@langchain/classic 1.0.23 → 1.0.24-dev-1773698445534

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_chain.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"violation_of_expectations_chain.cjs","names":["BaseChain","JsonOutputFunctionsParser","StringOutputParser","HumanMessage","PREDICT_NEXT_USER_MESSAGE_FUNCTION","PREDICT_NEXT_USER_MESSAGE_PROMPT","PREDICTION_VIOLATIONS_FUNCTION","PREDICTION_VIOLATIONS_PROMPT","GENERATE_REVISED_PREDICTION_PROMPT","GENERATE_FACTS_PROMPT"],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_chain.ts"],"sourcesContent":["import type { BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { ChatOpenAI } from \"@langchain/openai\";\nimport {\n  BaseMessage,\n  HumanMessage,\n  isBaseMessage,\n} from \"@langchain/core/messages\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport { StringOutputParser } from \"@langchain/core/output_parsers\";\nimport { CallbackManagerForChainRun } from \"@langchain/core/callbacks/manager\";\nimport { JsonOutputFunctionsParser } from \"../../../output_parsers/openai_functions.js\";\nimport { BaseChain, ChainInputs } from \"../../../chains/base.js\";\nimport {\n  GetPredictionViolationsResponse,\n  MessageChunkResult,\n  PREDICTION_VIOLATIONS_FUNCTION,\n  PREDICT_NEXT_USER_MESSAGE_FUNCTION,\n  PredictNextUserMessageResponse,\n} from \"./types.js\";\nimport {\n  GENERATE_FACTS_PROMPT,\n  GENERATE_REVISED_PREDICTION_PROMPT,\n  PREDICTION_VIOLATIONS_PROMPT,\n  PREDICT_NEXT_USER_MESSAGE_PROMPT,\n} from \"./violation_of_expectations_prompt.js\";\n\n/**\n * Interface for the input parameters of the ViolationOfExpectationsChain class.\n */\nexport interface ViolationOfExpectationsChainInput extends ChainInputs {\n  /**\n   * The retriever to use for retrieving stored\n   * thoughts and insights.\n   */\n  retriever: BaseRetrieverInterface;\n  /**\n   * The LLM to use\n   */\n  llm: ChatOpenAI;\n}\n\n/**\n * Chain that generates key insights/facts of a user based on a\n * a chat conversation with an AI.\n */\nexport class ViolationOfExpectationsChain\n  extends BaseChain\n  implements ViolationOfExpectationsChainInput\n{\n  static lc_name() {\n    return \"ViolationOfExpectationsChain\";\n  }\n\n  _chainType(): string {\n    return \"violation_of_expectation_chain\";\n  }\n\n  chatHistoryKey = \"chat_history\";\n\n  thoughtsKey = \"thoughts\";\n\n  get inputKeys() {\n    return [this.chatHistoryKey];\n  }\n\n  get outputKeys() {\n    return [this.thoughtsKey];\n  }\n\n  retriever: BaseRetrieverInterface;\n\n  llm: ChatOpenAI;\n\n  jsonOutputParser: JsonOutputFunctionsParser;\n\n  stringOutputParser: StringOutputParser;\n\n  constructor(fields: ViolationOfExpectationsChainInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llm = fields.llm;\n    this.jsonOutputParser = new JsonOutputFunctionsParser();\n    this.stringOutputParser = new StringOutputParser();\n  }\n\n  getChatHistoryString(chatHistory: BaseMessage[]): string {\n    return chatHistory\n      .map((chatMessage) => {\n        if (chatMessage._getType() === \"human\") {\n          return `Human: ${chatMessage.content}`;\n        } else if (chatMessage._getType() === \"ai\") {\n          return `AI: ${chatMessage.content}`;\n        } else {\n          return `${chatMessage.content}`;\n        }\n      })\n      .join(\"\\n\");\n  }\n\n  removeDuplicateStrings(strings: Array<string>): Array<string> {\n    return [...new Set(strings)];\n  }\n\n  /**\n   * This method breaks down the chat history into chunks of messages.\n   * Each chunk consists of a sequence of messages ending with an AI message and the subsequent user response, if any.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history to be chunked.\n   *\n   * @returns {MessageChunkResult[]} An array of message chunks. Each chunk includes a sequence of messages and the subsequent user response.\n   *\n   * @description\n   * The method iterates over the chat history and pushes each message into a temporary array.\n   * When it encounters an AI message, it checks for a subsequent user message.\n   * If a user message is found, it is considered as the user response to the AI message.\n   * If no user message is found after the AI message, the user response is undefined.\n   * The method then pushes the chunk (sequence of messages and user response) into the result array.\n   * This process continues until all messages in the chat history have been processed.\n   */\n  chunkMessagesByAIResponse(chatHistory: BaseMessage[]): MessageChunkResult[] {\n    const newArray: MessageChunkResult[] = [];\n    const tempArray: BaseMessage[] = [];\n\n    chatHistory.forEach((item, index) => {\n      tempArray.push(item);\n      if (item._getType() === \"ai\") {\n        let userResponse: BaseMessage | undefined = chatHistory[index + 1];\n        if (!userResponse || userResponse._getType() !== \"human\") {\n          userResponse = undefined;\n        }\n\n        newArray.push({\n          chunkedMessages: tempArray,\n          userResponse: userResponse\n            ? new HumanMessage(userResponse)\n            : undefined,\n        });\n      }\n    });\n\n    return newArray;\n  }\n\n  /**\n   * This method processes a chat history to generate insights about the user.\n   *\n   * @param {ChainValues} values - The input values for the chain. It should contain a key for chat history.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<ChainValues>} A promise that resolves to a list of insights about the user.