@livekit/agents-plugin-google 1.2.3 → 1.2.5

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  get provider(): string {\n    if (this.#opts.vertexai) {\n      return 'Vertex AI';\n    }\n    return 'Gemini';\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\nconst BLOCKED_REASONS = [\n  'SAFETY',\n  'SPII',\n  'PROHIBITED_CONTENT',\n  'BLOCKLIST',\n  'LANGUAGE',\n  'RECITATION',\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        // Check for blocked reasons first — safety-blocked responses often lack content.parts,\n        // so this must run before the no-content guard to avoid wasting retries.\n        if (\n          chunk.candidates?.[0]?.finishReason &&\n          BLOCKED_REASONS.includes(chunk.candidates[0].finishReason)\n        ) {\n          throw new APIStatusError({\n            message: `Google LLM: generation blocked - ${chunk.candidates[0].finishReason}`,\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 content in the response: ${JSON.stringify(chunk)}`);\n          if (retryable) {\n            throw new APIStatusError({\n              message: 'Google LLM: no content in the response',\n              options: {\n                retryable: true,\n                requestId,\n              },\n            });\n          }\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        const candidate = chunk.candidates[0];\n        const finishReason = candidate.finishReason;\n\n        let chunksYielded = false;\n        for (const part of candidate.content!.parts!) {\n          const chatChunk = this.#parsePart(requestId, part);\n          if (chatChunk) {\n            chunksYielded = true;\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (finishReason === 'STOP' && !chunksYielded && retryable) {\n          throw new APIStatusError({\n            message: 'Google LLM: no response generated',\n            options: {\n              retryable,\n              requestId,\n            },\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      if (error instanceof APIStatusError || error instanceof APIConnectionError) {\n        throw error;\n      }\n\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    if (!part.text) {\n      return null;\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,EAEA,IAAI,WAAmB;AACrB,QAAI,KAAK,MAAM,UAAU;AACvB,aAAO;AAAA,IACT;AACA,WAAO;AAAA,EACT;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;AAEA,MAAM,kBAAkB;AAAA,EACtB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;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;AArUvC;AAsUI,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;AAIA,cACE,iBAAM,eAAN,mBAAmB,OAAnB,mBAAuB,iBACvB,gBAAgB,SAAS,MAAM,WAAW,CAAC,EAAE,YAAY,GACzD;AACA,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,oCAAoC,MAAM,WAAW,CAAC,EAAE,YAAY;AAAA,YAC7E,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,+BAA+B,KAAK,UAAU,KAAK,CAAC,EAAE;AACvE,cAAI,WAAW;AACb,kBAAM,IAAI,eAAe;AAAA,cACvB,SAAS;AAAA,cACT,SAAS;AAAA,gBACP,WAAW;AAAA,gBACX;AAAA,cACF;AAAA,YACF,CAAC;AAAA,UACH;AACA;AAAA,QACF;AAEA,YAAI,MAAM,WAAW,SAAS,GAAG;AAC/B,eAAK,OAAO;AAAA,YACV;AAAA,UACF;AAAA,QACF;AAEA,cAAM,YAAY,MAAM,WAAW,CAAC;AACpC,cAAM,eAAe,UAAU;AAE/B,YAAI,gBAAgB;AACpB,mBAAW,QAAQ,UAAU,QAAS,OAAQ;AAC5C,gBAAM,YAAY,KAAK,WAAW,WAAW,IAAI;AACjD,cAAI,WAAW;AACb,4BAAgB;AAChB,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,iBAAiB,UAAU,CAAC,iBAAiB,WAAW;AAC1D,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS;AAAA,YACT,SAAS;AAAA,cACP;AAAA,cACA;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;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,UAAI,iBAAiB,kBAAkB,iBAAiB,oBAAoB;AAC1E,cAAM;AAAA,MACR;AAEA,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,QAAI,CAAC,KAAK,MAAM;AACd,aAAO;AAAA,IACT;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  get provider(): string {\n    if (this.#opts.vertexai) {\n      return 'Vertex AI';\n    }\n    return 'Gemini';\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. Gemini 3.1 models use `thinkingLevel`; Gemini 2.5 models use `thinkingBudget`. 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\nconst BLOCKED_REASONS = [\n  'SAFETY',\n  'SPII',\n  'PROHIBITED_CONTENT',\n  'BLOCKLIST',\n  'LANGUAGE',\n  'RECITATION',\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        // Check for blocked reasons first — safety-blocked responses often lack content.parts,\n        // so this must run before the no-content guard to avoid wasting retries.