@livekit/agents 1.1.0 → 1.2.0

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  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyModels, createAccessToken } from './utils.js';\n\nexport type AzureModels =\n  // | \"azure/gpt-5\"\n  // | \"azure/gpt-5-mini\"\n  // | \"azure/gpt-5-nano\"\n  'azure/gpt-4.1' | 'azure/gpt-4.1-mini' | 'azure/gpt-4.1-nano';\n// | \"azure/gpt-4o\"\n// | \"azure/gpt-4o-mini\"\n\n// https://inference-docs.cerebras.ai/models/overview\nexport type CerebrasModels =\n  // production models\n  | 'cerebras/llama3.1-8b'\n  | 'cerebras/llama-3.3-70b'\n  | 'cerebras/llama-4-scout-17b-16e-instruct'\n  | 'cerebras/gpt-oss-120b'\n  | 'cerebras/qwen-3-32b'\n  // preview models\n  | 'cerebras/llama-4-maverick-17b-128e-instruct'\n  | 'cerebras/qwen-3-235b-a22b-instruct-2507';\n\n// https://console.groq.com/docs/models\nexport type GroqModels =\n  // production models\n  | 'groq/llama-3.1-8b-instant'\n  | 'groq/llama-3.3-70b-versatile'\n  | 'groq/openai/gpt-oss-120b'\n  | 'groq/openai/gpt-oss-20b'\n  // preview models\n  | 'groq/meta-llama/llama-4-maverick-17b-128e-instruct'\n  | 'groq/meta-llama/llama-4-scout-17b-16e-instruct'\n  | 'groq/qwen/qwen3-32b';\n\n// https://www.baseten.co/library/tag/llms\nexport type BasetenModels =\n  | 'baseten/deepseek-ai/DeepSeek-V3-0324'\n  | 'baseten/meta-llama/Llama-4-Scout-17B-16E-Instruct'\n  | 'baseten/meta-llama/Llama-4-Maverick-17B-128E-Instruct'\n  | 'baseten/moonshotai/Kimi-K2-Instruct'\n  | 'baseten/openai/gpt-oss-120b'\n  | 'baseten/Qwen/Qwen3-235B-A22B-Instruct-2507';\n\nexport interface AzureOptions {\n  top_p?: number;\n}\n\nexport interface CerebrasOptions {\n  top_p?: number;\n}\n\nexport interface GroqOptions {\n  top_p?: number;\n}\n\nexport interface BasetenOptions {\n  top_p?: number;\n}\n\nexport type LLMModels = AzureModels | CerebrasModels | GroqModels | BasetenModels | AnyModels;\n\nexport type LLMOptions<T extends LLMModels> = T extends AzureModels\n  ? AzureOptions\n  : T extends CerebrasModels\n    ? CerebrasOptions\n    : T extends GroqOptions\n      ? GroqOptions\n      : T extends BasetenOptions\n        ? BasetenOptions\n        : Record<string, unknown>;\n\nexport type Verbosity = 'low' | 'medium' | 'high';\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport interface InferenceLLMOptions<TModel extends LLMModels> {\n  model: TModel;\n  temperature?: number;\n  parallelToolCalls?: boolean;\n  toolChoice?: llm.ToolChoice;\n  maxCompletionTokens?: number;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  verbosity?: Verbosity;\n  extraKwargs: Record<string, any>; // eslint-disable-line @typescript-eslint/no-explicit-any\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\nexport class LLM<TModel extends LLMModels> extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions<TModel>;\n\n  constructor(opts: {\n    model: TModel;\n    temperature?: number;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    maxCompletionTokens?: number;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    maxRetries?: number;\n    timeout?: number;\n    verbosity?: Verbosity;\n    extraKwargs?: LLMOptions<TModel>;\n  }) {\n    super();\n\n    const {\n      model,\n      temperature,\n      parallelToolCalls,\n      toolChoice,\n      maxCompletionTokens,\n      baseURL,\n      apiKey,\n      apiSecret,\n      maxRetries,\n      timeout,\n      verbosity,\n      extraKwargs,\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      temperature,\n      parallelToolCalls,\n      toolChoice,\n      verbosity,\n      maxCompletionTokens,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      extraKwargs: extraKwargs || {},\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      maxRetries: maxRetries || 0,\n      timeout: timeout || 15000,\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\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<TModel> {\n    let extras: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    if (this.opts.maxCompletionTokens !