@librechat/agents 3.3.13 → 3.4.0

package/dist/cjs/llm/google/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["ChatGoogleGenerativeAI","GenerativeAI","FunctionCallingMode","convertBaseMessagesToContent","dropUnsupportedModelTurnPrefill","mapGenerateContentResultToChatResult","convertResponseContentToChatGenerationChunk","ChatGenerationChunk","AIMessageChunk"],"sources":["../../../../src/llm/google/index.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/ban-ts-comment */\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport { ChatGoogleGenerativeAI } from '@langchain/google-genai';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport {\n  FunctionCallingMode,\n  GoogleGenerativeAI as GenerativeAI,\n} from '@google/generative-ai';\nimport type {\n  GenerateContentRequest,\n  SafetySetting,\n  ToolConfig,\n} from '@google/generative-ai';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { GeminiApiUsageMetadata, InputTokenDetails } from './types';\nimport type { GoogleClientOptions, GoogleThinkingConfig } from '@/types';\nimport {\n  convertResponseContentToChatGenerationChunk,\n  convertBaseMessagesToContent,\n  dropUnsupportedModelTurnPrefill,\n  mapGenerateContentResultToChatResult,\n} from './utils/common';\n\ntype GoogleToolConfigWithServerSideInvocations = ToolConfig & {\n  includeServerSideToolInvocations?: boolean;\n  functionCallingConfig?: Omit<\n    NonNullable<ToolConfig['functionCallingConfig']>,\n    'mode'\n  > & {\n    mode?:\n      | NonNullable<ToolConfig['functionCallingConfig']>['mode']\n      | 'VALIDATED';\n  };\n};\n\nexport class CustomChatGoogleGenerativeAI extends ChatGoogleGenerativeAI {\n  thinkingConfig?: GoogleThinkingConfig;\n  includeServerSideToolInvocations?: boolean;\n\n  /**\n   * Override to add gemini-3 model support for multimodal and function calling thought signatures\n   */\n  get _isMultimodalModel(): boolean {\n    return (\n      this.model.startsWith('gemini-1.5') ||\n      this.model.startsWith('gemini-2') ||\n      (this.model.startsWith('gemma-3-') &&\n        !this.model.startsWith('gemma-3-1b')) ||\n      this.model.startsWith('gemini-3')\n    );\n  }\n\n  constructor(fields: GoogleClientOptions) {\n    super(fields);\n\n    this.model = fields.model.replace(/^models\\//, '');\n\n    this.maxOutputTokens = fields.maxOutputTokens ?? this.maxOutputTokens;\n\n    if (this.maxOutputTokens != null && this.maxOutputTokens < 0) {\n      throw new Error('`maxOutputTokens` must be a positive integer');\n    }\n\n    this.temperature = fields.temperature ?? this.temperature;\n    if (\n      this.temperature != null &&\n      (this.temperature < 0 || this.temperature > 2)\n    ) {\n      throw new Error('`temperature` must be in the range of [0.0,2.0]');\n    }\n\n    this.topP = fields.topP ?? this.topP;\n    if (this.topP != null && this.topP < 0) {\n      throw new Error('`topP` must be a positive integer');\n    }\n\n    if (this.topP != null && this.topP > 1) {\n      throw new Error('`topP` must be below 1.');\n    }\n\n    this.topK = fields.topK ?? this.topK;\n    if (this.topK != null && this.topK < 0) {\n      throw new Error('`topK` must be a positive integer');\n    }\n\n    this.stopSequences = fields.stopSequences ?? this.stopSequences;\n\n    this.apiKey = fields.apiKey ?? getEnvironmentVariable('GOOGLE_API_KEY');\n    if (this.apiKey == null || this.apiKey === '') {\n      throw new Error(\n        'Please set an API key for Google GenerativeAI ' +\n          'in the environment variable GOOGLE_API_KEY ' +\n          'or in the `apiKey` field of the ' +\n          'ChatGoogleGenerativeAI constructor'\n      );\n    }\n\n    this.safetySettings = fields.safetySettings ?? this.safetySettings;\n    if (this.safetySettings && this.safetySettings.length > 0) {\n      const safetySettingsSet = new Set(\n        this.safetySettings.map((s) => s.category)\n      );\n      if (safetySettingsSet.size !== this.safetySettings.length) {\n        throw new Error(\n          'The categories in `safetySettings` array must be unique'\n        );\n      }\n    }\n\n    this.thinkingConfig = fields.thinkingConfig ?? this.thinkingConfig;\n    this.includeServerSideToolInvocations =\n      fields.includeServerSideToolInvocations ??\n      this.includeServerSideToolInvocations;\n\n    this.streaming = fields.streaming ?? this.streaming;\n    this.json = fields.json;\n\n    // @ts-ignore - Accessing private property from parent class\n    this.client = new GenerativeAI(this.apiKey).getGenerativeModel(\n      {\n        model: this.model,\n        safetySettings: this.safetySettings as SafetySetting[],\n        generationConfig: {\n          stopSequences: this.stopSequences,\n          maxOutputTokens: this.maxOutputTokens,\n          temperature: this.temperature,\n          topP: this.topP,\n          topK: this.topK,\n          ...(this.json != null\n            ? { responseMimeType: 'application/json' }\n            : {}),\n        },\n      },\n      {\n        apiVersion: fields.apiVersion,\n        baseUrl: fields.baseUrl,\n        customHeaders: fields.customHeaders,\n      }\n    );\n    this.streamUsage = fields.streamUsage ?? this.streamUsage;\n  }\n\n  static lc_name(): 'LibreChatGoogleGenerativeAI' {\n    return 'LibreChatGoogleGenerativeAI';\n  }\n\n  /**\n   * Helper function to convert Gemini API usage metadata to LangChain format\n   * Includes support for cached tokens and tier-based tracking for gemini-3-pro-preview\n   */\n  private _convertToUsageMetadata(\n    usageMetadata: GeminiApiUsageMetadata | undefined,\n    model: string\n  ): UsageMetadata | undefined {\n    if (!usageMetadata) {\n      return undefined;\n    }\n\n    const output: UsageMetadata = {\n      input_tokens: usageMetadata.promptTokenCount ?? 