\n   *\n   * @throws {Error} If the chat history key is not found in the input values or if the chat history is not an array of BaseMessages.\n   *\n   * @description\n   * The method performs the following steps:\n   * 1. Checks if the chat history key is present in the input values and if the chat history is an array of BaseMessages.\n   * 2. Breaks the chat history into chunks of messages.\n   * 3. For each chunk, it generates an initial prediction for the user's next message.\n   * 4. For each prediction, it generates insights and prediction violations, and regenerates the prediction based on the violations.\n   * 5. For each set of messages, it generates a fact/insight about the user.\n   * The method returns a list of these insights.\n   */\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    if (!(this.chatHistoryKey in values)) {\n      throw new Error(`Chat history key ${this.chatHistoryKey} not found`);\n    }\n\n    const chatHistory: unknown[] = values[this.chatHistoryKey];\n\n    const isEveryMessageBaseMessage = chatHistory.every((message) =>\n      isBaseMessage(message)\n    );\n    if (!isEveryMessageBaseMessage) {\n      throw new Error(\"Chat history must be an array of BaseMessages\");\n    }\n\n    const messageChunks = this.chunkMessagesByAIResponse(\n      chatHistory as BaseMessage[]\n    );\n\n    // Generate the initial prediction for every user message.\n    const userPredictions = await Promise.all(\n      messageChunks.map(async (chatHistoryChunk) => ({\n        userPredictions: await this.predictNextUserMessage(\n          chatHistoryChunk.chunkedMessages\n        ),\n        userResponse: chatHistoryChunk.userResponse,\n        runManager,\n      }))\n    );\n\n    // Generate insights, and prediction violations for every user message.\n    // This call also regenerates the prediction based on the violations.\n    const predictionViolations = await Promise.all(\n      userPredictions.map((prediction) =>\n        this.getPredictionViolations({\n          userPredictions: prediction.userPredictions,\n          userResponse: prediction.userResponse,\n          runManager,\n        })\n      )\n    );\n\n    // Generate a fact/insight about the user for every set of messages.\n    const insights = await Promise.all(\n      predictionViolations.map((violation) =>\n        this.generateFacts({\n          userResponse: violation.userResponse,\n          predictions: {\n            revisedPrediction: violation.revisedPrediction,\n            explainedPredictionErrors: violation.explainedPredictionErrors,\n          },\n        })\n      )\n    );\n\n    return {\n      insights,\n    };\n  }\n\n  /**\n   * This method predicts the next user message based on the chat history.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history based on which the next user message is predicted.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<PredictNextUserMessageResponse>} A promise that resolves to the predicted next user message, the user state, and any insights.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'userState', 'predictedUserMessage', and 'insights'.\n   */\n  private async predictNextUserMessage(\n    chatHistory: BaseMessage[],\n    runManager?: CallbackManagerForChainRun\n  ): Promise<PredictNextUserMessageResponse> {\n    const messageString = this.getChatHistoryString(chatHistory);\n\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICT_NEXT_USER_MESSAGE_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICT_NEXT_USER_MESSAGE_FUNCTION.name },\n      });\n\n    const chain = PREDICT_NEXT_USER_MESSAGE_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    const res = await chain.invoke(\n      {\n        chat_history: messageString,\n      },\n      runManager?.getChild(\"prediction\")\n    );\n\n    if (\n      !(\n        \"userState\" in res &&\n        \"predictedUserMessage\" in res &&\n        \"insights\" in res\n      )\n    ) {\n      throw new Error(`Invalid response from LLM: ${JSON.stringify(res)}`);\n    }\n\n    const predictionResponse = res as PredictNextUserMessageResponse;\n\n    // Query the retriever for relevant insights. Use the generates insights as a query.\n    const retrievedDocs = await this.retrieveRelevantInsights(\n      predictionResponse.insights\n    );\n    const relevantDocs = this.removeDuplicateStrings([\n      ...predictionResponse.insights,\n      ...retrievedDocs,\n    ]);\n\n    return {\n      ...predictionResponse,\n      insights: relevantDocs,\n    };\n  }\n\n  /**\n   * Retrieves relevant insights based on the provided insights.\n   *\n   * @param {Array<string>} insights - An array of insights to be used for retrieving relevant documents.\n   *\n   * @returns {Promise<Array<string>>} A promise that resolves to an array of relevant insights content.\n   */\n  private async retrieveRelevantInsights(\n    insights: Array<string>\n  ): Promise<Array<string>> {\n    // Only extract the first relevant doc from the retriever. We don't need more than one.\n    const relevantInsightsDocuments = await Promise.all(\n      insights.map(async (insight) => {\n        const relevantInsight = await this.retriever.invoke(insight);\n        return relevantInsight[0];\n      })\n    );\n\n    const relevantInsightsContent = relevantInsightsDocuments.map(\n      (document) => document.pageContent\n    );\n\n    return relevantInsightsContent;\n  }\n\n  /**\n   * This method generates prediction violations based on the predicted and actual user responses.