== undefined) {\n      extras.max_completion_tokens = this.opts.maxCompletionTokens;\n    }\n    if (this.opts.temperature !== undefined) {\n      extras.temperature = this.opts.temperature;\n    }\n    if (this.opts.verbosity !== undefined) {\n      extras.verbosity = this.opts.verbosity;\n    }\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined ? parallelToolCalls : this.opts.parallelToolCalls;\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      extras.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice = toolChoice !== undefined ? toolChoice : this.opts.toolChoice;\n    if (toolChoice) {\n      extras.tool_choice = toolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    extras = { ...extras, ...this.opts.extraKwargs };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      providerFmt: 'openai',\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      extraKwargs: extras,\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream<TModel extends LLMModels> extends llm.LLMStream {\n  private model: TModel;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private extraKwargs: Record<string, unknown>;\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<TModel>,\n    {\n      model,\n      providerFmt,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      extraKwargs,\n    }: {\n      model: TModel;\n      providerFmt: llm.ProviderFormat;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      extraKwargs: Record<string, any>;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.providerFmt = providerFmt;\n    this.extraKwargs = extraKwargs;\n    this.model = model;\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            type: 'function' as const,\n            function: {\n              name,\n              description: func.description,\n              parameters: llm.toJsonSchema(\n                func.parameters,\n              ) as unknown as OpenAI.Chat.Completions.ChatCompletionTool['function']['parameters'],\n            },\n          }))\n        : undefined;\n\n      const requestExtras: Record<string, unknown> = { ...this.extraKwargs };\n      if (!tools) {\n        delete requestExtras.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      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          ...requestExtras,\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.request_id,\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,eAMO;AACP,UAAqB;AAErB,mBAAkD;AAwElD,MAAM,mBAAmB;AAoBlB,MAAM,YAAsC,IAAI,IAAI;AAAA,EACjD;AAAA,EACA;AAAA,EAER,YAAY,MAaT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,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;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,aAAa,eAAe,CAAC;AAAA,IAC/B;AAEA,SAAK,SAAS,IAAI,cAAAA,QAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,YAAY,cAAc;AAAA,MAC1B,SAAS,WAAW;AAAA,IACtB,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQsB;AACpB,QAAI,SAAkC,EAAE,GAAI,eAAe,CAAC,EAAG;AAE/D,QAAI,KAAK,KAAK,wBAAwB,QAAW;AAC/C,aAAO,wBAAwB,KAAK,KAAK;AAAA,IAC3C;AACA,QAAI,KAAK,KAAK,gBAAgB,QAAW;AACvC,aAAO,cAAc,KAAK,KAAK;AAAA,IACjC;AACA,QAAI,KAAK,KAAK,cAAc,QAAW;AACrC,aAAO,YAAY,KAAK,KAAK;AAAA,IAC/B;AAEA,wBACE,sBAAsB,SAAY,oBAAoB,KAAK,KAAK;AAClE,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,aAAO,sBAAsB;AAAA,IAC/B;AAEA,iBAAa,eAAe,SAAY,aAAa,KAAK,KAAK;AAC/D,QAAI,YAAY;AACd,aAAO,cAAc;AAAA,IACvB;AAIA,aAAS,EAAE,GAAG,QAAQ,GAAG,KAAK,KAAK,YAAY;AAE/C,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,aAAa;AAAA,MACb,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa;AAAA,MACb,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAA4C,IAAI,UAAU;AAAA,EAC7D;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,EACF,GAUA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,cAAc;AACnB,SAAK,cAAc;AACnB,SAAK,QAAQ;AAAA,EACf;AAAA,EAEA,MAAgB,MAAqB;AA1RvC;AA6RI,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,OAAO;AAAA,QAClD,MAAM;AAAA,QACN,UAAU;AAAA,UACR;AAAA,UACA,aAAa,KAAK;AAAA,UAClB,YAAY,IAAI;AAAA,YACd,KAAK;AAAA,UACP;AAAA,QACF;AAAA,MACF,EAAE,IACF;AAEJ,YAAM,gBAAyC,EAAE,GAAG,KAAK,YAAY;AACrE,UAAI,CAAC,OAAO;AACV,eAAO,cAAc;AAAA,MACvB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,UAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;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 { APIConnectionError, APIStatusError, APITimeoutError } from '../_exceptions.