@livekit/agents 1.0.32 → 1.0.34

package/dist/inference/llm.cjs

@@ -251,8 +251,6 @@ class LLMStream extends llm.LLMStream {
           options: { retryable }
         });
       }
-    } finally {
-      this.queue.close();
     }
   }
   parseChoice(id, choice) {

package/dist/inference/llm.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels = 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    } finally {\n      this.queue.close();\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      return callChunk;\n    }\n\n    // Regular content message\n    if (!delta.content) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n          }),\n        ],\n      },\n    };\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAAmB;AACnB,eAOO;AACP,UAAqB;AAErB,mBAAkD;AAElD,MAAM,mBAAmB;AAqFlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,cAAAA,QAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEC,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AAlRvC;AAqRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AAEzE,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,UAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,cAAAD,QAAO,2BAA2B;AACrD,cAAM,IAAI,yBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,cAAAA,QAAO,UAAU;AAC3C,cAAM,IAAI,wBAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,4BAAmB;AAAA,UAC3B,aAAS,kBAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF,UAAE;AACA,WAAK,MAAM,MAAM;AAAA,IACnB;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AAAA,QAC1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAAA,QACpD,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,aAAO;AAAA,IACT;AAGA,QAAI,CAAC,MAAM,SAAS;AAClB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,UAChC,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;","names":["OpenAI","llm"]}
+{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels = 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      return callChunk;\n    }\n\n    // Regular content message\n    if (!delta.content) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n          }),\n        ],\n      },\n    };\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAAmB;AACnB,eAOO;AACP,UAAqB;AAErB,mBAAkD;AAElD,MAAM,mBAAmB;AAqFlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,cAAAA,QAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEC,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AAlRvC;AAqRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AAEzE,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,UAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,cAAAD,QAAO,2BAA2B;AACrD,cAAM,IAAI,yBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,cAAAA,QAAO,UAAU;AAC3C,cAAM,IAAI,wBAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,4BAAmB;AAAA,UAC3B,aAAS,kBAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AAAA,QAC1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAAA,QACpD,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,aAAO;AAAA,IACT;AAGA,QAAI,CAAC,MAAM,SAAS;AAClB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,UAChC,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;","names":["OpenAI","llm"]}

package/dist/inference/llm.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../src/inference/llm.ts"],"names":[],"mappings":"AAGA,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,EAKL,KAAK,MAAM,EAEZ,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,GAAG,MAAM,iBAAiB,CAAC;AACvC,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,EAAE,KAAK,SAAS,EAAqB,MAAM,YAAY,CAAC;AAI/D,MAAM,MAAM,YAAY,GACpB,cAAc,GACd,mBAAmB,GACnB,mBAAmB,GACnB,gBAAgB,GAChB,qBAAqB,GACrB,qBAAqB,GACrB,eAAe,GACf,oBAAoB,GACpB,qBAAqB,CAAC;AAE1B,MAAM,MAAM,YAAY,GAAG,8BAA8B,CAAC;AAE1D,MAAM,MAAM,UAAU,GAAG,+BAA+B,CAAC;AAEzD,MAAM,MAAM,UAAU,GAAG,6BAA6B,CAAC;AAEvD,MAAM,MAAM,cAAc,GAAG,yBAAyB,CAAC;AAEvD,KAAK,oCAAoC,GACvC,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,+BAA+B,CAAC,CAAC;AAClE,KAAK,gBAAgB,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0BAA0B,CAAC,gBAAgB,CAAC,CAAC;AACpG,KAAK,UAAU,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0BAA0B,CAAC,aAAa,CAAC,CAAC,CAAC;AAC5F,KAAK,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAE3C,MAAM,WAAW,qBAAsB,SAAQ,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC;IACpE,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,UAAU,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACpC,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAClC,UAAU,CAAC,EAAE,KAAK,CAAC,MAAM,GAAG,OAAO,CAAC,CAAC;IACrC,CAAC,CAAC,EAAE,MAAM,CAAC;IACX,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,UAAU,CAAC,EAAE,oCAAoC,GAAG,IAAI,CAAC;IACzD,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,gBAAgB,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IACzD,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,YAAY,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,MAAM,GAAG,OAAO,GAAG,UAAU,CAAC;IAClE,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IACzB,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB,kBAAkB,CAAC,EAAE,gBAAgB,CAAC;IAGtC,WAAW,CAAC,EAAE,UAAU,CAAC;CAG1B;AAED,MAAM,MAAM,SAAS,GACjB,YAAY,GACZ,YAAY,GACZ,UAAU,GACV,UAAU,GACV,cAAc,GACd,SAAS,CAAC;AAEd,MAAM,WAAW,mBAAmB;IAClC,KAAK,EAAE,SAAS,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,qBAAqB,CAAC;IACpC,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,qBAAa,GAAI,SAAQ,GAAG,CAAC,GAAG;IAC9B,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,IAAI,CAAsB;gBAEtB,IAAI,EAAE;QAChB,KAAK,EAAE,SAAS,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,YAAY,CAAC,EAAE,mBAAmB,CAAC,cAAc,CAAC,CAAC;QACnD,gBAAgB,CAAC,EAAE,OAAO,CAAC;KAC5B;IAyCD,KAAK,IAAI,MAAM;IAIf,IAAI,KAAK,