6B,IAAI,IAAI,OAAO,IAAI,WAAW,eAAe;AAAA,YACnF,SAAS;AAAA,cACP,YAAY,IAAI;AAAA,cAChB,WAAW;AAAA,YACb;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAEA,UAAI,IAAI,QAAQ,IAAI,QAAQ,KAAK;AAC/B,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,6BAA6B,IAAI,IAAI,OAAO,IAAI,WAAW,eAAe;AAAA,UACnF,SAAS;AAAA,YACP,YAAY,IAAI;AAAA,YAChB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH;AAEA,YAAM,IAAI,mBAAmB;AAAA,QAC3B,SAAS,2BAA2B,IAAI,WAAW,eAAe;AAAA,QAClE,SAAS;AAAA,UACP;AAAA,QACF;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,WAAW,IAAY,MAAwC;AAC7D,QAAI,KAAK,cAAc;AACrB,aAAO;AAAA,QACL;AAAA,QACA,OAAO;AAAA,UACL,MAAM;AAAA,UACN,WAAW;AAAA,YACT,IAAI,aAAa,OAAO;AAAA,cACtB,QAAQ,KAAK,aAAa,MAAM,UAAU,gBAAgB;AAAA,cAC1D,MAAM,KAAK,aAAa;AAAA,cACxB,MAAM,KAAK,UAAU,KAAK,aAAa,IAAK;AAAA;AAAA,cAE5C,kBAAkB,KAAK;AAAA,YACzB,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,SAAS,KAAK;AAAA,QACd,MAAM;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACF;","names":["llm"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@livekit/agents-plugin-google",
-  "version": "1.0.30",
+  "version": "1.0.32",
   "description": "Google Gemini plugin for LiveKit Node Agents",
   "main": "dist/index.js",
   "require": "dist/index.cjs",
@@ -29,19 +29,19 @@
     "@microsoft/api-extractor": "^7.35.0",
     "tsup": "^8.3.5",
     "typescript": "^5.0.0",
-    "@livekit/agents": "1.0.30",
-    "@livekit/agents-plugin-openai": "1.0.30",
-    "@livekit/agents-plugins-test": "1.0.30"
+    "@livekit/agents": "1.0.32",
+    "@livekit/agents-plugin-openai": "1.0.32",
+    "@livekit/agents-plugins-test": "1.0.32"
   },
   "dependencies": {
-    "@google/genai": "^1.13.0",
+    "@google/genai": "^1.34.0",
     "@livekit/mutex": "^1.1.1",
     "@types/json-schema": "^7.0.15",
     "json-schema": "^0.4.0"
   },
   "peerDependencies": {
     "@livekit/rtc-node": "^0.13.22",
-    "@livekit/agents": "1.0.30"
+    "@livekit/agents": "1.0.32"
   },
   "scripts": {
     "build": "tsup --onSuccess \"pnpm build:types\"",

package/src/beta/realtime/api_proto.ts

@@ -5,18 +5,58 @@ import type * as types from '@google/genai';
 
 /**
  * Supported Google Live API models
+ *
+ * Gemini API deprecations: https://ai.google.dev/gemini-api/docs/deprecations
+ * Gemini API release notes with preview deprecations: https://ai.google.dev/gemini-api/docs/changelog
+ * Live models: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/live-api
+ * VertexAI retirement: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions#retired-models
+ * Additional references:
+ * 1. https://github.com/kazunori279/adk-streaming-test/blob/main/test_report.md
  */
 export type LiveAPIModels =
-  | 'gemini-2.0-flash-exp'
-  // models supported on Gemini API
-  | 'gemini-2.0-flash-live-001'
-  | 'gemini-2.5-flash-preview-native-audio-dialog'
-  | 'gemini-2.5-flash-exp-native-audio-thinking-dialog';
+  // VertexAI models
+  | 'gemini-live-2.5-flash-native-audio' // GA https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash-live-api#live-2.5-flash
+  | 'gemini-live-2.5-flash-preview-native-audio-09-2025' // Public preview https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash-live-api#live-2.5-flash-preview
+  | 'gemini-live-2.5-flash-preview-native-audio' // still works, possibly an alias, but not mentioned in any docs or changelog
+  // Gemini API models
+  | 'gemini-2.5-flash-native-audio-preview-12-2025' // https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-live
+  | 'gemini-2.5-flash-native-audio-preview-09-2025' // https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-live
+  | 'gemini-2.0-flash-exp'; // still works in Gemini API but not VertexAI
 