0,\n      output_tokens:\n        (usageMetadata.candidatesTokenCount ?? 0) +\n        (usageMetadata.thoughtsTokenCount ?? 0),\n      total_tokens: usageMetadata.totalTokenCount ?? 0,\n    };\n\n    if (usageMetadata.cachedContentTokenCount) {\n      output.input_token_details ??= {};\n      output.input_token_details.cache_read =\n        usageMetadata.cachedContentTokenCount;\n    }\n\n    // gemini-3-pro-preview has bracket based tracking of tokens per request\n    if (model === 'gemini-3-pro-preview') {\n      const over200k = Math.max(\n        0,\n        (usageMetadata.promptTokenCount ?? 0) - 200000\n      );\n      const cachedOver200k = Math.max(\n        0,\n        (usageMetadata.cachedContentTokenCount ?? 0) - 200000\n      );\n      if (over200k) {\n        output.input_token_details = {\n          ...output.input_token_details,\n          over_200k: over200k,\n        } as InputTokenDetails;\n      }\n      if (cachedOver200k) {\n        output.input_token_details = {\n          ...output.input_token_details,\n          cache_read_over_200k: cachedOver200k,\n        } as InputTokenDetails;\n      }\n    }\n\n    return output;\n  }\n\n  invocationParams(\n    options?: this['ParsedCallOptions']\n  ): Omit<GenerateContentRequest, 'contents'> {\n    const params = super.invocationParams(options);\n    if (this.thinkingConfig) {\n      /** @ts-ignore */\n      this.client.generationConfig = {\n        /** @ts-ignore */\n        ...this.client.generationConfig,\n        /** @ts-ignore */\n        thinkingConfig: this.thinkingConfig,\n      };\n    }\n    if (\n      this.includeServerSideToolInvocations === true &&\n      Array.isArray(params.tools) &&\n      params.tools.length > 0\n    ) {\n      const toolConfig = params.toolConfig as\n        | GoogleToolConfigWithServerSideInvocations\n        | undefined;\n      const functionCallingConfig = toolConfig?.functionCallingConfig;\n      params.toolConfig = {\n        ...toolConfig,\n        ...(functionCallingConfig?.mode === FunctionCallingMode.AUTO\n          ? {\n            functionCallingConfig: {\n              ...functionCallingConfig,\n              mode: 'VALIDATED',\n            },\n          }\n          : {}),\n        includeServerSideToolInvocations: true,\n      } as ToolConfig;\n    }\n    return params;\n  }\n\n  async _generate(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<import('@langchain/core/outputs').ChatResult> {\n    const prompt = convertBaseMessagesToContent(\n      messages,\n      this._isMultimodalModel,\n      this.useSystemInstruction,\n      this.model\n    );\n    let actualPrompt = prompt;\n    if (prompt?.[0].role === 'system') {\n      const [systemInstruction] = prompt;\n      /** @ts-ignore */\n      this.client.systemInstruction = systemInstruction;\n      actualPrompt = prompt.slice(1);\n    }\n    actualPrompt = dropUnsupportedModelTurnPrefill(actualPrompt, this.model);\n    const parameters = this.invocationParams(options);\n    const request = {\n      ...parameters,\n      contents: actualPrompt,\n    };\n\n    const res = await this.caller.callWithOptions(\n      { signal: options.signal },\n      async () =>\n        /** @ts-ignore */\n        this.client.generateContent(request)\n    );\n\n    const response = res.response;\n    const usageMetadata = this._convertToUsageMetadata(\n      /** @ts-ignore */\n      response.usageMetadata,\n      this.model\n    );\n\n    /** @ts-ignore */\n    const generationResult = mapGenerateContentResultToChatResult(response, {\n      usageMetadata,\n    });\n\n    await runManager?.handleLLMNewToken(\n      generationResult.generations[0].text || '',\n      undefined,\n      undefined,\n      undefined,\n      undefined,\n      undefined\n    );\n    return generationResult;\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const prompt = convertBaseMessagesToContent(\n      messages,\n      this._isMultimodalModel,\n      this.useSystemInstruction,\n      this.model\n    );\n    let actualPrompt = prompt;\n    if (prompt?.[0].role === 'system') {\n      const [systemInstruction] = prompt;\n      /** @ts-ignore */\n      this.client.systemInstruction = systemInstruction;\n      actualPrompt = prompt.slice(1);\n    }\n    actualPrompt = dropUnsupportedModelTurnPrefill(actualPrompt, this.model);\n    const parameters = this.invocationParams(options);\n    const request = {\n      ...parameters,\n      contents: actualPrompt,\n    };\n    const stream = await this.caller.callWithOptions(\n      { signal: options.signal },\n      async () => {\n        /** @ts-ignore */\n        const { stream } = await this.client.generateContentStream(request);\n        return stream;\n      }\n    );\n\n    let index = 0;\n    let lastUsageMetadata: UsageMetadata | undefined;\n    for await (const response of stream) {\n      if (\n        'usageMetadata' in response &&\n        this.streamUsage !== false &&\n        options.streamUsage !