IAAI,MAAM,CAElB;IAED,MAAM,CAAC,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,GAAG;IAIhD,IAAI,CAAC,EACH,OAAO,EACP,OAAO,EACP,WAAyC,EACzC,iBAAiB,EACjB,UAAU,EAEV,WAAW,GACZ,EAAE;QACD,OAAO,EAAE,GAAG,CAAC,WAAW,CAAC;QACzB,OAAO,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC;QAC1B,WAAW,CAAC,EAAE,iBAAiB,CAAC;QAChC,iBAAiB,CAAC,EAAE,OAAO,CAAC;QAC5B,UAAU,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC;QAE5B,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;KACvC,GAAG,SAAS;CAwCd;AAED,qBAAa,SAAU,SAAQ,GAAG,CAAC,SAAS;IAC1C,OAAO,CAAC,KAAK,CAAY;IACzB,OAAO,CAAC,QAAQ,CAAC,CAAS;IAC1B,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,YAAY,CAA0B;IAC9C,OAAO,CAAC,gBAAgB,CAAU;IAElC,OAAO,CAAC,cAAc,CAAC,CAAiB;IACxC,OAAO,CAAC,UAAU,CAAC,CAAS;IAC5B,OAAO,CAAC,SAAS,CAAC,CAAS;IAC3B,OAAO,CAAC,OAAO,CAAC,CAAS;IACzB,OAAO,CAAC,eAAe,CAAC,CAAS;gBAG/B,GAAG,EAAE,GAAG,EACR,EACE,KAAK,EACL,QAAQ,EACR,MAAM,EACN,OAAO,EACP,OAAO,EACP,cAAc,EACd,WAAW,EACX,YAAY,EACZ,WAAW,EACX,gBAAgB,GACjB,EAAE;QACD,KAAK,EAAE,SAAS,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,MAAM,EAAE,MAAM,CAAC;QACf,OAAO,EAAE,GAAG,CAAC,WAAW,CAAC;QACzB,OAAO,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC;QAC1B,cAAc,CAAC,EAAE,cAAc,CAAC;QAChC,WAAW,EAAE,iBAAiB,CAAC;QAC/B,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;QACtC,WAAW,CAAC,EAAE,GAAG,CAAC,cAAc,CAAC;QACjC,gBAAgB,EAAE,OAAO,CAAC;KAC3B;cAYa,GAAG,IAAI,OAAO,CAAC,IAAI,CAAC;IAuHpC,OAAO,CAAC,WAAW;IAkFnB,OAAO,CAAC,0BAA0B;CAmBnC"}
+{"version":3,"file":"llm.d.ts","sourceRoot":"","sources":["../../src/inference/llm.ts"],"names":[],"mappings":"AAGA,OAAO,MAAM,MAAM,QAAQ,CAAC;AAC5B,OAAO,EAKL,KAAK,MAAM,EAEZ,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,GAAG,MAAM,iBAAiB,CAAC;AACvC,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,EAAE,KAAK,SAAS,EAAqB,MAAM,YAAY,CAAC;AAI/D,MAAM,MAAM,YAAY,GACpB,cAAc,GACd,mBAAmB,GACnB,mBAAmB,GACnB,gBAAgB,GAChB,qBAAqB,GACrB,qBAAqB,GACrB,eAAe,GACf,oBAAoB,GACpB,qBAAqB,CAAC;AAE1B,MAAM,MAAM,YAAY,GAAG,8BAA8B,CAAC;AAE1D,MAAM,MAAM,UAAU,GAAG,+BAA+B,CAAC;AAEzD,MAAM,MAAM,UAAU,GAAG,6BAA6B,CAAC;AAEvD,MAAM,MAAM,cAAc,GAAG,yBAAyB,CAAC;AAEvD,KAAK,oCAAoC,GACvC,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,+BAA+B,CAAC,CAAC;AAClE,KAAK,gBAAgB,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0BAA0B,CAAC,gBAAgB,CAAC,CAAC;AACpG,KAAK,UAAU,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,0BAA0B,CAAC,aAAa,CAAC,CAAC,CAAC;AAC5F,KAAK,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAE3C,MAAM,WAAW,qBAAsB,SAAQ,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC;IACpE,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,UAAU,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACpC,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAClC,UAAU,CAAC,EAAE,KAAK,CAAC,MAAM,GAAG,OAAO,CAAC,CAAC;IACrC,CAAC,CAAC,EAAE,MAAM,CAAC;IACX,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,UAAU,CAAC,EAAE,oCAAoC,GAAG,IAAI,CAAC;IACzD,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,gBAAgB,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IACzD,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,YAAY,CAAC,EAAE,MAAM,GAAG,SAAS,GAAG,MAAM,GAAG,OAAO,GAAG,UAAU,CAAC;IAClE,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM,EAAE,CAAC;IACzB,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB,kBAAkB,CAAC,EAAE,gBAAgB,CAAC;IAGtC,WAAW,CAAC,EAAE,UAAU,CAAC;CAG1B;AAED,MAAM,MAAM,SAAS,GACjB,YAAY,GACZ,YAAY,GACZ,UAAU,GACV,UAAU,GACV,cAAc,GACd,SAAS,CAAC;AAEd,MAAM,WAAW,mBAAmB;IAClC,KAAK,EAAE,SAAS,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,qBAAqB,CAAC;IACpC,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,qBAAa,GAAI,SAAQ,GAAG,CAAC,GAAG;IAC9B,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,IAAI,CAAsB;gBAEtB,IAAI,EAAE;QAChB,KAAK,EAAE,SAAS,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,YAAY,CAAC,EAAE,mBAAmB,CAAC,cAAc,CAAC,CAAC;QACnD,gBAAgB,CAAC,EAAE,OAAO,CAAC;KAC5B;IAyCD,KAAK,IAAI,MAAM;IAIf,IAAI,KAAK,IAAI,MAAM,CAElB;IAED,MAAM,CAAC,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,GAAG;IAIhD,IAAI,CAAC,EACH,OAAO,EACP,OAAO,EACP,WAAyC,EACzC,iBAAiB,EACjB,UAAU,EAEV,WAAW,GACZ,EAAE;QACD,OAAO,EAAE,GAAG,CAAC,WAAW,CAAC;QACzB,OAAO,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC;QAC1B,WAAW,CAAC,EAAE,iBAAiB,CAAC;QAChC,iBAAiB,CAAC,EAAE,OAAO,CAAC;QAC5B,UAAU,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC;QAE5B,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;KACvC,GAAG,SAAS;CAwCd;AAED,qBAAa,SAAU,SAAQ,GAAG,CAAC,SAAS;IAC1C,OAAO,CAAC,KAAK,CAAY;IACzB,OAAO,CAAC,QAAQ,CAAC,CAAS;IAC1B,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,YAAY,CAA0B;IAC9C,OAAO,CAAC,gBAAgB,CAAU;IAElC,OAAO,CAAC,cAAc,CAAC,CAAiB;IACxC,OAAO,CAAC,UAAU,CAAC,CAAS;IAC5B,OAAO,CAAC,SAAS,CAAC,CAAS;IAC3B,OAAO,CAAC,OAAO,CAAC,CAAS;IACzB,OAAO,CAAC,eAAe,CAAC,CAAS;gBAG/B,GAAG,EAAE,GAAG,EACR,EACE,KAAK,EACL,QAAQ,EACR,MAAM,EACN,OAAO,EACP,OAAO,EACP,cAAc,EACd,WAAW,EACX,YAAY,EACZ,WAAW,EACX,gBAAgB,GACjB,EAAE;QACD,KAAK,EAAE,SAAS,CAAC;QACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,MAAM,EAAE,MAAM,CAAC;QACf,OAAO,EAAE,GAAG,CAAC,WAAW,CAAC;QACzB,OAAO,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC;QAC1B,cAAc,CAAC,EAAE,cAAc,CAAC;QAChC,WAAW,EAAE,iBAAiB,CAAC;QAC/B,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;QACtC,WAAW,CAAC,EAAE,GAAG,CAAC,cAAc,CAAC;QACjC,gBAAgB,EAAE,OAAO,CAAC;KAC3B;cAYa,GAAG,IAAI,OAAO,CAAC,IAAI,CAAC;IAqHpC,OAAO,CAAC,WAAW;IAkFnB,OAAO,CAAC,0BAA0B;CAmBnC"}

package/dist/inference/llm.js

@@ -223,8 +223,6 @@ class LLMStream extends llm.LLMStream {
           options: { retryable }
         });
       }
-    } finally {
-      this.queue.close();
     }
   }
   parseChoice(id, choice) {