\n        if (\n          chunk.candidates?.[0]?.finishReason &&\n          BLOCKED_REASONS.includes(chunk.candidates[0].finishReason)\n        ) {\n          throw new APIStatusError({\n            message: `Google LLM: generation blocked - ${chunk.candidates[0].finishReason}`,\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 content in the response: ${JSON.stringify(chunk)}`);\n          if (retryable) {\n            throw new APIStatusError({\n              message: 'Google LLM: no content in the response',\n              options: {\n                retryable: true,\n                requestId,\n              },\n            });\n          }\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        const candidate = chunk.candidates[0];\n        const finishReason = candidate.finishReason;\n\n        let chunksYielded = false;\n        for (const part of candidate.content!.parts!) {\n          const chatChunk = this.#parsePart(requestId, part);\n          if (chatChunk) {\n            chunksYielded = true;\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (finishReason === 'STOP' && !chunksYielded && retryable) {\n          throw new APIStatusError({\n            message: 'Google LLM: no response generated',\n            options: {\n              retryable,\n              requestId,\n            },\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      if (error instanceof APIStatusError || error instanceof APIConnectionError) {\n        throw error;\n      }\n\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    if (!part.text) {\n      return null;\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,EAEA,IAAI,WAAmB;AACrB,QAAI,KAAK,MAAM,UAAU;AACvB,aAAO;AAAA,IACT;AACA,WAAO;AAAA,EACT;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;AAEA,MAAM,kBAAkB;AAAA,EACtB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;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;AArUvC;AAsUI,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;AAIA,cACE,iBAAM,eAAN,mBAAmB,OAAnB,mBAAuB,iBACvB,gBAAgB,SAAS,MAAM,WAAW,CAAC,EAAE,YAAY,GACzD;AACA,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS,oCAAoC,MAAM,WAAW,CAAC,EAAE,YAAY;AAAA,YAC7E,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,+BAA+B,KAAK,UAAU,KAAK,CAAC,EAAE;AACvE,cAAI,WAAW;AACb,kBAAM,IAAI,eAAe;AAAA,cACvB,SAAS;AAAA,cACT,SAAS;AAAA,gBACP,WAAW;AAAA,gBACX;AAAA,cACF;AAAA,YACF,CAAC;AAAA,UACH;AACA;AAAA,QACF;AAEA,YAAI,MAAM,WAAW,SAAS,GAAG;AAC/B,eAAK,OAAO;AAAA,YACV;AAAA,UACF;AAAA,QACF;AAEA,cAAM,YAAY,MAAM,WAAW,CAAC;AACpC,cAAM,eAAe,UAAU;AAE/B,YAAI,gBAAgB;AACpB,mBAAW,QAAQ,UAAU,QAAS,OAAQ;AAC5C,gBAAM,YAAY,KAAK,WAAW,WAAW,IAAI;AACjD,cAAI,WAAW;AACb,4BAAgB;AAChB,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,iBAAiB,UAAU,CAAC,iBAAiB,WAAW;AAC1D,gBAAM,IAAI,eAAe;AAAA,YACvB,SAAS;AAAA,YACT,SAAS;AAAA,cACP;AAAA,cACA;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;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,UAAI,iBAAiB,kBAAkB,iBAAiB,oBAAoB;AAC1E,cAAM;AAAA,MACR;AAEA,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,QAAI,CAAC,KAAK,MAAM;AACd,aAAO;AAAA,IACT;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": "1.2.3",
+  "version": "1.2.5",
   "description": "Google Gemini plugin for LiveKit Node Agents",
   "main": "dist/index.js",
   "require": "dist/index.cjs",
@@ -25,13 +25,13 @@
     "README.md"
   ],
   "devDependencies": {
-    "@livekit/rtc-node": "^0.13.24",
+    "@livekit/rtc-node": "^0.13.25",
     "@microsoft/api-extractor": "^7.35.0",
     "tsup": "^8.3.5",
     "typescript": "^5.0.0",
-    "@livekit/agents": "1.2.3",
-    "@livekit/agents-plugin-openai": "1.2.3",
-    "@livekit/agents-plugins-test": "1.2.3"
+    "@livekit/agents": "1.2.5",
+    "@livekit/agents-plugin-openai": "1.2.5",
+    "@livekit/agents-plugins-test": "1.2.5"
   },
   "dependencies": {
     "@google/genai": "^1.44.0",
@@ -40,8 +40,8 @@
     "json-schema": "^0.4.0"
   },
   "peerDependencies": {
-    "@livekit/rtc-node": "^0.13.24",
-    "@livekit/agents": "1.2.3"
+    "@livekit/rtc-node": "^0.13.25",
+    "@livekit/agents": "1.2.5"
   },
   "scripts": {
     "build": "tsup --onSuccess \"pnpm build:types\"",