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { DEFAULT_API_CONNECT_OPTIONS } from '../types.js';\nimport { type Expand, toError } from '../utils.js';\nimport { type AnyString, createAccessToken, getDefaultInferenceUrl } from './utils.js';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5.4'\n  | 'openai/gpt-5.3-chat-latest'\n  | 'openai/gpt-5.2'\n  | 'openai/gpt-5.2-chat-latest'\n  | 'openai/gpt-5.1'\n  | 'openai/gpt-5.1-chat-latest'\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 =\n  | 'google/gemini-3-pro'\n  | 'google/gemini-3-flash'\n  | 'google/gemini-2.5-pro'\n  | 'google/gemini-2.5-flash'\n  | 'google/gemini-2.5-flash-lite'\n  | 'google/gemini-2.0-flash'\n  | 'google/gemini-2.0-flash-lite';\n\nexport type MoonshotModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3' | 'deepseek-ai/deepseek-v3.2';\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 = OpenAIModels | GoogleModels | MoonshotModels | DeepSeekModels | 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 || getDefaultInferenceUrl();\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  get provider(): string {\n    return 'livekit';\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  private toolExtra?: Record<string, unknown>;\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    this.toolExtra = 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          // Note: We intentionally do NOT reset toolExtra here.\n          // For Gemini 3+, the thought_signature is only provided on the first tool call\n          // in a parallel batch, but must be applied to ALL tool calls in the batch.\n          // We preserve toolExtra so subsequent tool calls inherit the thought_signature.\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          // Extract extra from tool call (e.g., Google thought signatures)\n          // Only update toolExtra if this tool call has extra_content.\n          // Otherwise, inherit from previous tool call (for parallel Gemini tool calls).\n          const newToolExtra =\n            // eslint-disable-next-line @typescript-eslint/no-explicit-any\n            ((tool as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n          if (newToolExtra) {\n            this.toolExtra = newToolExtra;\n          }\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      // Reset toolExtra at the end of the response (not between parallel tool calls)\n      this.toolExtra = undefined;\n      return callChunk;\n    }\n\n    // Extract extra from delta (e.g., Google thought signatures on text parts)\n    const deltaExtra =\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n\n    // Regular content message\n    if (!delta.content && !deltaExtra) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        extra: deltaExtra,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    const toolExtra = this.toolExtra ? { ...this.toolExtra } : {};\n    const thoughtSignature = this.extractThoughtSignature(toolExtra);\n    const deltaExtra =\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        extra: deltaExtra,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n            extra: toolExtra,\n            thoughtSignature,\n          }),\n        ],\n      },\n    };\n  }\n\n  private extractThoughtSignature(extra?: Record<string, unknown>): string | undefined {\n    const googleExtra = extra?.google;\n    if (googleExtra && typeof googleExtra === 'object') {\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      return (googleExtra as any).thoughtSignature || (googleExtra as any).thought_signature;\n    }\n    return