package/dist/inference/llm.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels = 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    } finally {\n      this.queue.close();\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      return callChunk;\n    }\n\n    // Regular content message\n    if (!delta.content) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n          }),\n        ],\n      },\n    };\n  }\n}\n"],"mappings":"AAGA,OAAO,YAAY;AACnB;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,OACK;AACP,YAAY,SAAS;AAErB,SAAyB,yBAAyB;AAElD,MAAM,mBAAmB;AAqFlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,OAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AAlRvC;AAqRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AAEzE,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,MAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,OAAO,2BAA2B;AACrD,cAAM,IAAI,gBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,OAAO,UAAU;AAC3C,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,mBAAmB;AAAA,UAC3B,SAAS,QAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF,UAAE;AACA,WAAK,MAAM,MAAM;AAAA,IACnB;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AAAA,QAC1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAAA,QACpD,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,aAAO;AAAA,IACT;AAGA,QAAI,CAAC,MAAM,SAAS;AAClB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,UAChC,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;","names":["llm"]}
+{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels = 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      return callChunk;\n    }\n\n    // Regular content message\n    if (!delta.content) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n          }),\n        ],\n      },\n    };\n  }\n}\n"],"mappings":"AAGA,OAAO,YAAY;AACnB;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,OACK;AACP,YAAY,SAAS;AAErB,SAAyB,yBAAyB;AAElD,MAAM,mBAAmB;AAqFlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,OAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEA,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AAlRvC;AAqRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AAEzE,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,MAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,OAAO,2BAA2B;AACrD,cAAM,IAAI,gBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,OAAO,UAAU;AAC3C,cAAM,IAAI,eAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,mBAAmB;AAAA,UAC3B,SAAS,QAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AAAA,QAC1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAAA,QACpD,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,aAAO;AAAA,IACT;AAGA,QAAI,CAAC,MAAM,SAAS;AAClB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,UAChC,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;","names":["llm"]}

package/dist/llm/fallback_adapter.cjs

@@ -0,0 +1,278 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+var fallback_adapter_exports = {};
+__export(fallback_adapter_exports, {
+  FallbackAdapter: () => FallbackAdapter
+});
+module.exports = __toCommonJS(fallback_adapter_exports);
+var import_exceptions = require("../_exceptions.cjs");
+var import_log = require("../log.cjs");
+var import_types = require("../types.cjs");
+var import_llm = require("./llm.cjs");
+const DEFAULT_FALLBACK_API_CONNECT_OPTIONS = {
+  maxRetry: 0,
+  timeoutMs: import_types.DEFAULT_API_CONNECT_OPTIONS.timeoutMs,
+  retryIntervalMs: import_types.DEFAULT_API_CONNECT_OPTIONS.retryIntervalMs
+};
+class FallbackAdapter extends import_llm.LLM {
+  llms;
+  attemptTimeout;
+  maxRetryPerLLM;
+  retryInterval;
+  retryOnChunkSent;
+  /** @internal */
+  _status;
+  logger = (0, import_log.log)();
+  constructor(options) {
+    super();
+    if (!options.llms || options.llms.length < 1) {
+      throw new Error("at least one LLM instance must be provided.");
+    }
+    this.llms = options.llms;
+    this.attemptTimeout = options.attemptTimeout ?? 5;
+    this.maxRetryPerLLM = options.maxRetryPerLLM ?? 0;
+    this.retryInterval = options.retryInterval ?? 0.5;
+    this.retryOnChunkSent = options.retryOnChunkSent ?? false;
+    this._status = this.llms.map(() => ({
+      available: true,
+      recoveringTask: null
+    }));
+    for (const llm of this.llms) {
+      llm.on("metrics_collected", (metrics) => {
+        this.emit("metrics_collected", metrics);
+      });
+    }
+  }
+  get model() {
+    return "FallbackAdapter";
+  }
+  label() {
+    return "FallbackAdapter";
+  }
+  chat(opts) {
+    return new FallbackLLMStream(this, {
+      chatCtx: opts.chatCtx,
+      toolCtx: opts.toolCtx,
+      connOptions: opts.connOptions || DEFAULT_FALLBACK_API_CONNECT_OPTIONS,
+      parallelToolCalls: opts.parallelToolCalls,
+      toolChoice: opts.toolChoice,
+      extraKwargs: opts.extraKwargs
+    });
+  }
+  /**
+   * Emit availability changed event.
+   * @internal
+   */
+  _emitAvailabilityChanged(llm, available) {
+    const event = { llm, available };
+    this.emit(
+      "llm_availability_changed",
+      event
+    );
+  }
+}
+class FallbackLLMStream extends import_llm.LLMStream {
+  adapter;
+  parallelToolCalls;
+  toolChoice;
+  extraKwargs;
+  _currentStream;
+  _log = (0, import_log.log)();
+  constructor(adapter, opts) {
+    super(adapter, {
+      chatCtx: opts.chatCtx,
+      toolCtx: opts.toolCtx,
+      connOptions: opts.connOptions
+    });
+    this.adapter = adapter;
+    this.parallelToolCalls = opts.parallelToolCalls;
+    this.toolChoice = opts.toolChoice;
+    this.extraKwargs = opts.extraKwargs;
+  }
+  /**
+   * Override chatCtx to return current stream's context if available.