== false\n      ) {\n        lastUsageMetadata = this._convertToUsageMetadata(\n          response.usageMetadata as GeminiApiUsageMetadata | undefined,\n          this.model\n        );\n      }\n\n      const chunk = convertResponseContentToChatGenerationChunk(response, {\n        usageMetadata: undefined,\n        index,\n      });\n      index += 1;\n      if (!chunk) {\n        continue;\n      }\n\n      yield chunk;\n      await runManager?.handleLLMNewToken(\n        chunk.text || '',\n        undefined,\n        undefined,\n        undefined,\n        undefined,\n        { chunk }\n      );\n    }\n\n    if (lastUsageMetadata) {\n      const finalChunk = new ChatGenerationChunk({\n        text: '',\n        message: new AIMessageChunk({\n          content: '',\n          usage_metadata: lastUsageMetadata,\n        }),\n      });\n      yield finalChunk;\n      await runManager?.handleLLMNewToken(\n        finalChunk.text || '',\n        undefined,\n        undefined,\n        undefined,\n        undefined,\n        { chunk: finalChunk }\n      );\n    }\n  }\n}\n"],"mappings":";;;;;;;AAqCA,IAAa,+BAAb,cAAkDA,wBAAAA,uBAAuB;CACvE;CACA;;;;CAKA,IAAI,qBAA8B;EAChC,OACE,KAAK,MAAM,WAAW,YAAY,KAClC,KAAK,MAAM,WAAW,UAAU,KAC/B,KAAK,MAAM,WAAW,UAAU,KAC/B,CAAC,KAAK,MAAM,WAAW,YAAY,KACrC,KAAK,MAAM,WAAW,UAAU;CAEpC;CAEA,YAAY,QAA6B;EACvC,MAAM,MAAM;EAEZ,KAAK,QAAQ,OAAO,MAAM,QAAQ,aAAa,EAAE;EAEjD,KAAK,kBAAkB,OAAO,mBAAmB,KAAK;EAEtD,IAAI,KAAK,mBAAmB,QAAQ,KAAK,kBAAkB,GACzD,MAAM,IAAI,MAAM,8CAA8C;EAGhE,KAAK,cAAc,OAAO,eAAe,KAAK;EAC9C,IACE,KAAK,eAAe,SACnB,KAAK,cAAc,KAAK,KAAK,cAAc,IAE5C,MAAM,IAAI,MAAM,iDAAiD;EAGnE,KAAK,OAAO,OAAO,QAAQ,KAAK;EAChC,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,mCAAmC;EAGrD,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,yBAAyB;EAG3C,KAAK,OAAO,OAAO,QAAQ,KAAK;EAChC,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,mCAAmC;EAGrD,KAAK,gBAAgB,OAAO,iBAAiB,KAAK;EAElD,KAAK,SAAS,OAAO,WAAA,GAAA,0BAAA,uBAAA,CAAiC,gBAAgB;EACtE,IAAI,KAAK,UAAU,QAAQ,KAAK,WAAW,IACzC,MAAM,IAAI,MACR,6JAIF;EAGF,KAAK,iBAAiB,OAAO,kBAAkB,KAAK;EACpD,IAAI,KAAK,kBAAkB,KAAK,eAAe,SAAS;OAIlD,IAH0B,IAC5B,KAAK,eAAe,KAAK,MAAM,EAAE,QAAQ,CAEvB,CAAC,CAAC,SAAS,KAAK,eAAe,QACjD,MAAM,IAAI,MACR,yDACF;EAAA;EAIJ,KAAK,iBAAiB,OAAO,kBAAkB,KAAK;EACpD,KAAK,mCACH,OAAO,oCACP,KAAK;EAEP,KAAK,YAAY,OAAO,aAAa,KAAK;EAC1C,KAAK,OAAO,OAAO;EAGnB,KAAK,SAAS,IAAIC,sBAAAA,mBAAa,KAAK,MAAM,CAAC,CAAC,mBAC1C;GACE,OAAO,KAAK;GACZ,gBAAgB,KAAK;GACrB,kBAAkB;IAChB,eAAe,KAAK;IACpB,iBAAiB,KAAK;IACtB,aAAa,KAAK;IAClB,MAAM,KAAK;IACX,MAAM,KAAK;IACX,GAAI,KAAK,QAAQ,OACb,EAAE,kBAAkB,mBAAmB,IACvC,CAAC;GACP;EACF,GACA;GACE,YAAY,OAAO;GACnB,SAAS,OAAO;GAChB,eAAe,OAAO;EACxB,CACF;EACA,KAAK,cAAc,OAAO,eAAe,KAAK;CAChD;CAEA,OAAO,UAAyC;EAC9C,OAAO;CACT;;;;;CAMA,wBACE,eACA,OAC2B;EAC3B,IAAI,CAAC,eACH;EAGF,MAAM,SAAwB;GAC5B,cAAc,cAAc,oBAAoB;GAChD,gBACG,cAAc,wBAAwB,MACtC,cAAc,sBAAsB;GACvC,cAAc,cAAc,mBAAmB;EACjD;EAEA,IAAI,cAAc,yBAAyB;GACzC,OAAO,wBAAwB,CAAC;GAChC,OAAO,oBAAoB,aACzB,cAAc;EAClB;EAGA,IAAI,UAAU,wBAAwB;GACpC,MAAM,WAAW,KAAK,IACpB,IACC,cAAc,oBAAoB,KAAK,GAC1C;GACA,MAAM,iBAAiB,KAAK,IAC1B,IACC,cAAc,2BAA2B,KAAK,GACjD;GACA,IAAI,UACF,OAAO,sBAAsB;IAC3B,GAAG,OAAO;IACV,WAAW;GACb;GAEF,IAAI,gBACF,OAAO,sBAAsB;IAC3B,GAAG,OAAO;IACV,sBAAsB;GACxB;EAEJ;EAEA,OAAO;CACT;CAEA,iBACE,SAC0C;EAC1C,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK;;EAEP,KAAK,OAAO,mBAAmB;;GAE7B,GAAG,KAAK,OAAO;;GAEf,gBAAgB,KAAK;EACvB;EAEF,IACE,KAAK,qCAAqC,QAC1C,MAAM,QAAQ,OAAO,KAAK,KAC1B,OAAO,MAAM,SAAS,GACtB;GACA,MAAM,aAAa,OAAO;GAG1B,MAAM,wBAAwB,YAAY;GAC1C,OAAO,aAAa;IAClB,GAAG;IACH,GAAI,uBAAuB,SAASC,sBAAAA,oBAAoB,OACpD,EACA,uBAAuB;KACrB,GAAG;KACH,MAAM;IACR,EACF,IACE,CAAC;IACL,kCAAkC;GACpC;EACF;EACA,OAAO;CACT;CAEA,MAAM,UACJ,UACA,SACA,YACuD;EACvD,MAAM,SAASC,eAAAA,6BACb,UACA,KAAK,oBACL,KAAK,sBACL,KAAK,KACP;EACA,IAAI,eAAe;EACnB,IAAI,SAAS,EAAE,CAAC,SAAS,UAAU;GACjC,MAAM,CAAC,qBAAqB;;GAE5B,KAAK,OAAO,oBAAoB;GAChC,eAAe,OAAO,MAAM,CAAC;EAC/B;EACA,eAAeC,eAAAA,gCAAgC,cAAc,KAAK,KAAK;EAEvE,MAAM,UAAU;GACd,GAFiB,KAAK,iBAAiB,OAE3B;GACZ,UAAU;EACZ;EASA,MAAM,YAAW,MAPC,KAAK,OAAO,gBAC5B,EAAE,QAAQ,QAAQ,OAAO,GACzB,YAEE,KAAK,OAAO,gBAAgB,OAAO,CACvC,EAAA,CAEqB;;EAQrB,MAAM,mBAAmBC,eAAAA,qCAAqC,UAAU,EACtE,eARoB,KAAK;;GAEzB,SAAS;GACT,KAAK;EAKO,EACd,CAAC;EAED,MAAM,YAAY,kBAChB,iBAAiB,YAAY,EAAE,CAAC,QAAQ,IACxC,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,CACF;EACA,OAAO;CACT;CAEA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,MAAM,SAASF,eAAAA,6BACb,UACA,KAAK,oBACL,KAAK,sBACL,KAAK,KACP;EACA,IAAI,eAAe;EACnB,IAAI,SAAS,EAAE,CAAC,SAAS,UAAU;GACjC,MAAM,CAAC,qBAAqB;;GAE5B,KAAK,OAAO,oBAAoB;GAChC,eAAe,OAAO,MAAM,CAAC;EAC/B;EACA,eAAeC,eAAAA,gCAAgC,cAAc,KAAK,KAAK;EAEvE,MAAM,UAAU;GACd,GAFiB,KAAK,iBAAiB,OAE3B;GACZ,UAAU;EACZ;EACA,MAAM,SAAS,MAAM,KAAK,OAAO,gBAC/B,EAAE,QAAQ,QAAQ,OAAO,GACzB,YAAY;;GAEV,MAAM,EAAE,WAAW,MAAM,KAAK,OAAO,sBAAsB,OAAO;GAClE,OAAO;EACT,CACF;EAEA,IAAI,QAAQ;EACZ,IAAI;EACJ,WAAW,MAAM,YAAY,QAAQ;GACnC,IACE,mBAAmB,YACnB,KAAK,gBAAgB,SACrB,QAAQ,gBAAgB,OAExB,oBAAoB,KAAK,wBACvB,SAAS,eACT,KAAK,KACP;GAGF,MAAM,QAAQE,eAAAA,4CAA4C,UAAU;IAClE,eAAe,KAAA;IACf;GACF,CAAC;GACD,SAAS;GACT,IAAI,CAAC,OACH;GAGF,MAAM;GACN,MAAM,YAAY,kBAChB,MAAM,QAAQ,IACd,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,MAAM,CACV;EACF;EAEA,IAAI,mBAAmB;GACrB,MAAM,aAAa,IAAIC,wBAAAA,oBAAoB;IACzC,MAAM;IACN,SAAS,IAAIC,yBAAAA,eAAe;KAC1B,SAAS;KACT,gBAAgB;IAClB,CAAC;GACH,CAAC;GACD,MAAM;GACN,MAAM,YAAY,kBAChB,WAAW,QAAQ,IACnB,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,WAAW,CACtB;EACF;CACF;AACF"}