undefined;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAAmB;AACnB,wBAAoE;AACpE,UAAqB;AAErB,mBAA4C;AAC5C,mBAAqC;AACrC,IAAAA,gBAA0E;AA0FnE,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,eAAW,sCAAuB;AACpD,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,cAAAC,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,IAAI,WAAmB;AACrB,WAAO;AAAA,EACT;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,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;AArRvC;AAwRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AACzE,SAAK,YAAY;AAEjB,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,kCAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,cAAAA,QAAO,UAAU;AAC3C,cAAM,IAAI,iCAAe;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,qCAAmB;AAAA,UAC3B,aAAS,sBAAQ,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,QAK1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAIlD,gBAAM;AAAA;AAAA,YAEF,KAAa,iBAAyD;AAAA;AAC1E,cAAI,cAAc;AAChB,iBAAK,YAAY;AAAA,UACnB;AAAA,QACF,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;AAExD,WAAK,YAAY;AACjB,aAAO;AAAA,IACT;AAGA,UAAM;AAAA;AAAA,MAEF,MAAc,iBAAyD;AAAA;AAG3E,QAAI,CAAC,MAAM,WAAW,CAAC,YAAY;AACjC,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,OAAO;AAAA,MACT;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,UAAM,YAAY,KAAK,YAAY,EAAE,GAAG,KAAK,UAAU,IAAI,CAAC;AAC5D,UAAM,mBAAmB,KAAK,wBAAwB,SAAS;AAC/D,UAAM;AAAA;AAAA,MAEF,MAAc,iBAAyD;AAAA;AAE3E,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,OAAO;AAAA,QACP,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,YAC9B,OAAO;AAAA,YACP;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,wBAAwB,OAAqD;AACnF,UAAM,cAAc,+BAAO;AAC3B,QAAI,eAAe,OAAO,gBAAgB,UAAU;AAElD,aAAQ,YAAoB,oBAAqB,YAAoB;AAAA,IACvE;AACA,WAAO;AAAA,EACT;AACF;","names":["import_utils","OpenAI","llm"]}

package/dist/inference/llm.d.cts

@@ -1,61 +1,76 @@
 import OpenAI from 'openai';
 import * as llm from '../llm/index.js';
 import type { APIConnectOptions } from '../types.js';
-import { type AnyModels } from './utils.js';
-export type AzureModels = 'azure/gpt-4.1' | 'azure/gpt-4.1-mini' | 'azure/gpt-4.1-nano';
-export type CerebrasModels = 'cerebras/llama3.1-8b' | 'cerebras/llama-3.3-70b' | 'cerebras/llama-4-scout-17b-16e-instruct' | 'cerebras/gpt-oss-120b' | 'cerebras/qwen-3-32b' | 'cerebras/llama-4-maverick-17b-128e-instruct' | 'cerebras/qwen-3-235b-a22b-instruct-2507';
-export type GroqModels = 'groq/llama-3.1-8b-instant' | 'groq/llama-3.3-70b-versatile' | 'groq/openai/gpt-oss-120b' | 'groq/openai/gpt-oss-20b' | 'groq/meta-llama/llama-4-maverick-17b-128e-instruct' | 'groq/meta-llama/llama-4-scout-17b-16e-instruct' | 'groq/qwen/qwen3-32b';
-export type BasetenModels = 'baseten/deepseek-ai/DeepSeek-V3-0324' | 'baseten/meta-llama/Llama-4-Scout-17B-16E-Instruct' | 'baseten/meta-llama/Llama-4-Maverick-17B-128E-Instruct' | 'baseten/moonshotai/Kimi-K2-Instruct' | 'baseten/openai/gpt-oss-120b' | 'baseten/Qwen/Qwen3-235B-A22B-Instruct-2507';
-export interface AzureOptions {
-    top_p?: number;
-}
-export interface CerebrasOptions {
-    top_p?: number;
-}
-export interface GroqOptions {
-    top_p?: number;
-}
-export interface BasetenOptions {
+import { type Expand } from '../utils.js';
+import { type AnyString } from './utils.js';
+export type OpenAIModels = 'openai/gpt-5.4' | 'openai/gpt-5.3-chat-latest' | 'openai/gpt-5.2' | 'openai/gpt-5.2-chat-latest' | 'openai/gpt-5.1' | 'openai/gpt-5.1-chat-latest' | 'openai/gpt-5' | 'openai/gpt-5-mini' | 'openai/gpt-5-nano' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-4.1-nano' | 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-oss-120b';
+export type GoogleModels = 'google/gemini-3-pro' | 'google/gemini-3-flash' | 'google/gemini-2.5-pro' | 'google/gemini-2.5-flash' | 'google/gemini-2.5-flash-lite' | 'google/gemini-2.0-flash' | 'google/gemini-2.0-flash-lite';
+export type MoonshotModels = 'moonshotai/kimi-k2-instruct';
+export type DeepSeekModels = 'deepseek-ai/deepseek-v3' | 'deepseek-ai/deepseek-v3.2';
+type ChatCompletionPredictionContentParam = Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;
+type WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;