+   */
+  get chatCtx() {
+    var _a;
+    return ((_a = this._currentStream) == null ? void 0 : _a.chatCtx) ?? super.chatCtx;
+  }
+  /**
+   * Try to generate with a single LLM.
+   * Returns an async generator that yields chunks.
+   */
+  async *tryGenerate(llm, checkRecovery = false) {
+    const connOptions = {
+      ...this.connOptions,
+      maxRetry: this.adapter.maxRetryPerLLM,
+      timeoutMs: this.adapter.attemptTimeout * 1e3,
+      retryIntervalMs: this.adapter.retryInterval * 1e3
+    };
+    const stream = llm.chat({
+      chatCtx: super.chatCtx,
+      toolCtx: this.toolCtx,
+      connOptions,
+      parallelToolCalls: this.parallelToolCalls,
+      toolChoice: this.toolChoice,
+      extraKwargs: this.extraKwargs
+    });
+    let streamError;
+    const errorHandler = (ev) => {
+      streamError = ev.error;
+    };
+    llm.on("error", errorHandler);
+    try {
+      let shouldSetCurrent = !checkRecovery;
+      for await (const chunk of stream) {
+        if (shouldSetCurrent) {
+          shouldSetCurrent = false;
+          this._currentStream = stream;
+        }
+        yield chunk;
+      }
+      if (streamError) {
+        throw streamError;
+      }
+    } catch (error) {
+      if (error instanceof import_exceptions.APIError) {
+        if (checkRecovery) {
+          this._log.warn({ llm: llm.label(), error }, "recovery failed");
+        } else {
+          this._log.warn({ llm: llm.label(), error }, "failed, switching to next LLM");
+        }
+        throw error;
+      }
+      if (error instanceof Error && error.name === "AbortError") {
+        if (checkRecovery) {
+          this._log.warn({ llm: llm.label() }, "recovery timed out");
+        } else {
+          this._log.warn({ llm: llm.label() }, "timed out, switching to next LLM");
+        }
+        throw error;
+      }
+      if (checkRecovery) {
+        this._log.error({ llm: llm.label(), error }, "recovery unexpected error");
+      } else {
+        this._log.error({ llm: llm.label(), error }, "unexpected error, switching to next LLM");
+      }
+      throw error;
+    } finally {
+      llm.off("error", errorHandler);
+    }
+  }
+  /**
+   * Start background recovery task for an LLM.
+   */
+  tryRecovery(llm, index) {
+    const status = this.adapter._status[index];
+    if (status.recoveringTask !== null) {
+      return;
+    }
+    const recoverTask = async () => {
+      try {
+        for await (const _chunk of this.tryGenerate(llm, true)) {
+        }
+        status.available = true;
+        this._log.info({ llm: llm.label() }, "LLM recovered");
+        this.adapter._emitAvailabilityChanged(llm, true);
+      } catch {
+      } finally {
+        status.recoveringTask = null;
+      }
+    };
+    status.recoveringTask = recoverTask();
+  }
+  /**
+   * Main run method - iterates through LLMs with fallback logic.
+   */
+  async run() {
+    const startTime = Date.now();
+    const allFailed = this.adapter._status.every((s) => !s.available);
+    if (allFailed) {
+      this._log.error("all LLMs are unavailable, retrying...");
+    }
+    for (let i = 0; i < this.adapter.llms.length; i++) {
+      const llm = this.adapter.llms[i];
+      const status = this.adapter._status[i];
+      this._log.debug(
+        { llm: llm.label(), index: i, available: status.available, allFailed },
+        "checking LLM"
+      );
+      if (status.available || allFailed) {
+        let textSent = "";
+        const toolCallsSent = [];
+        try {
+          this._log.info({ llm: llm.label() }, "FallbackAdapter: Attempting provider");
+          let chunkCount = 0;
+          for await (const chunk of this.tryGenerate(llm, false)) {
+            chunkCount++;
+            if (chunk.delta) {
+              if (chunk.delta.content) {
+                textSent += chunk.delta.content;
+              }
+              if (chunk.delta.toolCalls) {
+                for (const tc of chunk.delta.toolCalls) {
+                  if (tc.name) {
+                    toolCallsSent.push(tc.name);
+                  }
+                }
+              }
+            }
+            this._log.debug({ llm: llm.label(), chunkCount }, "run: forwarding chunk to queue");
+            this.queue.put(chunk);
+          }
+          this._log.info(
+            { llm: llm.label(), totalChunks: chunkCount, textLength: textSent.length },
+            "FallbackAdapter: Provider succeeded"
+          );
+          return;
+        } catch (error) {
+          if (status.available) {
+            status.available = false;
+            this.adapter._emitAvailabilityChanged(llm, false);
+          }
+          if (textSent || toolCallsSent.length > 0) {
+            const extra = { textSent, toolCallsSent };
+            if (!this.adapter.retryOnChunkSent) {
+              this._log.error(
+                { llm: llm.label(), ...extra },
+                "failed after sending chunk, skip retrying. Set `retryOnChunkSent` to `true` to enable."
+              );
+              throw error;
+            }
+            this._log.warn(
+              { llm: llm.label(), ...extra },
+              "failed after sending chunk, retrying..."