+{"version":3,"file":"index.cjs","names":["ChatGoogleGenerativeAI","resolveStreamDelay","GenerativeAI","FunctionCallingMode","convertBaseMessagesToContent","dropUnsupportedModelTurnPrefill","mapGenerateContentResultToChatResult","smoothGenerationChunks","convertResponseContentToChatGenerationChunk","ChatGenerationChunk","AIMessageChunk"],"sources":["../../../../src/llm/google/index.ts"],"sourcesContent":["/* eslint-disable @typescript-eslint/ban-ts-comment */\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk } from '@langchain/core/outputs';\nimport { ChatGoogleGenerativeAI } from '@langchain/google-genai';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport {\n  FunctionCallingMode,\n  GoogleGenerativeAI as GenerativeAI,\n} from '@google/generative-ai';\nimport type {\n  GenerateContentRequest,\n  SafetySetting,\n  ToolConfig,\n} from '@google/generative-ai';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport type { GeminiApiUsageMetadata, InputTokenDetails } from './types';\nimport type { GoogleClientOptions, GoogleThinkingConfig } from '@/types';\nimport { smoothGenerationChunks } from '@/llm/stream/chunkAdapters';\nimport { resolveStreamDelay } from '@/llm/stream/smoother';\nimport {\n  convertResponseContentToChatGenerationChunk,\n  convertBaseMessagesToContent,\n  dropUnsupportedModelTurnPrefill,\n  mapGenerateContentResultToChatResult,\n} from './utils/common';\n\ntype GoogleToolConfigWithServerSideInvocations = ToolConfig & {\n  includeServerSideToolInvocations?: boolean;\n  functionCallingConfig?: Omit<\n    NonNullable<ToolConfig['functionCallingConfig']>,\n    'mode'\n  > & {\n    mode?:\n      | NonNullable<ToolConfig['functionCallingConfig']>['mode']\n      | 'VALIDATED';\n  };\n};\n\nexport class CustomChatGoogleGenerativeAI extends ChatGoogleGenerativeAI {\n  _lc_stream_delay: number;\n  thinkingConfig?: GoogleThinkingConfig;\n  includeServerSideToolInvocations?: boolean;\n\n  /**\n   * Override to add gemini-3 model support for multimodal and function calling thought signatures\n   */\n  get _isMultimodalModel(): boolean {\n    return (\n      this.model.startsWith('gemini-1.5') ||\n      this.model.startsWith('gemini-2') ||\n      (this.model.startsWith('gemma-3-') &&\n        !this.model.startsWith('gemma-3-1b')) ||\n      this.model.startsWith('gemini-3')\n    );\n  }\n\n  constructor(fields: GoogleClientOptions) {\n    super(fields);\n\n    this._lc_stream_delay = resolveStreamDelay(fields._lc_stream_delay);\n    this.model = fields.model.replace(/^models\\//, '');\n\n    this.maxOutputTokens = fields.maxOutputTokens ?? this.maxOutputTokens;\n\n    if (this.maxOutputTokens != null && this.maxOutputTokens < 0) {\n      throw new Error('`maxOutputTokens` must be a positive integer');\n    }\n\n    this.temperature = fields.temperature ?? this.temperature;\n    if (\n      this.temperature != null &&\n      (this.temperature < 0 || this.temperature > 2)\n    ) {\n      throw new Error('`temperature` must be in the range of [0.0,2.0]');\n    }\n\n    this.topP = fields.topP ?? this.topP;\n    if (this.topP != null && this.topP < 0) {\n      throw new Error('`topP` must be a positive integer');\n    }\n\n    if (this.topP != null && this.topP > 1) {\n      throw new Error('`topP` must be below 1.');\n    }\n\n    this.topK = fields.topK ?? this.topK;\n    if (this.topK != null && this.topK < 0) {\n      throw new Error('`topK` must be a positive integer');\n    }\n\n    this.stopSequences = fields.stopSequences ?? this.stopSequences;\n\n    this.apiKey = fields.apiKey ?? getEnvironmentVariable('GOOGLE_API_KEY');\n    if (this.apiKey == null || this.apiKey === '') {\n      throw new Error(\n        'Please set an API key for Google GenerativeAI ' +\n          'in the environment variable GOOGLE_API_KEY ' +\n          'or in the `apiKey` field of the ' +\n          'ChatGoogleGenerativeAI constructor'\n      );\n    }\n\n    this.safetySettings = fields.safetySettings ?? this.safetySettings;\n    if (this.safetySettings && this.safetySettings.length > 0) {\n      const safetySettingsSet = new Set(\n        this.safetySettings.map((s) => s.category)\n      );\n      if (safetySettingsSet.size !== this.safetySettings.length) {\n        throw new Error(\n          'The categories in `safetySettings` array must be unique'\n        );\n      }\n    }\n\n    this.thinkingConfig = fields.thinkingConfig ?? this.thinkingConfig;\n    this.includeServerSideToolInvocations =\n      fields.includeServerSideToolInvocations ??\n      this.includeServerSideToolInvocations;\n\n    this.streaming = fields.streaming ?? this.streaming;\n    this.json = fields.json;\n\n    // @ts-ignore - Accessing private property from parent class\n    this.client = new GenerativeAI(this.apiKey).getGenerativeModel(\n      {\n        model: this.model,\n        safetySettings: this.safetySettings as SafetySetting[],\n        generationConfig: {\n          stopSequences: this.stopSequences,\n          maxOutputTokens: this.maxOutputTokens,\n          temperature: this.temperature,\n          topP: this.topP,\n          topK: this.topK,\n          ...