+type ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;
+type Verbosity = 'low' | 'medium' | 'high';
+export interface ChatCompletionOptions extends Record<string, unknown> {
+    frequency_penalty?: number;
+    logit_bias?: Record<string, number>;
+    logprobs?: boolean;
+    max_completion_tokens?: number;
+    max_tokens?: number;
+    metadata?: Record<string, string>;
+    modalities?: Array<'text' | 'audio'>;
+    n?: number;
+    parallel_tool_calls?: boolean;
+    prediction?: ChatCompletionPredictionContentParam | null;
+    presence_penalty?: number;
+    prompt_cache_key?: string;
+    reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';
+    safety_identifier?: string;
+    seed?: number;
+    service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';
+    stop?: string | string[];
+    store?: boolean;
+    temperature?: number;
+    top_logprobs?: number;
     top_p?: number;
+    user?: string;
+    verbosity?: Verbosity;
+    web_search_options?: WebSearchOptions;
+    tool_choice?: ToolChoice;
 }
-export type LLMModels = AzureModels | CerebrasModels | GroqModels | BasetenModels | AnyModels;
-export type LLMOptions<T extends LLMModels> = T extends AzureModels ? AzureOptions : T extends CerebrasModels ? CerebrasOptions : T extends GroqOptions ? GroqOptions : T extends BasetenOptions ? BasetenOptions : Record<string, unknown>;
-export type Verbosity = 'low' | 'medium' | 'high';
-export interface InferenceLLMOptions<TModel extends LLMModels> {
-    model: TModel;
-    temperature?: number;
-    parallelToolCalls?: boolean;
-    toolChoice?: llm.ToolChoice;
-    maxCompletionTokens?: number;
+export type LLMModels = OpenAIModels | GoogleModels | MoonshotModels | DeepSeekModels | AnyString;
+export interface InferenceLLMOptions {
+    model: LLMModels;
+    provider?: string;
     baseURL: string;
     apiKey: string;
     apiSecret: string;
-    verbosity?: Verbosity;
-    extraKwargs: Record<string, any>;
+    modelOptions: ChatCompletionOptions;
+    strictToolSchema?: boolean;
 }
 export interface GatewayOptions {
     apiKey: string;
     apiSecret: string;
 }
-export declare class LLM<TModel extends LLMModels> extends llm.LLM {
+/**
+ * Livekit Cloud Inference LLM
+ */
+export declare class LLM extends llm.LLM {
     private client;
     private opts;
     constructor(opts: {
-        model: TModel;
-        temperature?: number;
-        parallelToolCalls?: boolean;
-        toolChoice?: llm.ToolChoice;
-        maxCompletionTokens?: number;
+        model: LLMModels;
+        provider?: string;
         baseURL?: string;
         apiKey?: string;
         apiSecret?: string;
-        maxRetries?: number;
-        timeout?: number;
-        verbosity?: Verbosity;
-        extraKwargs?: LLMOptions<TModel>;
+        modelOptions?: InferenceLLMOptions['modelOptions'];
+        strictToolSchema?: boolean;
     });
     label(): string;
     get model(): string;
+    get provider(): string;
+    static fromModelString(modelString: string): LLM;
     chat({ chatCtx, toolCtx, connOptions, parallelToolCalls, toolChoice, extraKwargs, }: {
         chatCtx: llm.ChatContext;
         toolCtx?: llm.ToolContext;
@@ -63,30 +78,37 @@ export declare class LLM<TModel extends LLMModels> extends llm.LLM {
         parallelToolCalls?: boolean;
         toolChoice?: llm.ToolChoice;
         extraKwargs?: Record<string, unknown>;
-    }): LLMStream<TModel>;
+    }): LLMStream;
 }
-export declare class LLMStream<TModel extends LLMModels> extends llm.LLMStream {
+export declare class LLMStream extends llm.LLMStream {
     private model;
+    private provider?;
     private providerFmt;
     private client;
-    private extraKwargs;
+    private modelOptions;
+    private strictToolSchema;
     private gatewayOptions?;
     private toolCallId?;
     private toolIndex?;
     private fncName?;
     private fncRawArguments?;
-    constructor(llm: LLM<TModel>, { model, providerFmt, client, chatCtx, toolCtx, gatewayOptions, connOptions, extraKwargs, }: {