+            );
+          }
+        }
+      }
+      this.tryRecovery(llm, i);
+    }
+    const duration = (Date.now() - startTime) / 1e3;
+    const labels = this.adapter.llms.map((l) => l.label()).join(", ");
+    throw new import_exceptions.APIConnectionError({
+      message: `all LLMs failed (${labels}) after ${duration.toFixed(2)}s`
+    });
+  }
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  FallbackAdapter
+});
+//# sourceMappingURL=fallback_adapter.cjs.map

package/dist/llm/fallback_adapter.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/llm/fallback_adapter.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport { APIConnectionError, APIError } from '../_exceptions.js';\nimport { log } from '../log.js';\nimport { type APIConnectOptions, DEFAULT_API_CONNECT_OPTIONS } from '../types.js';\nimport type { ChatContext } from './chat_context.js';\nimport type { ChatChunk } from './llm.js';\nimport { LLM, LLMStream } from './llm.js';\nimport type { ToolChoice, ToolContext } from './tool_context.js';\n\n/**\n * Default connection options for FallbackAdapter.\n * Uses max_retry=0 since fallback handles retries at a higher level.\n */\nconst DEFAULT_FALLBACK_API_CONNECT_OPTIONS: APIConnectOptions = {\n  maxRetry: 0,\n  timeoutMs: DEFAULT_API_CONNECT_OPTIONS.timeoutMs,\n  retryIntervalMs: DEFAULT_API_CONNECT_OPTIONS.retryIntervalMs,\n};\n\n/**\n * Internal status tracking for each LLM instance.\n */\ninterface LLMStatus {\n  available: boolean;\n  recoveringTask: Promise<void> | null;\n}\n\n/**\n * Event emitted when an LLM's availability changes.\n */\nexport interface AvailabilityChangedEvent {\n  llm: LLM;\n  available: boolean;\n}\n\n/**\n * Options for creating a FallbackAdapter.\n */\nexport interface FallbackAdapterOptions {\n  /** List of LLM instances to fallback to (in order). */\n  llms: LLM[];\n  /** Timeout for each LLM attempt in seconds. Defaults to 5.0. */\n  attemptTimeout?: number;\n  /** Internal retries per LLM before moving to next. Defaults to 0. */\n  maxRetryPerLLM?: number;\n  /** Interval between retries in seconds. Defaults to 0.5. */\n  retryInterval?: number;\n  /** Whether to retry when LLM fails after chunks are sent. Defaults to false. */\n  retryOnChunkSent?: boolean;\n}\n\n/**\n * FallbackAdapter is an LLM that can fallback to a different LLM if the current LLM fails.\n *\n * @example\n * ```typescript\n * const fallbackLLM = new FallbackAdapter({\n *   llms: [primaryLLM, secondaryLLM, tertiaryLLM],\n *   attemptTimeout: 5.0,\n *   maxRetryPerLLM: 1,\n * });\n * ```\n */\nexport class FallbackAdapter extends LLM {\n  readonly llms: LLM[];\n  readonly attemptTimeout: number;\n  readonly maxRetryPerLLM: number;\n  readonly retryInterval: number;\n  readonly retryOnChunkSent: boolean;\n\n  /** @internal */\n  _status: LLMStatus[];\n\n  private logger = log();\n\n  constructor(options: FallbackAdapterOptions) {\n    super();\n\n    if (!options.llms || options.llms.length < 1) {\n      throw new Error('at least one LLM instance must be provided.');\n    }\n\n    this.llms = options.llms;\n    this.attemptTimeout = options.attemptTimeout ?? 5.0;\n    this.maxRetryPerLLM = options.maxRetryPerLLM ?? 0;\n    this.retryInterval = options.retryInterval ?? 0.5;\n    this.retryOnChunkSent = options.retryOnChunkSent ?? false;\n\n    // Initialize status for each LLM\n    this._status = this.llms.map(() => ({\n      available: true,\n      recoveringTask: null,\n    }));\n\n    // Forward metrics_collected events from child LLMs\n    for (const llm of this.llms) {\n      llm.on('metrics_collected', (metrics) => {\n        this.emit('metrics_collected', metrics);\n      });\n    }\n  }\n\n  get model(): string {\n    return 'FallbackAdapter';\n  }\n\n  label(): string {\n    return 'FallbackAdapter';\n  }\n\n  chat(opts: {\n    chatCtx: ChatContext;\n    toolCtx?: ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: ToolChoice;\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    return new FallbackLLMStream(this, {\n      chatCtx: opts.chatCtx,\n      toolCtx: opts.toolCtx,\n      connOptions: opts.connOptions || DEFAULT_FALLBACK_API_CONNECT_OPTIONS,\n      parallelToolCalls: opts.parallelToolCalls,\n      toolChoice: opts.toolChoice,\n      extraKwargs: opts.extraKwargs,\n    });\n  }\n\n  /**\n   * Emit availability changed event.\n   * @internal\n   */\n  _emitAvailabilityChanged(llm: LLM, available: boolean): void {\n    const event: AvailabilityChangedEvent = { llm, available };\n    // Use type assertion for custom event\n    (this as unknown as { emit: (event: string, data: AvailabilityChangedEvent) => void }).emit(\n      'llm_availability_changed',\n      event,\n    );\n  }\n}\n\n/**\n * LLMStream implementation for FallbackAdapter.\n * Handles fallback logic between multiple LLM providers.