 /**
  * Available voice options for Google Realtime API
  */
-export type Voice = 'Puck' | 'Charon' | 'Kore' | 'Fenrir' | 'Aoede' | 'Leda' | 'Orus' | 'Zephyr';
+export type Voice =
+  | 'Achernar'
+  | 'Achird'
+  | 'Algenib'
+  | 'Algieba'
+  | 'Alnilam'
+  | 'Aoede'
+  | 'Autonoe'
+  | 'Callirrhoe'
+  | 'Charon'
+  | 'Despina'
+  | 'Enceladus'
+  | 'Erinome'
+  | 'Fenrir'
+  | 'Gacrux'
+  | 'Iapetus'
+  | 'Kore'
+  | 'Laomedeia'
+  | 'Leda'
+  | 'Orus'
+  | 'Pulcherrima'
+  | 'Puck'
+  | 'Rasalgethi'
+  | 'Sadachbia'
+  | 'Sadaltager'
+  | 'Schedar'
+  | 'Sulafat'
+  | 'Umbriel'
+  | 'Vindemiatrix'
+  | 'Zephyr'
+  | 'Zubenelgenubi';
 
 /**
  * Union type for all possible client events

package/src/beta/realtime/realtime_api.ts

@@ -102,6 +102,7 @@ interface RealtimeOptions {
   contextWindowCompression?: ContextWindowCompressionConfig;
   apiVersion?: string;
   geminiTools?: LLMTools;
+  thinkingConfig?: types.ThinkingConfig;
 }
 
 /**
@@ -136,6 +137,10 @@ export class RealtimeModel extends llm.RealtimeModel {
   /** @internal */
   _options: RealtimeOptions;
 
+  get model(): string {
+    return this._options.model;
+  }
+
   constructor(
     options: {
       /**
@@ -273,6 +278,14 @@ export class RealtimeModel extends llm.RealtimeModel {
        * Gemini-specific tools to use for the session
        */
       geminiTools?: LLMTools;
+
+      /**
+       * Thinking configuration for native audio models.
+       * If not set, the model's default thinking behavior is used.
+       * Use `\{ thinkingBudget: 0 \}` to disable thinking.
+       * Use `\{ thinkingBudget: -1 \}` for automatic/dynamic thinking.
+       */
+      thinkingConfig?: types.ThinkingConfig;
     } = {},
   ) {
     const inputAudioTranscription =
@@ -300,7 +313,9 @@ export class RealtimeModel extends llm.RealtimeModel {
     const vertexai = options.vertexai ?? false;
 
     // Model selection based on API type
-    const defaultModel = vertexai ? 'gemini-2.0-flash-exp' : 'gemini-2.0-flash-live-001';
+    const defaultModel = vertexai
+      ? 'gemini-live-2.5-flash-native-audio'
+      : 'gemini-2.5-flash-native-audio-preview-12-2025';
 
     this._options = {
       model: options.model || defaultModel,
@@ -330,6 +345,7 @@ export class RealtimeModel extends llm.RealtimeModel {
       contextWindowCompression: options.contextWindowCompression,
       apiVersion: options.apiVersion,
       geminiTools: options.geminiTools,
+      thinkingConfig: options.thinkingConfig,
     };
   }
 
@@ -934,9 +950,11 @@ export class RealtimeSession extends llm.RealtimeSession {
       unlock();
     }
 
+    // start new generation for serverContent or for standalone toolCalls (functionChannel closed)
     if (
       (!this.currentGeneration || this.currentGeneration._done) &&
-      (response.serverContent || response.toolCall)
+      (response.serverContent ||
+        (response.toolCall && this.currentGeneration?.functionChannel.closed !== false))
     ) {
       this.startNewGeneration();
     }
@@ -1034,7 +1052,7 @@ export class RealtimeSession extends llm.RealtimeSession {
     return obj;
   }
 