-        model: TModel;
-        providerFmt: llm.ProviderFormat;
+    private toolExtra?;
+    constructor(llm: LLM, { model, provider, client, chatCtx, toolCtx, gatewayOptions, connOptions, modelOptions, providerFmt, strictToolSchema, }: {
+        model: LLMModels;
+        provider?: string;
         client: OpenAI;
         chatCtx: llm.ChatContext;
         toolCtx?: llm.ToolContext;
         gatewayOptions?: GatewayOptions;
         connOptions: APIConnectOptions;
-        extraKwargs: Record<string, any>;
+        modelOptions: Record<string, unknown>;
+        providerFmt?: llm.ProviderFormat;
+        strictToolSchema: boolean;
     });
     protected run(): Promise<void>;
     private parseChoice;
     private createRunningToolCallChunk;
+    private extractThoughtSignature;
 }
+export {};
 //# sourceMappingURL=llm.d.ts.map

package/dist/inference/llm.d.ts

@@ -1,61 +1,76 @@
 import OpenAI from 'openai';
 import * as llm from '../llm/index.js';
 import type { APIConnectOptions } from '../types.js';
-import { type AnyModels } from './utils.js';
-export type AzureModels = 'azure/gpt-4.1' | 'azure/gpt-4.1-mini' | 'azure/gpt-4.1-nano';
-export type CerebrasModels = 'cerebras/llama3.1-8b' | 'cerebras/llama-3.3-70b' | 'cerebras/llama-4-scout-17b-16e-instruct' | 'cerebras/gpt-oss-120b' | 'cerebras/qwen-3-32b' | 'cerebras/llama-4-maverick-17b-128e-instruct' | 'cerebras/qwen-3-235b-a22b-instruct-2507';
-export type GroqModels = 'groq/llama-3.1-8b-instant' | 'groq/llama-3.3-70b-versatile' | 'groq/openai/gpt-oss-120b' | 'groq/openai/gpt-oss-20b' | 'groq/meta-llama/llama-4-maverick-17b-128e-instruct' | 'groq/meta-llama/llama-4-scout-17b-16e-instruct' | 'groq/qwen/qwen3-32b';
-export type BasetenModels = 'baseten/deepseek-ai/DeepSeek-V3-0324' | 'baseten/meta-llama/Llama-4-Scout-17B-16E-Instruct' | 'baseten/meta-llama/Llama-4-Maverick-17B-128E-Instruct' | 'baseten/moonshotai/Kimi-K2-Instruct' | 'baseten/openai/gpt-oss-120b' | 'baseten/Qwen/Qwen3-235B-A22B-Instruct-2507';
-export interface AzureOptions {
-    top_p?: number;
-}
-export interface CerebrasOptions {
-    top_p?: number;
-}
-export interface GroqOptions {
-    top_p?: number;
-}
-export interface BasetenOptions {
+import { type Expand } from '../utils.js';
+import { type AnyString } from './utils.js';
+export type OpenAIModels = 'openai/gpt-5.4' | 'openai/gpt-5.3-chat-latest' | 'openai/gpt-5.2' | 'openai/gpt-5.2-chat-latest' | 'openai/gpt-5.1' | 'openai/gpt-5.1-chat-latest' | 'openai/gpt-5' | 'openai/gpt-5-mini' | 'openai/gpt-5-nano' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-4.1-nano' | 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-oss-120b';
+export type GoogleModels = 'google/gemini-3-pro' | 'google/gemini-3-flash' | 'google/gemini-2.5-pro' | 'google/gemini-2.5-flash' | 'google/gemini-2.5-flash-lite' | 'google/gemini-2.0-flash' | 'google/gemini-2.0-flash-lite';
+export type MoonshotModels = 'moonshotai/kimi-k2-instruct';
+export type DeepSeekModels = 'deepseek-ai/deepseek-v3' | 'deepseek-ai/deepseek-v3.2';
+type ChatCompletionPredictionContentParam = Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;
+type WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;
+type ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;
+type Verbosity = 'low' | 'medium' | 'high';
+export interface ChatCompletionOptions extends Record<string, unknown> {
+    frequency_penalty?: number;
+    logit_bias?: Record<string, number>;
+    logprobs?: boolean;
+    max_completion_tokens?: number;
+    max_tokens?: number;
+    metadata?: Record<string, string>;
+    modalities?: Array<'text' | 'audio'>;
+    n?: number;
+    parallel_tool_calls?: boolean;
+    prediction?: ChatCompletionPredictionContentParam | null;
+    presence_penalty?: number;
+    prompt_cache_key?: string;
+    reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';
+    safety_identifier?: string;
+    seed?: number;
+    service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';