\n */\nclass FallbackLLMStream extends LLMStream {\n  private adapter: FallbackAdapter;\n  private parallelToolCalls?: boolean;\n  private toolChoice?: ToolChoice;\n  private extraKwargs?: Record<string, unknown>;\n  private _currentStream?: LLMStream;\n  private _log = log();\n\n  constructor(\n    adapter: FallbackAdapter,\n    opts: {\n      chatCtx: ChatContext;\n      toolCtx?: ToolContext;\n      connOptions: APIConnectOptions;\n      parallelToolCalls?: boolean;\n      toolChoice?: ToolChoice;\n      extraKwargs?: Record<string, unknown>;\n    },\n  ) {\n    super(adapter, {\n      chatCtx: opts.chatCtx,\n      toolCtx: opts.toolCtx,\n      connOptions: opts.connOptions,\n    });\n    this.adapter = adapter;\n    this.parallelToolCalls = opts.parallelToolCalls;\n    this.toolChoice = opts.toolChoice;\n    this.extraKwargs = opts.extraKwargs;\n  }\n\n  /**\n   * Override chatCtx to return current stream's context if available.\n   */\n  override get chatCtx(): ChatContext {\n    return this._currentStream?.chatCtx ?? super.chatCtx;\n  }\n\n  /**\n   * Try to generate with a single LLM.\n   * Returns an async generator that yields chunks.\n   */\n  private async *tryGenerate(\n    llm: LLM,\n    checkRecovery: boolean = false,\n  ): AsyncGenerator<ChatChunk, void, unknown> {\n    const connOptions: APIConnectOptions = {\n      ...this.connOptions,\n      maxRetry: this.adapter.maxRetryPerLLM,\n      timeoutMs: this.adapter.attemptTimeout * 1000,\n      retryIntervalMs: this.adapter.retryInterval * 1000,\n    };\n\n    const stream = llm.chat({\n      chatCtx: super.chatCtx,\n      toolCtx: this.toolCtx,\n      connOptions,\n      parallelToolCalls: this.parallelToolCalls,\n      toolChoice: this.toolChoice,\n      extraKwargs: this.extraKwargs,\n    });\n\n    // Listen for error events - child LLMs emit errors via their LLM instance, not the stream\n    let streamError: Error | undefined;\n    const errorHandler = (ev: { error: Error }) => {\n      streamError = ev.error;\n    };\n    llm.on('error', errorHandler);\n\n    try {\n      let shouldSetCurrent = !checkRecovery;\n      for await (const chunk of stream) {\n        if (shouldSetCurrent) {\n          shouldSetCurrent = false;\n          this._currentStream = stream;\n        }\n        yield chunk;\n      }\n\n      // If an error was emitted but not thrown through iteration, throw it now\n      if (streamError) {\n        throw streamError;\n      }\n    } catch (error) {\n      if (error instanceof APIError) {\n        if (checkRecovery) {\n          this._log.warn({ llm: llm.label(), error }, 'recovery failed');\n        } else {\n          this._log.warn({ llm: llm.label(), error }, 'failed, switching to next LLM');\n        }\n        throw error;\n      }\n\n      // Handle timeout errors\n      if (error instanceof Error && error.name === 'AbortError') {\n        if (checkRecovery) {\n          this._log.warn({ llm: llm.label() }, 'recovery timed out');\n        } else {\n          this._log.warn({ llm: llm.label() }, 'timed out, switching to next LLM');\n        }\n        throw error;\n      }\n\n      // Unexpected error\n      if (checkRecovery) {\n        this._log.error({ llm: llm.label(), error }, 'recovery unexpected error');\n      } else {\n        this._log.error({ llm: llm.label(), error }, 'unexpected error, switching to next LLM');\n      }\n      throw error;\n    } finally {\n      llm.off('error', errorHandler);\n    }\n  }\n\n  /**\n   * Start background recovery task for an LLM.\n   */\n  private tryRecovery(llm: LLM, index: number): void {\n    const status = this.adapter._status[index]!;\n\n    // Skip if already recovering\n    if (status.recoveringTask !== null) {\n      return;\n    }\n\n    const recoverTask = async (): Promise<void> => {\n      try {\n        // Try to generate (just iterate to check if it works)\n        // eslint-disable-next-line @typescript-eslint/no-unused-vars\n        for await (const _chunk of this.tryGenerate(llm, true)) {\n          // Just consume the stream to verify it works\n        }\n\n        // Recovery successful\n        status.available = true;\n        this._log.info({ llm: llm.label() }, 'LLM recovered');\n        this.adapter._emitAvailabilityChanged(llm, true);\n      } catch {\n        // Recovery failed, stay unavailable\n      } finally {\n        status.recoveringTask = null;\n      }\n    };\n\n    // Fire and forget\n    status.recoveringTask = recoverTask();\n  }\n\n  /**\n   * Main run method - iterates through LLMs with fallback logic.