-  private markCurrentGenerationDone(): void {
+  private markCurrentGenerationDone(keepFunctionChannelOpen: boolean = false): void {
     if (!this.currentGeneration || this.currentGeneration._done) {
       return;
     }
@@ -1076,7 +1094,9 @@ export class RealtimeSession extends llm.RealtimeSession {
 
     gen.textChannel.close();
     gen.audioChannel.close();
-    gen.functionChannel.close();
+    if (!keepFunctionChannelOpen) {
+      gen.functionChannel.close();
+    }
     gen.messageChannel.close();
     gen._done = true;
   }
@@ -1095,6 +1115,7 @@ export class RealtimeSession extends llm.RealtimeSession {
     const opts = this.options;
 
     const config: types.LiveConnectConfig = {
+      thinkingConfig: opts.thinkingConfig,
       responseModalities: opts.responseModalities,
       systemInstruction: opts.instructions
         ? {
@@ -1156,6 +1177,11 @@ export class RealtimeSession extends llm.RealtimeSession {
   }
 
   private startNewGeneration(): void {
+    // close functionChannel of previous generation if still open (no toolCall arrived)
+    if (this.currentGeneration && !this.currentGeneration.functionChannel.closed) {
+      this.currentGeneration.functionChannel.close();
+    }
+
     if (this.currentGeneration && !this.currentGeneration._done) {
       this.#logger.warn('Starting new generation while another is active. Finalizing previous.');
       this.markCurrentGenerationDone();
@@ -1196,6 +1222,7 @@ export class RealtimeSession extends llm.RealtimeSession {
       messageStream: this.currentGeneration.messageChannel.stream(),
       functionStream: this.currentGeneration.functionChannel.stream(),
       userInitiated: false,
+      responseId,
     };
 
     if (this.pendingGenerationFut && !this.pendingGenerationFut.done) {
@@ -1233,6 +1260,11 @@ export class RealtimeSession extends llm.RealtimeSession {
       const turn = serverContent.modelTurn;
 
       for (const part of turn.parts || []) {
+        // bypass reasoning/thought output
+        if (part.thought) {
+          continue;
+        }
+
         if (part.text) {
           gen.outputText += part.text;
           gen.textChannel.write(part.text);
@@ -1301,7 +1333,8 @@ export class RealtimeSession extends llm.RealtimeSession {
     }
 
     if (serverContent.turnComplete) {
-      this.markCurrentGenerationDone();
+      // keep functionChannel open for potential late-arriving toolCalls
+      this.markCurrentGenerationDone(true);
     }
   }
 
@@ -1313,14 +1346,26 @@ export class RealtimeSession extends llm.RealtimeSession {
 
     const gen = this.currentGeneration;
 
+    if (gen.functionChannel.closed) {
+      this.#logger.warn('received tool call but functionChannel is already closed.');
+      return;
+    }
+
     for (const fc of toolCall.functionCalls || []) {
-      gen.functionChannel.write({
-        callId: fc.id || shortuuid('fnc-call-'),
-        name: fc.name,
-        args: fc.args ? JSON.stringify(fc.args) : '',
-      } as llm.FunctionCall);
+      if (!fc.name) {
+        this.#logger.warn('received function call without name, skipping');
+        continue;
+      }
+      gen.functionChannel.write(
+        llm.FunctionCall.create({
+          callId: fc.id || shortuuid('fnc-call-'),
+          name: fc.name,
+          args: fc.args ? JSON.stringify(fc.args) : '',
+        }),
+      );
     }
 
+    gen.functionChannel.close();
     this.markCurrentGenerationDone();
   }
 

package/src/llm.ts

@@ -449,6 +449,8 @@ export class LLMStream extends llm.LLMStream {
               callId: part.functionCall.id || shortuuid('function_call_'),
               name: part.functionCall.name!,
               args: JSON.stringify(part.functionCall.args!),
+              // Preserve thought signature for Gemini 3+ thinking mode
+              thoughtSignature: part.thoughtSignature,
             }),
           ],
         },