+    stop?: string | string[];
+    store?: boolean;
+    temperature?: number;
+    top_logprobs?: number;
     top_p?: number;
+    user?: string;
+    verbosity?: Verbosity;
+    web_search_options?: WebSearchOptions;
+    tool_choice?: ToolChoice;
 }
-export type LLMModels = AzureModels | CerebrasModels | GroqModels | BasetenModels | AnyModels;
-export type LLMOptions<T extends LLMModels> = T extends AzureModels ? AzureOptions : T extends CerebrasModels ? CerebrasOptions : T extends GroqOptions ? GroqOptions : T extends BasetenOptions ? BasetenOptions : Record<string, unknown>;
-export type Verbosity = 'low' | 'medium' | 'high';
-export interface InferenceLLMOptions<TModel extends LLMModels> {
-    model: TModel;
-    temperature?: number;
-    parallelToolCalls?: boolean;
-    toolChoice?: llm.ToolChoice;
-    maxCompletionTokens?: number;
+export type LLMModels = OpenAIModels | GoogleModels | MoonshotModels | DeepSeekModels | AnyString;
+export interface InferenceLLMOptions {
+    model: LLMModels;
+    provider?: string;
     baseURL: string;
     apiKey: string;
     apiSecret: string;
-    verbosity?: Verbosity;
-    extraKwargs: Record<string, any>;
+    modelOptions: ChatCompletionOptions;
+    strictToolSchema?: boolean;
 }
 export interface GatewayOptions {
     apiKey: string;
     apiSecret: string;
 }
-export declare class LLM<TModel extends LLMModels> extends llm.LLM {
+/**
+ * Livekit Cloud Inference LLM
+ */
+export declare class LLM extends llm.LLM {
     private client;
     private opts;
     constructor(opts: {
-        model: TModel;
-        temperature?: number;
-        parallelToolCalls?: boolean;
-        toolChoice?: llm.ToolChoice;
-        maxCompletionTokens?: number;
+        model: LLMModels;
+        provider?: string;
         baseURL?: string;
         apiKey?: string;
         apiSecret?: string;
-        maxRetries?: number;
-        timeout?: number;
-        verbosity?: Verbosity;
-        extraKwargs?: LLMOptions<TModel>;
+        modelOptions?: InferenceLLMOptions['modelOptions'];
+        strictToolSchema?: boolean;
     });
     label(): string;
     get model(): string;
+    get provider(): string;
+    static fromModelString(modelString: string): LLM;
     chat({ chatCtx, toolCtx, connOptions, parallelToolCalls, toolChoice, extraKwargs, }: {
         chatCtx: llm.ChatContext;
         toolCtx?: llm.ToolContext;
@@ -63,30 +78,37 @@ export declare class LLM<TModel extends LLMModels> extends llm.LLM {
         parallelToolCalls?: boolean;
         toolChoice?: llm.ToolChoice;
         extraKwargs?: Record<string, unknown>;
-    }): LLMStream<TModel>;
+    }): LLMStream;
 }
-export declare class LLMStream<TModel extends LLMModels> extends llm.LLMStream {
+export declare class LLMStream extends llm.LLMStream {
     private model;
+    private provider?;
     private providerFmt;
     private client;
-    private extraKwargs;
+    private modelOptions;
+    private strictToolSchema;
     private gatewayOptions?;
     private toolCallId?;
     private toolIndex?;
     private fncName?;
     private fncRawArguments?;
-    constructor(llm: LLM<TModel>, { model, providerFmt, client, chatCtx, toolCtx, gatewayOptions, connOptions, extraKwargs, }: {
-        model: TModel;
-        providerFmt: llm.ProviderFormat;
+    private toolExtra?;
+    constructor(llm: LLM, { model, provider, client, chatCtx, toolCtx, gatewayOptions, connOptions, modelOptions, providerFmt, strictToolSchema, }: {
+        model: LLMModels;
+        provider?: string;
         client: OpenAI;
         chatCtx: llm.ChatContext;
         toolCtx?: llm.ToolContext;
         gatewayOptions?: GatewayOptions;
         connOptions: APIConnectOptions;
-        extraKwargs: Record<string, any>;
+        modelOptions: Record<string, unknown>;
+        providerFmt?: llm.ProviderFormat;
+        strictToolSchema: boolean;
     });
     protected run(): Promise<void>;
     private parseChoice;
     private createRunningToolCallChunk;
+    private extractThoughtSignature;