(this.json != null\n            ? { responseMimeType: 'application/json' }\n            : {}),\n        },\n      },\n      {\n        apiVersion: fields.apiVersion,\n        baseUrl: fields.baseUrl,\n        customHeaders: fields.customHeaders,\n      }\n    );\n    this.streamUsage = fields.streamUsage ?? this.streamUsage;\n  }\n\n  static lc_name(): 'LibreChatGoogleGenerativeAI' {\n    return 'LibreChatGoogleGenerativeAI';\n  }\n\n  /**\n   * Helper function to convert Gemini API usage metadata to LangChain format\n   * Includes support for cached tokens and tier-based tracking for gemini-3-pro-preview\n   */\n  private _convertToUsageMetadata(\n    usageMetadata: GeminiApiUsageMetadata | undefined,\n    model: string\n  ): UsageMetadata | undefined {\n    if (!usageMetadata) {\n      return undefined;\n    }\n\n    const output: UsageMetadata = {\n      input_tokens: usageMetadata.promptTokenCount ?? 0,\n      output_tokens:\n        (usageMetadata.candidatesTokenCount ?? 0) +\n        (usageMetadata.thoughtsTokenCount ?? 0),\n      total_tokens: usageMetadata.totalTokenCount ?? 0,\n    };\n\n    if (usageMetadata.cachedContentTokenCount) {\n      output.input_token_details ??= {};\n      output.input_token_details.cache_read =\n        usageMetadata.cachedContentTokenCount;\n    }\n\n    // gemini-3-pro-preview has bracket based tracking of tokens per request\n    if (model === 'gemini-3-pro-preview') {\n      const over200k = Math.max(\n        0,\n        (usageMetadata.promptTokenCount ?? 0) - 200000\n      );\n      const cachedOver200k = Math.max(\n        0,\n        (usageMetadata.cachedContentTokenCount ?? 0) - 200000\n      );\n      if (over200k) {\n        output.input_token_details = {\n          ...output.input_token_details,\n          over_200k: over200k,\n        } as InputTokenDetails;\n      }\n      if (cachedOver200k) {\n        output.input_token_details = {\n          ...output.input_token_details,\n          cache_read_over_200k: cachedOver200k,\n        } as InputTokenDetails;\n      }\n    }\n\n    return output;\n  }\n\n  invocationParams(\n    options?: this['ParsedCallOptions']\n  ): Omit<GenerateContentRequest, 'contents'> {\n    const params = super.invocationParams(options);\n    if (this.thinkingConfig) {\n      /** @ts-ignore */\n      this.client.generationConfig = {\n        /** @ts-ignore */\n        ...this.client.generationConfig,\n        /** @ts-ignore */\n        thinkingConfig: this.thinkingConfig,\n      };\n    }\n    if (\n      this.includeServerSideToolInvocations === true &&\n      Array.isArray(params.tools) &&\n      params.tools.length > 0\n    ) {\n      const toolConfig = params.toolConfig as\n        | GoogleToolConfigWithServerSideInvocations\n        | undefined;\n      const functionCallingConfig = toolConfig?.functionCallingConfig;\n      params.toolConfig = {\n        ...toolConfig,\n        ...(functionCallingConfig?.mode === FunctionCallingMode.AUTO\n          ? {\n            functionCallingConfig: {\n              ...functionCallingConfig,\n              mode: 'VALIDATED',\n            },\n          }\n          : {}),\n        includeServerSideToolInvocations: true,\n      } as ToolConfig;\n    }\n    return params;\n  }\n\n  async _generate(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<import('@langchain/core/outputs').ChatResult> {\n    const prompt = convertBaseMessagesToContent(\n      messages,\n      this._isMultimodalModel,\n      this.useSystemInstruction,\n      this.model\n    );\n    let actualPrompt = prompt;\n    if (prompt?.[0].role === 'system') {\n      const [systemInstruction] = prompt;\n      /** @ts-ignore */\n      this.client.systemInstruction = systemInstruction;\n      actualPrompt = prompt.slice(1);\n    }\n    actualPrompt = dropUnsupportedModelTurnPrefill(actualPrompt, this.model);\n    const parameters = this.invocationParams(options);\n    const request = {\n      ...parameters,\n      contents: actualPrompt,\n    };\n\n    const res = await this.caller.callWithOptions(\n      { signal: options.signal },\n      async () =>\n        /** @ts-ignore */\n        this.client.generateContent(request)\n    );\n\n    const response = res.response;\n    const usageMetadata = this._convertToUsageMetadata(\n      /** @ts-ignore */\n      response.usageMetadata,\n      this.model\n    );\n\n    /** @ts-ignore */\n    const generationResult = mapGenerateContentResultToChatResult(response, {\n      usageMetadata,\n    });\n\n    await runManager?.handleLLMNewToken(\n      generationResult.generations[0].text || '',\n      undefined,\n      undefined,\n      undefined,\n      undefined,\n      undefined\n    );\n    return generationResult;\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    yield* smoothGenerationChunks({\n      chunks: this._streamProviderChunks(messages, options),\n      delayMs: this._lc_stream_delay,\n      signal: options.signal,\n      runManager,\n    });\n  }\n\n  private async *_streamProviderChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions']\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const prompt = convertBaseMessagesToContent(\n      messages,\n      this._isMultimodalModel,\n      this.useSystemInstruction,\n      this.model\n    );\n    let actualPrompt = prompt;\n    if (prompt?.[0].role === 'system') {\n      const [systemInstruction] = prompt;\n      /** @ts-ignore */\n      this.client.systemInstruction = systemInstruction;\n      actualPrompt = prompt.slice(1);\n    }\n    actualPrompt = dropUnsupportedModelTurnPrefill(actualPrompt, this.model);\n    const parameters = this.invocationParams(options);\n    const request = {\n      ...parameters,\n      contents: actualPrompt,\n    };\n    const stream = await this.caller.callWithOptions(\n      { signal: options.signal },\n      async () => {\n        /** @ts-ignore */\n        const { stream } = await this.client.generateContentStream(request);\n        return stream;\n      }\n    );\n\n    let index = 0;\n    let lastUsageMetadata: UsageMetadata | undefined;\n    for await (const response of stream) {\n      if (\n        'usageMetadata' in response &&\n        this.streamUsage !