\n   */\n  protected async run(): Promise<void> {\n    const startTime = Date.now();\n\n    // Check if all LLMs are unavailable\n    const allFailed = this.adapter._status.every((s) => !s.available);\n    if (allFailed) {\n      this._log.error('all LLMs are unavailable, retrying...');\n    }\n\n    for (let i = 0; i < this.adapter.llms.length; i++) {\n      const llm = this.adapter.llms[i]!;\n      const status = this.adapter._status[i]!;\n\n      this._log.debug(\n        { llm: llm.label(), index: i, available: status.available, allFailed },\n        'checking LLM',\n      );\n\n      if (status.available || allFailed) {\n        let textSent = '';\n        const toolCallsSent: string[] = [];\n\n        try {\n          this._log.info({ llm: llm.label() }, 'FallbackAdapter: Attempting provider');\n\n          let chunkCount = 0;\n          for await (const chunk of this.tryGenerate(llm, false)) {\n            chunkCount++;\n            // Track what's been sent\n            if (chunk.delta) {\n              if (chunk.delta.content) {\n                textSent += chunk.delta.content;\n              }\n              if (chunk.delta.toolCalls) {\n                for (const tc of chunk.delta.toolCalls) {\n                  if (tc.name) {\n                    toolCallsSent.push(tc.name);\n                  }\n                }\n              }\n            }\n\n            // Forward chunk to queue\n            this._log.debug({ llm: llm.label(), chunkCount }, 'run: forwarding chunk to queue');\n            this.queue.put(chunk);\n          }\n\n          // Success!\n          this._log.info(\n            { llm: llm.label(), totalChunks: chunkCount, textLength: textSent.length },\n            'FallbackAdapter: Provider succeeded',\n          );\n          return;\n        } catch (error) {\n          // Mark as unavailable if it was available before\n          if (status.available) {\n            status.available = false;\n            this.adapter._emitAvailabilityChanged(llm, false);\n          }\n\n          // Check if we sent data before failing\n          if (textSent || toolCallsSent.length > 0) {\n            const extra = { textSent, toolCallsSent };\n\n            if (!this.adapter.retryOnChunkSent) {\n              this._log.error(\n                { llm: llm.label(), ...extra },\n                'failed after sending chunk, skip retrying. Set `retryOnChunkSent` to `true` to enable.',\n              );\n              throw error;\n            }\n\n            this._log.warn(\n              { llm: llm.label(), ...extra },\n              'failed after sending chunk, retrying...',\n            );\n          }\n        }\n      }\n\n      // Trigger background recovery for this LLM\n      this.tryRecovery(llm, i);\n    }\n\n    // All LLMs failed\n    const duration = (Date.now() - startTime) / 1000;\n    const labels = this.adapter.llms.map((l) => l.label()).join(', ');\n    throw new APIConnectionError({\n      message: `all LLMs failed (${labels}) after ${duration.toFixed(2)}s`,\n    });\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,wBAA6C;AAC7C,iBAAoB;AACpB,mBAAoE;AAGpE,iBAA+B;AAO/B,MAAM,uCAA0D;AAAA,EAC9D,UAAU;AAAA,EACV,WAAW,yCAA4B;AAAA,EACvC,iBAAiB,yCAA4B;AAC/C;AA8CO,MAAM,wBAAwB,eAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGT;AAAA,EAEQ,aAAS,gBAAI;AAAA,EAErB,YAAY,SAAiC;AAC3C,UAAM;AAEN,QAAI,CAAC,QAAQ,QAAQ,QAAQ,KAAK,SAAS,GAAG;AAC5C,YAAM,IAAI,MAAM,6CAA6C;AAAA,IAC/D;AAEA,SAAK,OAAO,QAAQ;AACpB,SAAK,iBAAiB,QAAQ,kBAAkB;AAChD,SAAK,iBAAiB,QAAQ,kBAAkB;AAChD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,mBAAmB,QAAQ,oBAAoB;AAGpD,SAAK,UAAU,KAAK,KAAK,IAAI,OAAO;AAAA,MAClC,WAAW;AAAA,MACX,gBAAgB;AAAA,IAClB,EAAE;AAGF,eAAW,OAAO,KAAK,MAAM;AAC3B,UAAI,GAAG,qBAAqB,CAAC,YAAY;AACvC,aAAK,KAAK,qBAAqB,OAAO;AAAA,MACxC,CAAC;AAAA,IACH;AAAA,EACF;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO;AAAA,EACT;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,KAAK,MAOS;AACZ,WAAO,IAAI,kBAAkB,MAAM;AAAA,MACjC,SAAS,KAAK;AAAA,MACd,SAAS,KAAK;AAAA,MACd,aAAa,KAAK,eAAe;AAAA,MACjC,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,aAAa,KAAK;AAAA,IACpB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,yBAAyB,KAAU,WAA0B;AAC3D,UAAM,QAAkC,EAAE,KAAK,UAAU;AAEzD,IAAC,KAAsF;AAAA,MACrF;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAMA,MAAM,0BAA0B,qBAAU;AAAA,EAChC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA,WAAO,gBAAI;AAAA,EAEnB,YACE,SACA,MAQA;AACA,UAAM,SAAS;AAAA,MACb,SAAS,KAAK;AAAA,MACd,SAAS,KAAK;AAAA,MACd,aAAa,KAAK;AAAA,IACpB,CAAC;AACD,SAAK,UAAU;AACf,SAAK,oBAAoB,KAAK;AAC9B,SAAK,aAAa,KAAK;AACvB,SAAK,cAAc,KAAK;AAAA,EAC1B;AAAA;AAAA;AAAA;AAAA,EAKA,IAAa,UAAuB;AArLtC;AAsLI,aAAO,UAAK,mBAAL,mBAAqB,YAAW,MAAM;AAAA,EAC/C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,OAAe,YACb,KACA,gBAAyB,OACiB;AAC1C,UAAM,cAAiC;AAAA,MACrC,GAAG,KAAK;AAAA,MACR,UAAU,KAAK,QAAQ;AAAA,MACvB,WAAW,KAAK,QAAQ,iBAAiB;AAAA,MACzC,iBAAiB,KAAK,QAAQ,gBAAgB;AAAA,IAChD;AAEA,UAAM,SAAS,IAAI,KAAK;AAAA,MACtB,SAAS,MAAM;AAAA,MACf,SAAS,KAAK;AAAA,MACd;AAAA,MACA,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,aAAa,KAAK;AAAA,IACpB,CAAC;AAGD,QAAI;AACJ,UAAM,eAAe,CAAC,OAAyB;AAC7C,oBAAc,GAAG;AAAA,IACnB;AACA,QAAI,GAAG,SAAS,YAAY;AAE5B,QAAI;AACF,UAAI,mBAAmB,CAAC;AACxB,uBAAiB,SAAS,QAAQ;AAChC,YAAI,kBAAkB;AACpB,6BAAmB;AACnB,eAAK,iBAAiB;AAAA,QACxB;AACA,cAAM;AAAA,MACR;AAGA,UAAI,aAAa;AACf,cAAM;AAAA,MACR;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,4BAAU;AAC7B,YAAI,eAAe;AACjB,eAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,GAAG,MAAM,GAAG,iBAAiB;AAAA,QAC/D,OAAO;AACL,eAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,GAAG,MAAM,GAAG,+BAA+B;AAAA,QAC7E;AACA,cAAM;AAAA,MACR;AAGA,UAAI,iBAAiB,SAAS,MAAM,SAAS,cAAc;AACzD,YAAI,