\n   * It also generates a revised prediction based on the identified violations.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {PredictNextUserMessageResponse} params.userPredictions - The predicted user message, user state, and insights.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<{ userResponse: BaseMessage | undefined; revisedPrediction: string; explainedPredictionErrors: Array<string>; }>} A promise that resolves to an object containing the actual user response, the revised prediction, and the explained prediction errors.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'violationExplanation', 'explainedPredictionErrors', and 'accuratePrediction'.\n   */\n  private async getPredictionViolations({\n    userPredictions,\n    userResponse,\n    runManager,\n  }: {\n    userPredictions: PredictNextUserMessageResponse;\n    userResponse?: BaseMessage;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<GetPredictionViolationsResponse> {\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICTION_VIOLATIONS_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICTION_VIOLATIONS_FUNCTION.name },\n      });\n\n    const chain = PREDICTION_VIOLATIONS_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = (await chain.invoke(\n      {\n        predicted_output: userPredictions.predictedUserMessage,\n        actual_output: userResponse?.content ?? \"\",\n        user_insights: userPredictions.insights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_violations\")\n    )) as Awaited<{\n      violationExplanation: string;\n      explainedPredictionErrors: Array<string>;\n      accuratePrediction: boolean;\n    }>;\n\n    // Generate a revised prediction based on violations.\n    const revisedPrediction = await this.generateRevisedPrediction({\n      originalPrediction: userPredictions.predictedUserMessage,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n      userInsights: userPredictions.insights,\n      runManager,\n    });\n\n    return {\n      userResponse,\n      revisedPrediction,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n    };\n  }\n\n  /**\n   * This method generates a revised prediction based on the original prediction, explained prediction errors, and user insights.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {string} params.originalPrediction - The original prediction made by the model.\n   * @param {Array<string>} params.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {Array<string>} params.userInsights - An array of insights about the user.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a revised prediction.\n   */\n  private async generateRevisedPrediction({\n    originalPrediction,\n    explainedPredictionErrors,\n    userInsights,\n    runManager,\n  }: {\n    originalPrediction: string;\n    explainedPredictionErrors: Array<string>;\n    userInsights: Array<string>;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const revisedPredictionChain = GENERATE_REVISED_PREDICTION_PROMPT.pipe(\n      this.llm\n    ).pipe(this.stringOutputParser);\n\n    const revisedPredictionRes = await revisedPredictionChain.invoke(\n      {\n        prediction: originalPrediction,\n        explained_prediction_errors: explainedPredictionErrors.join(\"\\n\"),\n        user_insights: userInsights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_revision\")\n    );\n\n    return revisedPredictionRes;\n  }\n\n  /**\n   * This method generates facts or insights about the user based on the revised prediction, explained prediction errors, and the user's response.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {Object} params.predictions - The revised prediction and explained prediction errors.\n   * @param {string} params.predictions.revisedPrediction - The revised prediction made by the model.\n   * @param {Array<string>} params.predictions.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a string containing the generated facts or insights about the user.\n   */\n  private async generateFacts({\n    userResponse,\n    predictions,\n    runManager,\n  }: {\n    userResponse?: BaseMessage;\n    /**\n     * Optional if the prediction was accurate.\n     */\n    predictions: {\n      revisedPrediction: string;\n      explainedPredictionErrors: Array<string>;\n    };\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const chain = GENERATE_FACTS_PROMPT.pipe(this.llm).pipe(\n      this.stringOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = await chain.invoke(\n      {\n        prediction_violations: predictions.explainedPredictionErrors.join(\"\\n\"),\n        prediction: predictions.revisedPrediction,\n        user_message: userResponse?.content ?? \"\",\n      },\n      runManager?.getChild(\"generate_facts\")\n    );\n\n    return res;\n  }\n\n  /**\n   * Static method that creates a ViolationOfExpectationsChain instance from a\n   * ChatOpenAI and retriever. It also accepts optional options\n   * to customize the chain.\n   *\n   * @param llm The ChatOpenAI instance.\n   * @param retriever The retriever used for similarity search.\n   * @param options Optional options to customize the chain.\n   *\n   * @returns A new instance of ViolationOfExpectationsChain.