== false &&\n        options.streamUsage !== false\n      ) {\n        lastUsageMetadata = this._convertToUsageMetadata(\n          response.usageMetadata as GeminiApiUsageMetadata | undefined,\n          this.model\n        );\n      }\n\n      const chunk = convertResponseContentToChatGenerationChunk(response, {\n        usageMetadata: undefined,\n        index,\n      });\n      index += 1;\n      if (!chunk) {\n        continue;\n      }\n\n      yield chunk;\n    }\n\n    if (lastUsageMetadata) {\n      yield new ChatGenerationChunk({\n        text: '',\n        message: new AIMessageChunk({\n          content: '',\n          usage_metadata: lastUsageMetadata,\n        }),\n      });\n    }\n  }\n}\n"],"mappings":";;;;;;;;;AAuCA,IAAa,+BAAb,cAAkDA,wBAAAA,uBAAuB;CACvE;CACA;CACA;;;;CAKA,IAAI,qBAA8B;EAChC,OACE,KAAK,MAAM,WAAW,YAAY,KAClC,KAAK,MAAM,WAAW,UAAU,KAC/B,KAAK,MAAM,WAAW,UAAU,KAC/B,CAAC,KAAK,MAAM,WAAW,YAAY,KACrC,KAAK,MAAM,WAAW,UAAU;CAEpC;CAEA,YAAY,QAA6B;EACvC,MAAM,MAAM;EAEZ,KAAK,mBAAmBC,iBAAAA,mBAAmB,OAAO,gBAAgB;EAClE,KAAK,QAAQ,OAAO,MAAM,QAAQ,aAAa,EAAE;EAEjD,KAAK,kBAAkB,OAAO,mBAAmB,KAAK;EAEtD,IAAI,KAAK,mBAAmB,QAAQ,KAAK,kBAAkB,GACzD,MAAM,IAAI,MAAM,8CAA8C;EAGhE,KAAK,cAAc,OAAO,eAAe,KAAK;EAC9C,IACE,KAAK,eAAe,SACnB,KAAK,cAAc,KAAK,KAAK,cAAc,IAE5C,MAAM,IAAI,MAAM,iDAAiD;EAGnE,KAAK,OAAO,OAAO,QAAQ,KAAK;EAChC,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,mCAAmC;EAGrD,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,yBAAyB;EAG3C,KAAK,OAAO,OAAO,QAAQ,KAAK;EAChC,IAAI,KAAK,QAAQ,QAAQ,KAAK,OAAO,GACnC,MAAM,IAAI,MAAM,mCAAmC;EAGrD,KAAK,gBAAgB,OAAO,iBAAiB,KAAK;EAElD,KAAK,SAAS,OAAO,WAAA,GAAA,0BAAA,uBAAA,CAAiC,gBAAgB;EACtE,IAAI,KAAK,UAAU,QAAQ,KAAK,WAAW,IACzC,MAAM,IAAI,MACR,6JAIF;EAGF,KAAK,iBAAiB,OAAO,kBAAkB,KAAK;EACpD,IAAI,KAAK,kBAAkB,KAAK,eAAe,SAAS;OAIlD,IAH0B,IAC5B,KAAK,eAAe,KAAK,MAAM,EAAE,QAAQ,CAEvB,CAAC,CAAC,SAAS,KAAK,eAAe,QACjD,MAAM,IAAI,MACR,yDACF;EAAA;EAIJ,KAAK,iBAAiB,OAAO,kBAAkB,KAAK;EACpD,KAAK,mCACH,OAAO,oCACP,KAAK;EAEP,KAAK,YAAY,OAAO,aAAa,KAAK;EAC1C,KAAK,OAAO,OAAO;EAGnB,KAAK,SAAS,IAAIC,sBAAAA,mBAAa,KAAK,MAAM,CAAC,CAAC,mBAC1C;GACE,OAAO,KAAK;GACZ,gBAAgB,KAAK;GACrB,kBAAkB;IAChB,eAAe,KAAK;IACpB,iBAAiB,KAAK;IACtB,aAAa,KAAK;IAClB,MAAM,KAAK;IACX,MAAM,KAAK;IACX,GAAI,KAAK,QAAQ,OACb,EAAE,kBAAkB,mBAAmB,IACvC,CAAC;GACP;EACF,GACA;GACE,YAAY,OAAO;GACnB,SAAS,OAAO;GAChB,eAAe,OAAO;EACxB,CACF;EACA,KAAK,cAAc,OAAO,eAAe,KAAK;CAChD;CAEA,OAAO,UAAyC;EAC9C,OAAO;CACT;;;;;CAMA,wBACE,eACA,OAC2B;EAC3B,IAAI,CAAC,eACH;EAGF,MAAM,SAAwB;GAC5B,cAAc,cAAc,oBAAoB;GAChD,gBACG,cAAc,wBAAwB,MACtC,cAAc,sBAAsB;GACvC,cAAc,cAAc,mBAAmB;EACjD;EAEA,IAAI,cAAc,yBAAyB;GACzC,OAAO,wBAAwB,CAAC;GAChC,OAAO,oBAAoB,aACzB,cAAc;EAClB;EAGA,IAAI,UAAU,wBAAwB;GACpC,MAAM,WAAW,KAAK,IACpB,IACC,cAAc,oBAAoB,KAAK,GAC1C;GACA,MAAM,iBAAiB,KAAK,IAC1B,IACC,cAAc,2BAA2B,KAAK,GACjD;GACA,IAAI,UACF,OAAO,sBAAsB;IAC3B,GAAG,OAAO;IACV,WAAW;GACb;GAEF,IAAI,gBACF,OAAO,sBAAsB;IAC3B,GAAG,OAAO;IACV,sBAAsB;GACxB;EAEJ;EAEA,OAAO;CACT;CAEA,iBACE,SAC0C;EAC1C,MAAM,SAAS,MAAM,iBAAiB,OAAO;EAC7C,IAAI,KAAK;;EAEP,KAAK,OAAO,mBAAmB;;GAE7B,GAAG,KAAK,OAAO;;GAEf,gBAAgB,KAAK;EACvB;EAEF,IACE,KAAK,qCAAqC,QAC1C,MAAM,QAAQ,OAAO,KAAK,KAC1B,OAAO,MAAM,SAAS,GACtB;GACA,MAAM,aAAa,OAAO;GAG1B,MAAM,wBAAwB,YAAY;GAC1C,OAAO,aAAa;IAClB,GAAG;IACH,GAAI,uBAAuB,SAASC,sBAAAA,oBAAoB,OACpD,EACA,uBAAuB;KACrB,GAAG;KACH,MAAM;IACR,EACF,IACE,CAAC;IACL,kCAAkC;GACpC;EACF;EACA,OAAO;CACT;CAEA,MAAM,UACJ,UACA,SACA,YACuD;EACvD,MAAM,SAASC,eAAAA,6BACb,UACA,KAAK,oBACL,KAAK,sBACL,KAAK,KACP;EACA,IAAI,eAAe;EACnB,IAAI,SAAS,EAAE,CAAC,SAAS,UAAU;GACjC,MAAM,CAAC,qBAAqB;;GAE5B,KAAK,OAAO,oBAAoB;GAChC,eAAe,OAAO,MAAM,CAAC;EAC/B;EACA,eAAeC,eAAAA,gCAAgC,cAAc,KAAK,KAAK;EAEvE,MAAM,UAAU;GACd,GAFiB,KAAK,iBAAiB,OAE3B;GACZ,UAAU;EACZ;EASA,MAAM,YAAW,MAPC,KAAK,OAAO,gBAC5B,EAAE,QAAQ,QAAQ,OAAO,GACzB,YAEE,KAAK,OAAO,gBAAgB,OAAO,CACvC,EAAA,CAEqB;;EAQrB,MAAM,mBAAmBC,eAAAA,qCAAqC,UAAU,EACtE,eARoB,KAAK;;GAEzB,SAAS;GACT,KAAK;EAKO,EACd,CAAC;EAED,MAAM,YAAY,kBAChB,iBAAiB,YAAY,EAAE,CAAC,QAAQ,IACxC,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,CACF;EACA,OAAO;CACT;CAEA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,OAAOC,sBAAAA,uBAAuB;GAC5B,QAAQ,KAAK,sBAAsB,UAAU,OAAO;GACpD,SAAS,KAAK;GACd,QAAQ,QAAQ;GAChB;EACF,CAAC;CACH;CAEA,OAAe,sBACb,UACA,SACqC;EACrC,MAAM,SAASH,eAAAA,6BACb,UACA,KAAK,oBACL,KAAK,sBACL,KAAK,KACP;EACA,IAAI,eAAe;EACnB,IAAI,SAAS,EAAE,CAAC,SAAS,UAAU;GACjC,MAAM,CAAC,qBAAqB;;GAE5B,KAAK,OAAO,oBAAoB;GAChC,eAAe,OAAO,MAAM,CAAC;EAC/B;EACA,eAAeC,eAAAA,gCAAgC,cAAc,KAAK,KAAK;EAEvE,MAAM,UAAU;GACd,GAFiB,KAAK,iBAAiB,OAE3B;GACZ,UAAU;EACZ;EACA,MAAM,SAAS,MAAM,KAAK,OAAO,gBAC/B,EAAE,QAAQ,QAAQ,OAAO,GACzB,YAAY;;GAEV,MAAM,EAAE,WAAW,MAAM,KAAK,OAAO,sBAAsB,OAAO;GAClE,OAAO;EACT,CACF;EAEA,IAAI,QAAQ;EACZ,IAAI;EACJ,WAAW,MAAM,YAAY,QAAQ;GACnC,IACE,mBAAmB,YACnB,KAAK,gBAAgB,SACrB,QAAQ,gBAAgB,OAExB,oBAAoB,KAAK,wBACvB,SAAS,eACT,KAAK,KACP;GAGF,MAAM,QAAQG,eAAAA,4CAA4C,UAAU;IAClE,eAAe,KAAA;IACf;GACF,CAAC;GACD,SAAS;GACT,IAAI,CAAC,OACH;GAGF,MAAM;EACR;EAEA,IAAI,mBACF,MAAM,IAAIC,wBAAAA,oBAAoB;GAC5B,MAAM;GACN,SAAS,IAAIC,yBAAAA,eAAe;IAC1B,SAAS;IACT,gBAAgB;GAClB,CAAC;EACH,CAAC;CAEL;AACF"}

package/dist/cjs/llm/mistral/index.cjs

@@ -0,0 +1,26 @@