package/src/beta/realtime/realtime_api.ts

@@ -288,8 +288,8 @@ export class RealtimeModel extends llm.RealtimeModel {
       /**
        * 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.
+       * Gemini 3.1 live models use `thinkingLevel`.
+       * Gemini 2.5 live models use `thinkingBudget`.
        */
       thinkingConfig?: types.ThinkingConfig;
     } = {},
@@ -311,6 +311,10 @@ export class RealtimeModel extends llm.RealtimeModel {
       autoToolReplyGeneration: true,
       audioOutput: options.modalities?.includes(Modality.AUDIO) ?? true,
       manualFunctionCalls: false,
+      midSessionChatCtxUpdate: false,
+      midSessionInstructionsUpdate: true,
+      midSessionToolsUpdate: false,
+      perResponseToolChoice: false,
     });
 
     // Environment variable fallbacks
@@ -543,6 +547,13 @@ export class RealtimeSession extends llm.RealtimeSession {
   }
 
   async updateInstructions(instructions: string): Promise<void> {
+    if (this.options.model === 'gemini-3.1-flash-live-preview') {
+      this.#logger.warn(
+        'updateInstructions is not compatible with gemini-3.1-flash-live-preview and will be ignored.',
+      );
+      this.options.instructions = instructions;
+      return;
+    }
     if (this.options.instructions === undefined || this.options.instructions !== instructions) {
       this.options.instructions = instructions;
       this.markRestartNeeded();
@@ -550,6 +561,13 @@ export class RealtimeSession extends llm.RealtimeSession {
   }
 
   async updateChatCtx(chatCtx: llm.ChatContext): Promise<void> {
+    if (this.options.model === 'gemini-3.1-flash-live-preview') {
+      this.#logger.warn(
+        'updateChatCtx is not compatible with gemini-3.1-flash-live-preview and will be ignored.',
+      );
+      this._chatCtx = chatCtx.copy();
+      return;
+    }
     const unlock = await this.sessionLock.lock();
     try {
       if (!this.activeSession) {
@@ -684,9 +702,11 @@ export class RealtimeSession extends llm.RealtimeSession {
   async generateReply(instructions?: string): Promise<llm.GenerationCreatedEvent> {
     if (this.options.model === 'gemini-3.1-flash-live-preview') {
       this.#logger.warn(
-        'generateReply is not compatible with gemini-3.1-flash-live-preview and will be ignored.',
+        'generateReply is not compatible with gemini-3.1-flash-live-preview. Use a Gemini 2.5 live model for voice-agent flows that require programmatic reply generation.',
+      );
+      throw new Error(
+        "generateReply is not compatible with 'gemini-3.1-flash-live-preview'; use a Gemini 2.5 live model for voice-agent flows that require programmatic reply generation.",
       );
-      throw new Error("generateReply is not compatible with 'gemini-3.1-flash-live-preview'");
     }
 
     if (this.pendingGenerationFut && !this.pendingGenerationFut.done) {
@@ -794,6 +814,7 @@ export class RealtimeSession extends llm.RealtimeSession {
     this.#closed = true;
 
     this.sessionShouldClose.set();
+    this.inputResampler?.close();
 
     await this.closeActiveSession();
 
@@ -1644,6 +1665,7 @@ export class RealtimeSession extends llm.RealtimeSession {
     if (this.inputResampler) {
       if (frame.sampleRate !== this.inputResamplerInputRate) {
         // input audio changed to a different sample rate
+        this.inputResampler.close();
         this.inputResampler = undefined;
         this.inputResamplerInputRate = undefined;
       }

package/src/llm.ts

@@ -80,7 +80,7 @@ export class LLM extends llm.LLM {
    * @param presencePenalty - Penalizes the model for generating previously mentioned concepts. Defaults to undefined.
    * @param frequencyPenalty - Penalizes the model for repeating words. Defaults to undefined.
    * @param toolChoice - Specifies whether to use tools during response generation. Defaults to "auto".
-   * @param thinkingConfig - The thinking configuration for response generation. Defaults to undefined.
+   * @param thinkingConfig - The thinking configuration for response generation. Gemini 3.1 models use `thinkingLevel`; Gemini 2.5 models use `thinkingBudget`. Defaults to undefined.
    * @param automaticFunctionCallingConfig - The automatic function calling configuration for response generation. Defaults to undefined.
    * @param geminiTools - The Gemini-specific tools to use for the session.
    * @param httpOptions - The HTTP options to use for the session.