\n   */\n  static fromLLM(\n    llm: ChatOpenAI,\n    retriever: BaseRetrieverInterface,\n    options?: Partial<\n      Omit<ViolationOfExpectationsChainInput, \"llm\" | \"retriever\">\n    >\n  ): ViolationOfExpectationsChain {\n    return new this({\n      retriever,\n      llm,\n      ...options,\n    });\n  }\n}\n"],"mappings":";;;;;;;;;;;;;AA6CA,IAAa,+BAAb,cACUA,uBAEV;CACE,OAAO,UAAU;AACf,SAAO;;CAGT,aAAqB;AACnB,SAAO;;CAGT,iBAAiB;CAEjB,cAAc;CAEd,IAAI,YAAY;AACd,SAAO,CAAC,KAAK,eAAe;;CAG9B,IAAI,aAAa;AACf,SAAO,CAAC,KAAK,YAAY;;CAG3B;CAEA;CAEA;CAEA;CAEA,YAAY,QAA2C;AACrD,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,OAAO;AAClB,OAAK,mBAAmB,IAAIC,oDAA2B;AACvD,OAAK,qBAAqB,IAAIC,mDAAoB;;CAGpD,qBAAqB,aAAoC;AACvD,SAAO,YACJ,KAAK,gBAAgB;AACpB,OAAI,YAAY,UAAU,KAAK,QAC7B,QAAO,UAAU,YAAY;YACpB,YAAY,UAAU,KAAK,KACpC,QAAO,OAAO,YAAY;OAE1B,QAAO,GAAG,YAAY;IAExB,CACD,KAAK,KAAK;;CAGf,uBAAuB,SAAuC;AAC5D,SAAO,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC;;;;;;;;;;;;;;;;;;CAmB9B,0BAA0B,aAAkD;EAC1E,MAAM,WAAiC,EAAE;EACzC,MAAM,YAA2B,EAAE;AAEnC,cAAY,SAAS,MAAM,UAAU;AACnC,aAAU,KAAK,KAAK;AACpB,OAAI,KAAK,UAAU,KAAK,MAAM;IAC5B,IAAI,eAAwC,YAAY,QAAQ;AAChE,QAAI,CAAC,gBAAgB,aAAa,UAAU,KAAK,QAC/C,gBAAe;AAGjB,aAAS,KAAK;KACZ,iBAAiB;KACjB,cAAc,eACV,IAAIC,sCAAa,aAAa,GAC9B;KACL,CAAC;;IAEJ;AAEF,SAAO;;;;;;;;;;;;;;;;;;;;;CAsBT,MAAM,MACJ,QACA,YACsB;AACtB,MAAI,EAAE,KAAK,kBAAkB,QAC3B,OAAM,IAAI,MAAM,oBAAoB,KAAK,eAAe,YAAY;EAGtE,MAAM,cAAyB,OAAO,KAAK;AAK3C,MAAI,CAH8B,YAAY,OAAO,wDACrC,QAAQ,CACvB,CAEC,OAAM,IAAI,MAAM,gDAAgD;EAGlE,MAAM,gBAAgB,KAAK,0BACzB,YACD;EAGD,MAAM,kBAAkB,MAAM,QAAQ,IACpC,cAAc,IAAI,OAAO,sBAAsB;GAC7C,iBAAiB,MAAM,KAAK,uBAC1B,iBAAiB,gBAClB;GACD,cAAc,iBAAiB;GAC/B;GACD,EAAE,CACJ;EAID,MAAM,uBAAuB,MAAM,QAAQ,IACzC,gBAAgB,KAAK,eACnB,KAAK,wBAAwB;GAC3B,iBAAiB,WAAW;GAC5B,cAAc,WAAW;GACzB;GACD,CAAC,CACH,CACF;AAeD,SAAO,EACL,UAbe,MAAM,QAAQ,IAC7B,qBAAqB,KAAK,cACxB,KAAK,cAAc;GACjB,cAAc,UAAU;GACxB,aAAa;IACX,mBAAmB,UAAU;IAC7B,2BAA2B,UAAU;IACtC;GACF,CAAC,CACH,CACF,EAIA;;;;;;;;;;;;CAaH,MAAc,uBACZ,aACA,YACyC;EACzC,MAAM,gBAAgB,KAAK,qBAAqB,YAAY;EAE5D,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAACC,iDAAmC,CAAC,CAC/C,WAAW,EACV,eAAe,EAAE,MAAMA,iDAAmC,MAAM,EACjE,CAAC;EAMJ,MAAM,MAAM,MAJEC,0EAAiC,KAAK,iBAAiB,CAAC,KACpE,KAAK,iBACN,CAEuB,OACtB,EACE,cAAc,eACf,EACD,YAAY,SAAS,aAAa,CACnC;AAED,MACE,EACE,eAAe,OACf,0BAA0B,OAC1B,cAAc,KAGhB,OAAM,IAAI,MAAM,8BAA8B,KAAK,UAAU,IAAI,GAAG;EAGtE,MAAM,qBAAqB;EAG3B,MAAM,gBAAgB,MAAM,KAAK,yBAC/B,mBAAmB,SACpB;EACD,MAAM,eAAe,KAAK,uBAAuB,CAC/C,GAAG,mBAAmB,UACtB,GAAG,cACJ,CAAC;AAEF,SAAO;GACL,GAAG;GACH,UAAU;GACX;;;;;;;;;CAUH,MAAc,yBACZ,UACwB;AAaxB,UAXkC,MAAM,QAAQ,IAC9C,SAAS,IAAI,OAAO,YAAY;AAE9B,WADwB,MAAM,KAAK,UAAU,OAAO,QAAQ,EACrC;IACvB,CACH,EAEyD,KACvD,aAAa,SAAS,YACxB;;;;;;;;;;;;;;;CAkBH,MAAc,wBAAwB,EACpC,iBACA,cACA,cAK2C;EAC3C,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAACC,6CAA+B,CAAC,CAC3C,WAAW,EACV,eAAe,EAAE,MAAMA,6CAA+B,MAAM,EAC7D,CAAC;EAEJ,MAAM,QAAQC,sEAA6B,KAAK,iBAAiB,CAAC,KAChE,KAAK,iBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;EAEzE,MAAM,MAAO,MAAM,MAAM,OACvB;GACE,kBAAkB,gBAAgB;GAClC,eAAe,cAAc,WAAW;GACxC,eAAe,gBAAgB,SAAS,KAAK,KAAK;GACnD,EACD,YAAY,SAAS,wBAAwB,CAC9C;AAcD,SAAO;GACL;GACA,mBATwB,MAAM,KAAK,0BAA0B;IAC7D,oBAAoB,gBAAgB;IACpC,2BAA2B,IAAI;IAC/B,cAAc,gBAAgB;IAC9B;IACD,CAAC;GAKA,2BAA2B,IAAI;GAChC;;;;;;;;;;;;;CAcH,MAAc,0BAA0B,EACtC,oBACA,2BACA,cACA,cAMkB;AAclB,SAT6B,MAJEC,4EAAmC,KAChE,KAAK,IACN,CAAC,KAAK,KAAK,mBAAmB,CAE2B,OACxD;GACE,YAAY;GACZ,6BAA6B,0BAA0B,KAAK,KAAK;GACjE,eAAe,aAAa,KAAK,KAAK;GACvC,EACD,YAAY,SAAS,sBAAsB,CAC5C;;;;;;;;;;;;;;CAiBH,MAAc,cAAc,EAC1B,cACA,aACA,cAWkB;EAClB,MAAM,QAAQC,+DAAsB,KAAK,KAAK,IAAI,CAAC,KACjD,KAAK,mBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;AAWzE,SATY,MAAM,MAAM,OACtB;GACE,uBAAuB,YAAY,0BAA0B,KAAK,KAAK;GACvE,YAAY,YAAY;GACxB,cAAc,cAAc,WAAW;GACxC,EACD,YAAY,SAAS,iBAAiB,CACvC;;;;;;;;;;;;;CAgBH,OAAO,QACL,KACA,WACA,SAG8B;AAC9B,SAAO,IAAI,KAAK;GACd;GACA;GACA,GAAG;GACJ,CAAC"}
+{"version":3,"file":"violation_of_expectations_chain.cjs","names":["BaseChain","JsonOutputFunctionsParser","StringOutputParser","HumanMessage","PREDICT_NEXT_USER_MESSAGE_FUNCTION","PREDICT_NEXT_USER_MESSAGE_PROMPT","PREDICTION_VIOLATIONS_FUNCTION","PREDICTION_VIOLATIONS_PROMPT","GENERATE_REVISED_PREDICTION_PROMPT","GENERATE_FACTS_PROMPT"],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_chain.ts"],"sourcesContent":["import type { BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { ChatOpenAI } from \"@langchain/openai\";\nimport {\n  BaseMessage,\n  HumanMessage,\n  isBaseMessage,\n} from \"@langchain/core/messages\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport { StringOutputParser } from \"@langchain/core/output_parsers\";\nimport { CallbackManagerForChainRun } from \"@langchain/core/callbacks/manager\";\nimport { JsonOutputFunctionsParser } from \"../../../output_parsers/openai_functions.js\";\nimport { BaseChain, ChainInputs } from \"../../../chains/base.js\";\nimport {\n  GetPredictionViolationsResponse,\n  MessageChunkResult,\n  PREDICTION_VIOLATIONS_FUNCTION,\n  PREDICT_NEXT_USER_MESSAGE_FUNCTION,\n  PredictNextUserMessageResponse,\n} from \"./types.js\";\nimport {\n  GENERATE_FACTS_PROMPT,\n  GENERATE_REVISED_PREDICTION_PROMPT,\n  PREDICTION_VIOLATIONS_PROMPT,\n  PREDICT_NEXT_USER_MESSAGE_PROMPT,\n} from \"./violation_of_expectations_prompt.js\";\n\n/**\n * Interface for the input parameters of the ViolationOfExpectationsChain class.\n */\nexport interface ViolationOfExpectationsChainInput extends ChainInputs {\n  /**\n   * The retriever to use for retrieving stored\n   * thoughts and insights.\n   */\n  retriever: BaseRetrieverInterface;\n  /**\n   * The LLM to use\n   */\n  llm: ChatOpenAI;\n}\n\n/**\n * Chain that generates key insights/facts of a user based on a\n * a chat conversation with an AI.