+const require_smoother = require("../stream/smoother.cjs");
+const require_chunkAdapters = require("../stream/chunkAdapters.cjs");
+let _langchain_mistralai = require("@langchain/mistralai");
+//#region src/llm/mistral/index.ts
+var CustomChatMistralAI = class extends _langchain_mistralai.ChatMistralAI {
+	_lc_stream_delay;
+	static lc_name() {
+		return "LibreChatMistralAI";
+	}
+	constructor(fields) {
+		super(fields);
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
+	}
+	async *_streamResponseChunks(messages, options, runManager) {
+		yield* require_chunkAdapters.smoothGenerationChunks({
+			chunks: super._streamResponseChunks(messages, options, void 0),
+			delayMs: this._lc_stream_delay,
+			signal: options.signal,
+			runManager
+		});
+	}
+};
+//#endregion
+exports.CustomChatMistralAI = CustomChatMistralAI;
+
+//# sourceMappingURL=index.cjs.map

package/dist/cjs/llm/mistral/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.cjs","names":["ChatMistralAI","resolveStreamDelay","smoothGenerationChunks"],"sources":["../../../../src/llm/mistral/index.ts"],"sourcesContent":["import { ChatMistralAI } from '@langchain/mistralai';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatGenerationChunk } from '@langchain/core/outputs';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport type { MistralAIClientOptions } from '@/types';\nimport { smoothGenerationChunks } from '@/llm/stream/chunkAdapters';\nimport { resolveStreamDelay } from '@/llm/stream/smoother';\n\nexport class CustomChatMistralAI extends ChatMistralAI {\n  _lc_stream_delay: number;\n\n  static lc_name(): 'LibreChatMistralAI' {\n    return 'LibreChatMistralAI';\n  }\n\n  constructor(fields?: MistralAIClientOptions) {\n    super(fields);\n    this._lc_stream_delay = resolveStreamDelay(fields?._lc_stream_delay);\n  }\n\n  async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    yield* smoothGenerationChunks({\n      chunks: super._streamResponseChunks(messages, options, undefined),\n      delayMs: this._lc_stream_delay,\n      signal: options.signal,\n      runManager,\n    });\n  }\n}\n"],"mappings":";;;;AAQA,IAAa,sBAAb,cAAyCA,qBAAAA,cAAc;CACrD;CAEA,OAAO,UAAgC;EACrC,OAAO;CACT;CAEA,YAAY,QAAiC;EAC3C,MAAM,MAAM;EACZ,KAAK,mBAAmBC,iBAAAA,mBAAmB,QAAQ,gBAAgB;CACrE;CAEA,OAAO,sBACL,UACA,SACA,YACqC;EACrC,OAAOC,sBAAAA,uBAAuB;GAC5B,QAAQ,MAAM,sBAAsB,UAAU,SAAS,KAAA,CAAS;GAChE,SAAS,KAAK;GACd,QAAQ,QAAQ;GAChB;EACF,CAAC;CACH;AACF"}

package/dist/cjs/llm/openai/index.cjs

@@ -2,27 +2,19 @@ const require_cache = require("../../messages/cache.cjs");
 const require_core = require("../../messages/core.cjs");
 const require_streamedToolCallSeals = require("../../tools/streamedToolCallSeals.cjs");
 const require_intentArg = require("../../tools/intentArg.cjs");
+const require_smoother = require("../stream/smoother.cjs");
+const require_chunkAdapters = require("../stream/chunkAdapters.cjs");
 const require_index = require("./utils/index.cjs");
 const require_streamMetadata = require("./streamMetadata.cjs");
 let _langchain_openai = require("@langchain/openai");
 let _langchain_core_messages = require("@langchain/core/messages");
+let _langchain_core_outputs = require("@langchain/core/outputs");
 let openai = require("openai");
 let _langchain_xai = require("@langchain/xai");
-let _langchain_core_outputs = require("@langchain/core/outputs");
 let _langchain_deepseek = require("@langchain/deepseek");
 let _langchain_core_utils_function_calling = require("@langchain/core/utils/function_calling");
 //#region src/llm/openai/index.ts
 const iife = (fn) => fn();
-const STREAM_CHUNK_MIN_SIZE = 4;
-const STREAM_BOUNDARIES = new Set([
-	" ",
-	".",
-	",",
-	"!",
-	"?",
-	";",
-	":"
-]);
 function isHeaders(headers) {
 	return typeof Headers !== "undefined" && headers !== null && typeof headers === "object" && Object.prototype.toString.call(headers) === "[object Headers]";
 }
@@ -504,38 +496,60 @@ function getCustomOpenAIClientOptions(owner, options) {
 		...options
 	};
 }
-function findStreamChunkBoundary(text, minSize) {
-	if (minSize >= text.length) return text.length;
-	for (let position = minSize; position < text.length; position++) if (STREAM_BOUNDARIES.has(text[position])) return position + 1;
-	return text.length;
-}
-function splitStreamToken(text) {
-	const chunks = [];
-	let currentIndex = 0;
-	while (currentIndex < text.length) {
-		const chunkSize = findStreamChunkBoundary(text.slice(currentIndex), STREAM_CHUNK_MIN_SIZE);
-		chunks.push(text.slice(currentIndex, currentIndex + chunkSize));
-		currentIndex += chunkSize;
-	}
-	return chunks;
-}
-function splitTextGenerationChunk(chunk) {
+/**
+* Classifies a generation chunk for the smoothing engine:
+* - splittable: plain visible text (string content equal to `chunk.text`, no
+*   logprobs / finish_reason) — sliced adaptively at the pacing cadence.
+*   ANY logprobs value blocks splitting here (this family only attaches
+*   logprobs on request; the DeepSeek suite pins chunks with them staying
+*   intact) — deliberately stricter than `stream/chunkAdapters.ts`, where
+*   google-common's always-present empty logprobs must not block.
+* - atomic: text- or reasoning-bearing chunks whose metadata cannot survive
+*   slicing — paced as one piece, never split (legacy parity: these were
+*   emitted whole but still paced).
+* - passthrough: tool-call deltas, usage-only, finish_reason and other
+*   metadata chunks — strict FIFO, zero delay.