eAAe;AACjB,eAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,EAAE,GAAG,oBAAoB;AAAA,QAC3D,OAAO;AACL,eAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,EAAE,GAAG,kCAAkC;AAAA,QACzE;AACA,cAAM;AAAA,MACR;AAGA,UAAI,eAAe;AACjB,aAAK,KAAK,MAAM,EAAE,KAAK,IAAI,MAAM,GAAG,MAAM,GAAG,2BAA2B;AAAA,MAC1E,OAAO;AACL,aAAK,KAAK,MAAM,EAAE,KAAK,IAAI,MAAM,GAAG,MAAM,GAAG,yCAAyC;AAAA,MACxF;AACA,YAAM;AAAA,IACR,UAAE;AACA,UAAI,IAAI,SAAS,YAAY;AAAA,IAC/B;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKQ,YAAY,KAAU,OAAqB;AACjD,UAAM,SAAS,KAAK,QAAQ,QAAQ,KAAK;AAGzC,QAAI,OAAO,mBAAmB,MAAM;AAClC;AAAA,IACF;AAEA,UAAM,cAAc,YAA2B;AAC7C,UAAI;AAGF,yBAAiB,UAAU,KAAK,YAAY,KAAK,IAAI,GAAG;AAAA,QAExD;AAGA,eAAO,YAAY;AACnB,aAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,EAAE,GAAG,eAAe;AACpD,aAAK,QAAQ,yBAAyB,KAAK,IAAI;AAAA,MACjD,QAAQ;AAAA,MAER,UAAE;AACA,eAAO,iBAAiB;AAAA,MAC1B;AAAA,IACF;AAGA,WAAO,iBAAiB,YAAY;AAAA,EACtC;AAAA;AAAA;AAAA;AAAA,EAKA,MAAgB,MAAqB;AACnC,UAAM,YAAY,KAAK,IAAI;AAG3B,UAAM,YAAY,KAAK,QAAQ,QAAQ,MAAM,CAAC,MAAM,CAAC,EAAE,SAAS;AAChE,QAAI,WAAW;AACb,WAAK,KAAK,MAAM,uCAAuC;AAAA,IACzD;AAEA,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,QAAQ,KAAK;AACjD,YAAM,MAAM,KAAK,QAAQ,KAAK,CAAC;AAC/B,YAAM,SAAS,KAAK,QAAQ,QAAQ,CAAC;AAErC,WAAK,KAAK;AAAA,QACR,EAAE,KAAK,IAAI,MAAM,GAAG,OAAO,GAAG,WAAW,OAAO,WAAW,UAAU;AAAA,QACrE;AAAA,MACF;AAEA,UAAI,OAAO,aAAa,WAAW;AACjC,YAAI,WAAW;AACf,cAAM,gBAA0B,CAAC;AAEjC,YAAI;AACF,eAAK,KAAK,KAAK,EAAE,KAAK,IAAI,MAAM,EAAE,GAAG,sCAAsC;AAE3E,cAAI,aAAa;AACjB,2BAAiB,SAAS,KAAK,YAAY,KAAK,KAAK,GAAG;AACtD;AAEA,gBAAI,MAAM,OAAO;AACf,kBAAI,MAAM,MAAM,SAAS;AACvB,4BAAY,MAAM,MAAM;AAAA,cAC1B;AACA,kBAAI,MAAM,MAAM,WAAW;AACzB,2BAAW,MAAM,MAAM,MAAM,WAAW;AACtC,sBAAI,GAAG,MAAM;AACX,kCAAc,KAAK,GAAG,IAAI;AAAA,kBAC5B;AAAA,gBACF;AAAA,cACF;AAAA,YACF;AAGA,iBAAK,KAAK,MAAM,EAAE,KAAK,IAAI,MAAM,GAAG,WAAW,GAAG,gCAAgC;AAClF,iBAAK,MAAM,IAAI,KAAK;AAAA,UACtB;AAGA,eAAK,KAAK;AAAA,YACR,EAAE,KAAK,IAAI,MAAM,GAAG,aAAa,YAAY,YAAY,SAAS,OAAO;AAAA,YACzE;AAAA,UACF;AACA;AAAA,QACF,SAAS,OAAO;AAEd,cAAI,OAAO,WAAW;AACpB,mBAAO,YAAY;AACnB,iBAAK,QAAQ,yBAAyB,KAAK,KAAK;AAAA,UAClD;AAGA,cAAI,YAAY,cAAc,SAAS,GAAG;AACxC,kBAAM,QAAQ,EAAE,UAAU,cAAc;AAExC,gBAAI,CAAC,KAAK,QAAQ,kBAAkB;AAClC,mBAAK,KAAK;AAAA,gBACR,EAAE,KAAK,IAAI,MAAM,GAAG,GAAG,MAAM;AAAA,gBAC7B;AAAA,cACF;AACA,oBAAM;AAAA,YACR;AAEA,iBAAK,KAAK;AAAA,cACR,EAAE,KAAK,IAAI,MAAM,GAAG,GAAG,MAAM;AAAA,cAC7B;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAGA,WAAK,YAAY,KAAK,CAAC;AAAA,IACzB;AAGA,UAAM,YAAY,KAAK,IAAI,IAAI,aAAa;AAC5C,UAAM,SAAS,KAAK,QAAQ,KAAK,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,EAAE,KAAK,IAAI;AAChE,UAAM,IAAI,qCAAmB;AAAA,MAC3B,SAAS,oBAAoB,MAAM,WAAW,SAAS,QAAQ,CAAC,CAAC;AAAA,IACnE,CAAC;AAAA,EACH;AACF;","names":[]}

package/dist/llm/fallback_adapter.d.cts

@@ -0,0 +1,73 @@
+import { type APIConnectOptions } from '../types.js';
+import type { ChatContext } from './chat_context.js';
+import { LLM, LLMStream } from './llm.js';
+import type { ToolChoice, ToolContext } from './tool_context.js';
+/**
+ * Internal status tracking for each LLM instance.
+ */
+interface LLMStatus {
+    available: boolean;
+    recoveringTask: Promise<void> | null;
+}
+/**
+ * Event emitted when an LLM's availability changes.
+ */
+export interface AvailabilityChangedEvent {
+    llm: LLM;
+    available: boolean;
+}
+/**
+ * Options for creating a FallbackAdapter.
+ */
+export interface FallbackAdapterOptions {
+    /** List of LLM instances to fallback to (in order). */
+    llms: LLM[];
+    /** Timeout for each LLM attempt in seconds. Defaults to 5.0. */
+    attemptTimeout?: number;
+    /** Internal retries per LLM before moving to next. Defaults to 0. */
+    maxRetryPerLLM?: number;
+    /** Interval between retries in seconds. Defaults to 0.5. */
+    retryInterval?: number;
+    /** Whether to retry when LLM fails after chunks are sent. Defaults to false. */
+    retryOnChunkSent?: boolean;
+}
+/**
+ * FallbackAdapter is an LLM that can fallback to a different LLM if the current LLM fails.
+ *
+ * @example
+ * ```typescript
+ * const fallbackLLM = new FallbackAdapter({
+ *   llms: [primaryLLM, secondaryLLM, tertiaryLLM],
+ *   attemptTimeout: 5.0,
+ *   maxRetryPerLLM: 1,
+ * });
+ * ```
+ */
+export declare class FallbackAdapter extends LLM {
+    readonly llms: LLM[];
+    readonly attemptTimeout: number;
+    readonly maxRetryPerLLM: number;
+    readonly retryInterval: number;
+    readonly retryOnChunkSent: boolean;
+    /** @internal */
+    _status: LLMStatus[];
+    private logger;
+    constructor(options: FallbackAdapterOptions);
+    get model(): string;
+    label(): string;
+    chat(opts: {
+        chatCtx: ChatContext;
+        toolCtx?: ToolContext;
+        connOptions?: APIConnectOptions;
+        parallelToolCalls?: boolean;
+        toolChoice?: ToolChoice;
+        extraKwargs?: Record<string, unknown>;
+    }): LLMStream;
+    /**
+     * Emit availability changed event.
+     * @internal
+     */
+    _emitAvailabilityChanged(llm: LLM, available: boolean): void;
+}
+export {};
+//# sourceMappingURL=fallback_adapter.d.ts.map