\n */\nexport class ViolationOfExpectationsChain\n  extends BaseChain\n  implements ViolationOfExpectationsChainInput\n{\n  static lc_name() {\n    return \"ViolationOfExpectationsChain\";\n  }\n\n  _chainType(): string {\n    return \"violation_of_expectation_chain\";\n  }\n\n  chatHistoryKey = \"chat_history\";\n\n  thoughtsKey = \"thoughts\";\n\n  get inputKeys() {\n    return [this.chatHistoryKey];\n  }\n\n  get outputKeys() {\n    return [this.thoughtsKey];\n  }\n\n  retriever: BaseRetrieverInterface;\n\n  llm: ChatOpenAI;\n\n  jsonOutputParser: JsonOutputFunctionsParser;\n\n  stringOutputParser: StringOutputParser;\n\n  constructor(fields: ViolationOfExpectationsChainInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llm = fields.llm;\n    this.jsonOutputParser = new JsonOutputFunctionsParser();\n    this.stringOutputParser = new StringOutputParser();\n  }\n\n  getChatHistoryString(chatHistory: BaseMessage[]): string {\n    return chatHistory\n      .map((chatMessage) => {\n        if (chatMessage._getType() === \"human\") {\n          return `Human: ${chatMessage.content}`;\n        } else if (chatMessage._getType() === \"ai\") {\n          return `AI: ${chatMessage.content}`;\n        } else {\n          return `${chatMessage.content}`;\n        }\n      })\n      .join(\"\\n\");\n  }\n\n  removeDuplicateStrings(strings: Array<string>): Array<string> {\n    return [...new Set(strings)];\n  }\n\n  /**\n   * This method breaks down the chat history into chunks of messages.\n   * Each chunk consists of a sequence of messages ending with an AI message and the subsequent user response, if any.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history to be chunked.\n   *\n   * @returns {MessageChunkResult[]} An array of message chunks. Each chunk includes a sequence of messages and the subsequent user response.\n   *\n   * @description\n   * The method iterates over the chat history and pushes each message into a temporary array.\n   * When it encounters an AI message, it checks for a subsequent user message.\n   * If a user message is found, it is considered as the user response to the AI message.\n   * If no user message is found after the AI message, the user response is undefined.\n   * The method then pushes the chunk (sequence of messages and user response) into the result array.\n   * This process continues until all messages in the chat history have been processed.\n   */\n  chunkMessagesByAIResponse(chatHistory: BaseMessage[]): MessageChunkResult[] {\n    const newArray: MessageChunkResult[] = [];\n    const tempArray: BaseMessage[] = [];\n\n    chatHistory.forEach((item, index) => {\n      tempArray.push(item);\n      if (item._getType() === \"ai\") {\n        let userResponse: BaseMessage | undefined = chatHistory[index + 1];\n        if (!userResponse || userResponse._getType() !== \"human\") {\n          userResponse = undefined;\n        }\n\n        newArray.push({\n          chunkedMessages: tempArray,\n          userResponse: userResponse\n            ? new HumanMessage(userResponse)\n            : undefined,\n        });\n      }\n    });\n\n    return newArray;\n  }\n\n  /**\n   * This method processes a chat history to generate insights about the user.\n   *\n   * @param {ChainValues} values - The input values for the chain. It should contain a key for chat history.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<ChainValues>} A promise that resolves to a list of insights about the user.\n   *\n   * @throws {Error} If the chat history key is not found in the input values or if the chat history is not an array of BaseMessages.\n   *\n   * @description\n   * The method performs the following steps:\n   * 1. Checks if the chat history key is present in the input values and if the chat history is an array of BaseMessages.\n   * 2. Breaks the chat history into chunks of messages.\n   * 3. For each chunk, it generates an initial prediction for the user's next message.\n   * 4. For each prediction, it generates insights and prediction violations, and regenerates the prediction based on the violations.\n   * 5. For each set of messages, it generates a fact/insight about the user.\n   * The method returns a list of these insights.\n   */\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    if (!(this.chatHistoryKey in values)) {\n      throw new Error(`Chat history key ${this.chatHistoryKey} not found`);\n    }\n\n    const chatHistory: unknown[] = values[this.chatHistoryKey];\n\n    const isEveryMessageBaseMessage = chatHistory.every((message) =>\n      isBaseMessage(message)\n    );\n    if (!isEveryMessageBaseMessage) {\n      throw new Error(\"Chat history must be an array of BaseMessages\");\n    }\n\n    const messageChunks = this.chunkMessagesByAIResponse(\n      chatHistory as BaseMessage[]\n    );\n\n    // Generate the initial prediction for every user message.\n    const userPredictions = await Promise.all(\n      messageChunks.map(async (chatHistoryChunk) => ({\n        userPredictions: await this.predictNextUserMessage(\n          chatHistoryChunk.chunkedMessages\n        ),\n        userResponse: chatHistoryChunk.userResponse,\n        runManager,\n      }))\n    );\n\n    // Generate insights, and prediction violations for every user message.\n    // This call also regenerates the prediction based on the violations.