+*/
+function toSmoothItem(chunk) {
 	const { message } = chunk;
-	if (!chunk.text || !(message instanceof _langchain_core_messages.AIMessageChunk) || typeof message.content !== "string" || message.content !== chunk.text || chunk.generationInfo?.logprobs != null || chunk.generationInfo?.finish_reason != null) return [chunk];
-	const tokenChunks = splitStreamToken(chunk.text);
-	if (tokenChunks.length <= 1) return [chunk];
-	let emittedUsage = false;
-	return tokenChunks.map((token) => {
-		const usageMetadata = emittedUsage && message.usage_metadata != null ? void 0 : message.usage_metadata;
-		if (message.usage_metadata != null && !emittedUsage) emittedUsage = true;
-		return new _langchain_core_outputs.ChatGenerationChunk({
-			text: token,
-			generationInfo: chunk.generationInfo,
-			message: new _langchain_core_messages.AIMessageChunk(Object.assign({}, message, {
-				content: token,
-				usage_metadata: usageMetadata
-			}))
-		});
+	const isMessageChunk = message instanceof _langchain_core_messages.AIMessageChunk;
+	if (Boolean(chunk.text) && isMessageChunk && typeof message.content === "string" && message.content === chunk.text && chunk.generationInfo?.logprobs == null && chunk.generationInfo?.finish_reason == null && !require_chunkAdapters.hasReasoningKwargs(message) && !require_chunkAdapters.hasToolCallChunks(message)) return {
+		text: chunk.text,
+		smooth: true,
+		emit: (piece) => cloneGenerationChunkPiece(chunk, piece)
+	};
+	const pacedText = chunk.text || (isMessageChunk ? require_chunkAdapters.getReasoningKwargsText(message) : "");
+	if (pacedText !== "") return {
+		text: pacedText,
+		smooth: true,
+		atomic: true,
+		emit: () => chunk
+	};
+	return {
+		text: "",
+		smooth: false,
+		emit: () => chunk
+	};
+}
+/**
+* Usage metadata, additional kwargs and response metadata survive only on
+* the first piece: the aggregator's dict merge concatenates string fields
+* and sums usage, so replication across pieces corrupts them. Unlike the
+* generic adapter, `generationInfo` stays on every piece — per-piece token
+* indices ride in it and `dropRepeatedScalarMetadata` owns repetition there.
+*/
+function cloneGenerationChunkPiece(chunk, piece) {
+	if (piece.isFirst && piece.isLast) return chunk;
+	const message = chunk.message;
+	return new _langchain_core_outputs.ChatGenerationChunk({
+		text: piece.text,
+		generationInfo: chunk.generationInfo,
+		message: new _langchain_core_messages.AIMessageChunk(Object.assign({}, message, {
+			content: piece.text,
+			usage_metadata: piece.isFirst ? message.usage_metadata : void 0,
+			additional_kwargs: piece.isFirst ? message.additional_kwargs : {},
+			response_metadata: piece.isFirst ? message.response_metadata : {}
+		}))
 	});
 }
 async function emitStreamChunkCallback(chunk, runManager) {
@@ -549,42 +563,30 @@ function getStreamChunkTokenIndices(chunk) {
 		completion
 	};
 }
+/**
+* Adaptive smoothing adapter for the OpenAI chat-model family, layered over
+* the shared `smoothStream` engine. Keeps the historical signature so every
+* `_streamResponseChunks` call site is unchanged.
+*
+* `seenScalarMetadata`: when provided, de-duplicates repeated scalar metadata
+* just before emitting, so token callbacks and the yielded chunk observe the
+* same cleaned data. Omitted by callers that wrap this stream and finalize
+* downstream (e.g. `ChatOpenRouter`, which needs the raw `finish_reason` as
+* its flush signal and de-duplicates after its own processing).
+*/
 async function* delayStreamChunks(chunks, delay, signal, runManager, seenScalarMetadata) {
-	let lastYieldedAt;
-	for await (const chunk of chunks) {
-		const outputChunks = delay != null && delay > 0 ? splitTextGenerationChunk(chunk) : [chunk];
-		for (const outputChunk of outputChunks) {
-			signal?.throwIfAborted();
-			if (delay != null && delay > 0 && lastYieldedAt != null) {
-				const timeSinceLastYield = Date.now() - lastYieldedAt;
-				const timeToWait = Math.max(0, delay - timeSinceLastYield);
-				if (timeToWait > 0) await sleepWithAbort(timeToWait, signal);
-			}
-			signal?.throwIfAborted();
-			lastYieldedAt = Date.now();
-			if (seenScalarMetadata != null) require_streamMetadata.dropRepeatedScalarMetadata(outputChunk, seenScalarMetadata);
-			await emitStreamChunkCallback(outputChunk, runManager);
-			signal?.throwIfAborted();
-			yield outputChunk;
-		}
-	}
-}
-async function sleepWithAbort(delay, signal) {
-	if (delay <= 0) return;
-	signal?.throwIfAborted();
-	await new Promise((resolve, reject) => {
-		const timeout = setTimeout(() => {
-			signal?.removeEventListener("abort", onAbort);
-			resolve();
-		}, delay);
-		const onAbort = () => {
-			clearTimeout(timeout);
-			signal?.removeEventListener("abort", onAbort);
-			reject(signal?.reason ?? /* @__PURE__ */ new Error("AbortError: User aborted request."));
-		};
-		signal?.addEventListener("abort", onAbort, { once: true });
-		if (signal?.aborted === true) onAbort();
+	const smoothed = require_smoother.smoothStream({
+		source: (async function* () {
+			for await (const chunk of chunks) yield toSmoothItem(chunk);
+		})(),
+		delayMs: delay != null && delay > 0 ? delay : 0,
+		signal
 	});
+	for await (const outputChunk of smoothed) {
+		if (seenScalarMetadata != null) require_streamMetadata.dropRepeatedScalarMetadata(outputChunk, seenScalarMetadata);
+		await emitStreamChunkCallback(outputChunk, runManager);
+		yield outputChunk;
+	}
 }
 function createAbortHandler(controller) {
 	return function() {
@@ -1184,7 +1186,7 @@ var ChatOpenAI = class extends _langchain_openai.ChatOpenAI {
 	_lc_stream_delay;
 	constructor(fields) {
 		super(withLibreChatOpenAIFields(fields));
-		this._lc_stream_delay = fields?._lc_stream_delay;
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
 	}
 	get exposedClient() {
 		return getExposedOpenAIClient(this.completions, this.responses, this._useResponsesApi(void 0));
@@ -1237,7 +1239,7 @@ var AzureChatOpenAI = class extends _langchain_openai.AzureChatOpenAI {
 		super(fields);
 		this.completions = new LibreChatAzureOpenAICompletions(fields);
 		this.responses = new LibreChatAzureOpenAIResponses(fields);
-		this._lc_stream_delay = fields?._lc_stream_delay;
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
 	}
 	get exposedClient() {
 		return getExposedOpenAIClient(this.completions, this.responses, this._useResponsesApi(void 0));
@@ -1309,7 +1311,7 @@ var ChatDeepSeek = class extends _langchain_deepseek.ChatDeepSeek {
 	_lc_stream_delay;
 	constructor(fields) {
 		super(fields);
-		this._lc_stream_delay = fields?._lc_stream_delay;
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
 	}
 	get exposedClient() {
 		return this.client;
@@ -1572,7 +1574,7 @@ var ChatXAI = class extends _langchain_xai.ChatXAI {
 	_lc_stream_delay;
 	constructor(fields) {
 		super(fields);
-		this._lc_stream_delay = fields?._lc_stream_delay;
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
 		const customBaseURL = fields?.configuration?.baseURL ?? fields?.clientConfig?.baseURL;
 		if (customBaseURL != null && customBaseURL) {
 			this.clientConfig = {