 }
+export {};
 //# sourceMappingURL=llm.d.ts.map

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;AAQ5B,OAAO,KAAK,GAAG,MAAM,iBAAiB,CAAC;AACvC,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AACrD,OAAO,EAAE,KAAK,SAAS,EAAqB,MAAM,YAAY,CAAC;AAE/D,MAAM,MAAM,WAAW,GAIrB,eAAe,GAAG,oBAAoB,GAAG,oBAAoB,CAAC;AAKhE,MAAM,MAAM,cAAc,GAEtB,sBAAsB,GACtB,wBAAwB,GACxB,yCAAyC,GACzC,uBAAuB,GACvB,qBAAqB,GAErB,6CAA6C,GAC7C,yCAAyC,CAAC;AAG9C,MAAM,MAAM,UAAU,GAElB,2BAA2B,GAC3B,8BAA8B,GAC9B,0BAA0B,GAC1B,yBAAyB,GAEzB,oDAAoD,GACpD,gDAAgD,GAChD,qBAAqB,CAAC;AAG1B,MAAM,MAAM,aAAa,GACrB,sCAAsC,GACtC,mDAAmD,GACnD,uDAAuD,GACvD,qCAAqC,GACrC,6BAA6B,GAC7B,4CAA4C,CAAC;AAEjD,MAAM,WAAW,YAAY;IAC3B,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,eAAe;IAC9B,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,WAAW;IAC1B,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,cAAc;IAC7B,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,MAAM,SAAS,GAAG,WAAW,GAAG,cAAc,GAAG,UAAU,GAAG,aAAa,GAAG,SAAS,CAAC;AAE9F,MAAM,MAAM,UAAU,CAAC,CAAC,SAAS,SAAS,IAAI,CAAC,SAAS,WAAW,GAC/D,YAAY,GACZ,CAAC,SAAS,cAAc,GACtB,eAAe,GACf,CAAC,SAAS,WAAW,GACnB,WAAW,GACX,CAAC,SAAS,cAAc,GACtB,cAAc,GACd,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AAElC,MAAM,MAAM,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAGlD,MAAM,WAAW,mBAAmB,CAAC,MAAM,SAAS,SAAS;IAC3D,KAAK,EAAE,MAAM,CAAC;IACd,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,UAAU,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC;IAC5B,mBAAmB,CAAC,EAAE,MAAM,CAAC;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB,WAAW,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;CAClC;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,qBAAa,GAAG,CAAC,MAAM,SAAS,SAAS,CAAE,SAAQ,GAAG,CAAC,GAAG;IACxD,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,IAAI,CAA8B;gBAE9B,IAAI,EAAE;QAChB,KAAK,EAAE,MAAM,CAAC;QACd,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,iBAAiB,CAAC,EAAE,OAAO,CAAC;QAC5B,UAAU,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC;QAC5B,mBAAmB,CAAC,EAAE,MAAM,CAAC;QAC7B,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,UAAU,CAAC,EAAE,MAAM,CAAC;QACpB,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,SAAS,CAAC,EAAE,SAAS,CAAC;QACtB,WAAW,CAAC,EAAE,UAAU,CAAC,MAAM,CAAC,CAAC;KAClC;IAkDD,KAAK,IAAI,MAAM;IAIf,IAAI,KAAK,IAAI,MAAM,CAElB;IAED,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,CAAC,MAAM,CAAC;CA0CtB;AAED,qBAAa,SAAS,CAAC,MAAM,SAAS,SAAS,CAAE,SAAQ,GAAG,CAAC,SAAS;IACpE,OAAO,CAAC,KAAK,CAAS;IACtB,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,WAAW,CAA0B;IAE7C,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,CAAC,MAAM,CAAC,EAChB,EACE,KAAK,EACL,WAAW,EACX,MAAM,EACN,OAAO,EACP,OAAO,EACP,cAAc,EACd,WAAW,EACX,WAAW,GACZ,EAAE;QACD,KAAK,EAAE,MAAM,CAAC;QACd,WAAW,EAAE,GAAG,CAAC,cAAc,CAAC;QAChC,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,WAAW,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;KAClC;cAUa,GAAG,IAAI,OAAO,CAAC,IAAI,CAAC;IAqGpC,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;AAE5B,OAAO,KAAK,GAAG,MAAM,iBAAiB,CAAC;AACvC,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAErD,OAAO,EAAE,KAAK,MAAM,EAAW,MAAM,aAAa,CAAC;AACnD,OAAO,EAAE,KAAK,SAAS,EAA6C,MAAM,YAAY,CAAC;AAEvF,MAAM,MAAM,YAAY,GACpB,gBAAgB,GAChB,4BAA4B,GAC5B,gBAAgB,GAChB,4BAA4B,GAC5B,gBAAgB,GAChB,4BAA4B,GAC5B,cAAc,GACd,mBAAmB,GACnB,mBAAmB,GACnB,gBAAgB,GAChB,qBAAqB,GACrB,qBAAqB,GACrB,eAAe,GACf,oBAAoB,GACpB,qBAAqB,CAAC;AAE1B,MAAM,MAAM,YAAY,GACpB,qBAAqB,GACrB,uBAAuB,GACvB,uBAAuB,GACvB,yBAAyB,GACzB,8BAA8B,GAC9B,yBAAyB,GACzB,8BAA8B,CAAC;AAEnC,MAAM,MAAM,cAAc,GAAG,6BAA6B,CAAC;AAE3D,MAAM,MAAM,cAAc,GAAG,yBAAyB,GAAG,2BAA2B,CAAC;AAErF,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,GAAG,YAAY,GAAG,YAAY,GAAG,cAAc,GAAG,cAAc,GAAG,SAAS,CAAC;AAElG,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,IAAI,QAAQ,IAAI,MAAM,CAErB;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;IACjC,OAAO,CAAC,SAAS,CAAC,CAA0B;gBAG1C,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;IAsHpC,OAAO,CAAC,WAAW;IAuGnB,OAAO,CAAC,0BAA0B;IA6BlC,OAAO,CAAC,uBAAuB;CAQhC"}