\n    const predictionViolations = await Promise.all(\n      userPredictions.map((prediction) =>\n        this.getPredictionViolations({\n          userPredictions: prediction.userPredictions,\n          userResponse: prediction.userResponse,\n          runManager,\n        })\n      )\n    );\n\n    // Generate a fact/insight about the user for every set of messages.\n    const insights = await Promise.all(\n      predictionViolations.map((violation) =>\n        this.generateFacts({\n          userResponse: violation.userResponse,\n          predictions: {\n            revisedPrediction: violation.revisedPrediction,\n            explainedPredictionErrors: violation.explainedPredictionErrors,\n          },\n        })\n      )\n    );\n\n    return {\n      insights,\n    };\n  }\n\n  /**\n   * This method predicts the next user message based on the chat history.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history based on which the next user message is predicted.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<PredictNextUserMessageResponse>} A promise that resolves to the predicted next user message, the user state, and any insights.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'userState', 'predictedUserMessage', and 'insights'.\n   */\n  private async predictNextUserMessage(\n    chatHistory: BaseMessage[],\n    runManager?: CallbackManagerForChainRun\n  ): Promise<PredictNextUserMessageResponse> {\n    const messageString = this.getChatHistoryString(chatHistory);\n\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICT_NEXT_USER_MESSAGE_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICT_NEXT_USER_MESSAGE_FUNCTION.name },\n      });\n\n    const chain = PREDICT_NEXT_USER_MESSAGE_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    const res = await chain.invoke(\n      {\n        chat_history: messageString,\n      },\n      runManager?.getChild(\"prediction\")\n    );\n\n    if (\n      !(\n        \"userState\" in res &&\n        \"predictedUserMessage\" in res &&\n        \"insights\" in res\n      )\n    ) {\n      throw new Error(`Invalid response from LLM: ${JSON.stringify(res)}`);\n    }\n\n    const predictionResponse = res as PredictNextUserMessageResponse;\n\n    // Query the retriever for relevant insights. Use the generates insights as a query.\n    const retrievedDocs = await this.retrieveRelevantInsights(\n      predictionResponse.insights\n    );\n    const relevantDocs = this.removeDuplicateStrings([\n      ...predictionResponse.insights,\n      ...retrievedDocs,\n    ]);\n\n    return {\n      ...predictionResponse,\n      insights: relevantDocs,\n    };\n  }\n\n  /**\n   * Retrieves relevant insights based on the provided insights.\n   *\n   * @param {Array<string>} insights - An array of insights to be used for retrieving relevant documents.\n   *\n   * @returns {Promise<Array<string>>} A promise that resolves to an array of relevant insights content.\n   */\n  private async retrieveRelevantInsights(\n    insights: Array<string>\n  ): Promise<Array<string>> {\n    // Only extract the first relevant doc from the retriever. We don't need more than one.\n    const relevantInsightsDocuments = await Promise.all(\n      insights.map(async (insight) => {\n        const relevantInsight = await this.retriever.invoke(insight);\n        return relevantInsight[0];\n      })\n    );\n\n    const relevantInsightsContent = relevantInsightsDocuments.map(\n      (document) => document.pageContent\n    );\n\n    return relevantInsightsContent;\n  }\n\n  /**\n   * This method generates prediction violations based on the predicted and actual user responses.\n   * It also generates a revised prediction based on the identified violations.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {PredictNextUserMessageResponse} params.userPredictions - The predicted user message, user state, and insights.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<{ userResponse: BaseMessage | undefined; revisedPrediction: string; explainedPredictionErrors: Array<string>; }>} A promise that resolves to an object containing the actual user response, the revised prediction, and the explained prediction errors.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'violationExplanation', 'explainedPredictionErrors', and 'accuratePrediction'.\n   */\n  private async getPredictionViolations({\n    userPredictions,\n    userResponse,\n    runManager,\n  }: {\n    userPredictions: PredictNextUserMessageResponse;\n    userResponse?: BaseMessage;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<GetPredictionViolationsResponse> {\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICTION_VIOLATIONS_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICTION_VIOLATIONS_FUNCTION.name },\n      });\n\n    const chain = PREDICTION_VIOLATIONS_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = (await chain.invoke(\n      {\n        predicted_output: userPredictions.predictedUserMessage,\n        actual_output: userResponse?.content ?? \"\",\n        user_insights: userPredictions.insights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_violations\")\n    )) as Awaited<{\n      violationExplanation: string;\n      explainedPredictionErrors: Array<string>;\n      accuratePrediction: boolean;\n    }>;\n\n    // Generate a revised prediction based on violations.\n    const revisedPrediction = await this.generateRevisedPrediction({\n      originalPrediction: userPredictions.predictedUserMessage,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n      userInsights: userPredictions.insights,\n      runManager,\n    });\n\n    return {\n      userResponse,\n      revisedPrediction,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n    };\n  }\n\n  /**\n   * This method generates a revised prediction based on the original prediction, explained prediction errors, and user insights.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {string} params.originalPrediction - The original prediction made by the model.\n   * @param {Array<string>} params.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {Array<string>} params.userInsights - An array of insights about the user.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a revised prediction.\n   */\n  private async generateRevisedPrediction({\n    originalPrediction,\n    explainedPredictionErrors,\n    userInsights,\n    runManager,\n  }: {\n    originalPrediction: string;\n    explainedPredictionErrors: Array<string>;\n    userInsights: Array<string>;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const revisedPredictionChain = GENERATE_REVISED_PREDICTION_PROMPT.pipe(\n      this.llm\n    ).pipe(this.stringOutputParser);\n\n    const revisedPredictionRes = await revisedPredictionChain.invoke(\n      {\n        prediction: originalPrediction,\n        explained_prediction_errors: explainedPredictionErrors.join(\"\\n\"),\n        user_insights: userInsights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_revision\")\n    );\n\n    return revisedPredictionRes;\n  }\n\n  /**\n   * This method generates facts or insights about the user based on the revised prediction, explained prediction errors, and the user's response.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {Object} params.predictions - The revised prediction and explained prediction errors.\n   * @param {string} params.predictions.revisedPrediction - The revised prediction made by the model.\n   * @param {Array<string>} params.predictions.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a string containing the generated facts or insights about the user.\n   */\n  private async generateFacts({\n    userResponse,\n    predictions,\n    runManager,\n  }: {\n    userResponse?: BaseMessage;\n    /**\n     * Optional if the prediction was accurate.\n     */\n    predictions: {\n      revisedPrediction: string;\n      explainedPredictionErrors: Array<string>;\n    };\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const chain = GENERATE_FACTS_PROMPT.pipe(this.llm).pipe(\n      this.stringOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = await chain.invoke(\n      {\n        prediction_violations: predictions.explainedPredictionErrors.join(\"\\n\"),\n        prediction: predictions.revisedPrediction,\n        user_message: userResponse?.content ?? \"\",\n      },\n      runManager?.getChild(\"generate_facts\")\n    );\n\n    return res;\n  }\n\n  /**\n   * Static method that creates a ViolationOfExpectationsChain instance from a\n   * ChatOpenAI and retriever. It also accepts optional options\n   * to customize the chain.\n   *\n   * @param llm The ChatOpenAI instance.\n   * @param retriever The retriever used for similarity search.\n   * @param options Optional options to customize the chain.\n   *\n   * @returns A new instance of ViolationOfExpectationsChain.\n   */\n  static fromLLM(\n    llm: ChatOpenAI,\n    retriever: BaseRetrieverInterface,\n    options?: Partial<\n      Omit<ViolationOfExpectationsChainInput, \"llm\" | \"retriever\">\n    >\n  ): ViolationOfExpectationsChain {\n    return new this({\n      retriever,\n      llm,\n      ...options,\n    });\n  }\n}\n"],"mappings":";;;;;;;;;;;;AA6CA,IAAa,+BAAb,cACUA,aAAAA,UAEV;CACE,OAAO,UAAU;AACf,SAAO;;CAGT,aAAqB;AACnB,SAAO;;CAGT,iBAAiB;CAEjB,cAAc;CAEd,IAAI,YAAY;AACd,SAAO,CAAC,KAAK,eAAe;;CAG9B,IAAI,aAAa;AACf,SAAO,CAAC,KAAK,YAAY;;CAG3B;CAEA;CAEA;CAEA;CAEA,YAAY,QAA2C;AACrD,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,OAAO;AAClB,OAAK,mBAAmB,IAAIC,yBAAAA,2BAA2B;AACvD,OAAK,qBAAqB,IAAIC,+BAAAA,oBAAoB;;CAGpD,qBAAqB,aAAoC;AACvD,SAAO,YACJ,KAAK,gBAAgB;AACpB,OAAI,YAAY,UAAU,KAAK,QAC7B,QAAO,UAAU,YAAY;YACpB,YAAY,UAAU,KAAK,KACpC,QAAO,OAAO,YAAY;OAE1B,QAAO,GAAG,YAAY;IAExB,CACD,KAAK,KAAK;;CAGf,uBAAuB,SAAuC;AAC5D,SAAO,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC;;;;;;;;;;;;;;;;;;CAmB9B,0BAA0B,aAAkD;EAC1E,MAAM,WAAiC,EAAE;EACzC,MAAM,YAA2B,EAAE;AAEnC,cAAY,SAAS,MAAM,UAAU;AACnC,aAAU,KAAK,KAAK;AACpB,OAAI,KAAK,UAAU,KAAK,MAAM;IAC5B,IAAI,eAAwC,YAAY,QAAQ;AAChE,QAAI,CAAC,gBAAgB,aAAa,UAAU,KAAK,QAC/C,gBAAe,KAAA;AAGjB,aAAS,KAAK;KACZ,iBAAiB;KACjB,cAAc,eACV,IAAIC,yBAAAA,aAAa,aAAa,GAC9B,KAAA;KACL,CAAC;;IAEJ;AAEF,SAAO;;;;;;;;;;;;;;;;;;;;;CAsBT,MAAM,MACJ,QACA,YACsB;AACtB,MAAI,EAAE,KAAK,kBAAkB,QAC3B,OAAM,IAAI,MAAM,oBAAoB,KAAK,eAAe,YAAY;EAGtE,MAAM,cAAyB,OAAO,KAAK;AAK3C,MAAI,CAH8B,YAAY,OAAO,aAAA,GAAA,yBAAA,eACrC,QAAQ,CACvB,CAEC,OAAM,IAAI,MAAM,gDAAgD;EAGlE,MAAM,gBAAgB,KAAK,0BACzB,YACD;EAGD,MAAM,kBAAkB,MAAM,QAAQ,IACpC,cAAc,IAAI,OAAO,sBAAsB;GAC7C,iBAAiB,MAAM,KAAK,uBAC1B,iBAAiB,gBAClB;GACD,cAAc,iBAAiB;GAC/B;GACD,EAAE,CACJ;EAID,MAAM,uBAAuB,MAAM,QAAQ,IACzC,gBAAgB,KAAK,eACnB,KAAK,wBAAwB;GAC3B,iBAAiB,WAAW;GAC5B,cAAc,WAAW;GACzB;GACD,CAAC,CACH,CACF;AAeD,SAAO,EACL,UAbe,MAAM,QAAQ,IAC7B,qBAAqB,KAAK,cACxB,KAAK,cAAc;GACjB,cAAc,UAAU;GACxB,aAAa;IACX,mBAAmB,UAAU;IAC7B,2BAA2B,UAAU;IACtC;GACF,CAAC,CACH,CACF,EAIA;;;;;;;;;;;;CAaH,MAAc,uBACZ,aACA,YACyC;EACzC,MAAM,gBAAgB,KAAK,qBAAqB,YAAY;EAE5D,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAACC,cAAAA,mCAAmC,CAAC,CAC/C,WAAW,EACV,eAAe,EAAE,MAAMA,cAAAA,mCAAmC,MAAM,EACjE,CAAC;EAMJ,MAAM,MAAM,MAJEC,yCAAAA,iCAAiC,KAAK,iBAAiB,CAAC,KACpE,KAAK,iBACN,CAEuB,OACtB,EACE,cAAc,eACf,EACD,YAAY,SAAS,aAAa,CACnC;AAED,MACE,EACE,eAAe,OACf,0BAA0B,OAC1B,cAAc,KAGhB,OAAM,IAAI,MAAM,8BAA8B,KAAK,UAAU,IAAI,GAAG;EAGtE,MAAM,qBAAqB;EAG3B,MAAM,gBAAgB,MAAM,KAAK,yBAC/B,mBAAmB,SACpB;EACD,MAAM,eAAe,KAAK,uBAAuB,CAC/C,GAAG,mBAAmB,UACtB,GAAG,cACJ,CAAC;AAEF,SAAO;GACL,GAAG;GACH,UAAU;GACX;;;;;;;;;CAUH,MAAc,yBACZ,UACwB;AAaxB,UAXkC,MAAM,QAAQ,IAC9C,SAAS,IAAI,OAAO,YAAY;AAE9B,WADwB,MAAM,KAAK,UAAU,OAAO,QAAQ,EACrC;IACvB,CACH,EAEyD,KACvD,aAAa,SAAS,YACxB;;;;;;;;;;;;;;;CAkBH,MAAc,wBAAwB,EACpC,iBACA,cACA,cAK2C;EAC3C,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAACC,cAAAA,+BAA+B,CAAC,CAC3C,WAAW,EACV,eAAe,EAAE,MAAMA,cAAAA,+BAA+B,MAAM,EAC7D,CAAC;EAEJ,MAAM,QAAQC,yCAAAA,6BAA6B,KAAK,iBAAiB,CAAC,KAChE,KAAK,iBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;EAEzE,MAAM,MAAO,MAAM,MAAM,OACvB;GACE,kBAAkB,gBAAgB;GAClC,eAAe,cAAc,WAAW;GACxC,eAAe,gBAAgB,SAAS,KAAK,KAAK;GACnD,EACD,YAAY,SAAS,wBAAwB,CAC9C;AAcD,SAAO;GACL;GACA,mBATwB,MAAM,KAAK,0BAA0B;IAC7D,oBAAoB,gBAAgB;IACpC,2BAA2B,IAAI;IAC/B,cAAc,gBAAgB;IAC9B;IACD,CAAC;GAKA,2BAA2B,IAAI;GAChC;;;;;;;;;;;;;CAcH,MAAc,0BAA0B,EACtC,oBACA,2BACA,cACA,cAMkB;AAclB,SAT6B,MAJEC,yCAAAA,mCAAmC,KAChE,KAAK,IACN,CAAC,KAAK,KAAK,mBAAmB,CAE2B,OACxD;GACE,YAAY;GACZ,6BAA6B,0BAA0B,KAAK,KAAK;GACjE,eAAe,aAAa,KAAK,KAAK;GACvC,EACD,YAAY,SAAS,sBAAsB,CAC5C;;;;;;;;;;;;;;CAiBH,MAAc,cAAc,EAC1B,cACA,aACA,cAWkB;EAClB,MAAM,QAAQC,yCAAAA,sBAAsB,KAAK,KAAK,IAAI,CAAC,KACjD,KAAK,mBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;AAWzE,SATY,MAAM,MAAM,OACtB;GACE,uBAAuB,YAAY,0BAA0B,KAAK,KAAK;GACvE,YAAY,YAAY;GACxB,cAAc,cAAc,WAAW;GACxC,EACD,YAAY,SAAS,iBAAiB,CACvC;;;;;;;;;;;;;CAgBH,OAAO,QACL,KACA,WACA,SAG8B;AAC9B,SAAO,IAAI,KAAK;GACd;GACA;GACA,GAAG;GACJ,CAAC"}

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_chain.js

@@ -4,7 +4,6 @@ import { PREDICTION_VIOLATIONS_FUNCTION, PREDICT_NEXT_USER_MESSAGE_FUNCTION } fr
 import { GENERATE_FACTS_PROMPT, GENERATE_REVISED_PREDICTION_PROMPT, PREDICTION_VIOLATIONS_PROMPT, PREDICT_NEXT_USER_MESSAGE_PROMPT } from "./violation_of_expectations_prompt.js";
 import { HumanMessage, isBaseMessage } from "@langchain/core/messages";
 import { StringOutputParser } from "@langchain/core/output_parsers";
-
 //#region src/experimental/chains/violation_of_expectations/violation_of_expectations_chain.ts
 /**
 * Chain that generates key insights/facts of a user based on a
@@ -246,7 +245,7 @@ var ViolationOfExpectationsChain = class extends BaseChain {
 		});
 	}
 };
-
 //#endregion
 export { ViolationOfExpectationsChain };
+
 //# sourceMappingURL=violation_of_expectations_chain.js.map

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_chain.js.map

@@ -1 +1 @@
-{"version":3,"file":"violation_of_expectations_chain.js","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_chain.ts"],"sourcesContent":["import type { BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { ChatOpenAI } from \"@langchain/openai\";\nimport {\n  BaseMessage,\n  HumanMessage,\n  isBaseMessage,\n} from \"@langchain/core/messages\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport { StringOutputParser } from \"@langchain/core/output_parsers\";\nimport { CallbackManagerForChainRun } from \"@langchain/core/callbacks/manager\";\nimport { JsonOutputFunctionsParser } from \"../../../output_parsers/openai_functions.js\";\nimport { BaseChain, ChainInputs } from \"../../../chains/base.js\";\nimport {\n  GetPredictionViolationsResponse,\n  MessageChunkResult,\n  PREDICTION_VIOLATIONS_FUNCTION,\n  PREDICT_NEXT_USER_MESSAGE_FUNCTION,\n  PredictNextUserMessageResponse,\n} from \"./types.js\";\nimport {\n  GENERATE_FACTS_PROMPT,\n  GENERATE_REVISED_PREDICTION_PROMPT,\n  PREDICTION_VIOLATIONS_PROMPT,\n  PREDICT_NEXT_USER_MESSAGE_PROMPT,\n} from \"./violation_of_expectations_prompt.js\";\n\n/**\n * Interface for the input parameters of the ViolationOfExpectationsChain class.\n */\nexport interface ViolationOfExpectationsChainInput extends ChainInputs {\n  /**\n   * The retriever to use for retrieving stored\n   * thoughts and insights.\n   */\n  retriever: BaseRetrieverInterface;\n  /**\n   * The LLM to use\n   */\n  llm: ChatOpenAI;\n}\n\n/**\n * Chain that generates key insights/facts of a user based on a\n * a chat conversation with an AI.\n */\nexport class ViolationOfExpectationsChain\n  extends BaseChain\n  implements ViolationOfExpectationsChainInput\n{\n  static lc_name() {\n    return \"ViolationOfExpectationsChain\";\n  }\n\n  _chainType(): string {\n    return \"violation_of_expectation_chain\";\n  }\n\n  chatHistoryKey = \"chat_history\";\n\n  thoughtsKey = \"thoughts\";\n\n  get inputKeys() {\n    return [this.chatHistoryKey];\n  }\n\n  get outputKeys() {\n    return [this.thoughtsKey];\n  }\n\n  retriever: BaseRetrieverInterface;\n\n  llm: ChatOpenAI;\n\n  jsonOutputParser: JsonOutputFunctionsParser;\n\n  stringOutputParser: StringOutputParser;\n\n  constructor(fields: ViolationOfExpectationsChainInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llm = fields.llm;\n    this.jsonOutputParser = new JsonOutputFunctionsParser();\n    this.stringOutputParser = new StringOutputParser();\n  }\n\n  getChatHistoryString(chatHistory: BaseMessage[]): string {\n    return chatHistory\n      .map((chatMessage) => {\n        if (chatMessage._getType() === \"human\") {\n          return `Human: ${chatMessage.content}`;\n        } else if (chatMessage._getType() === \"ai\") {\n          return `AI: ${chatMessage.content}`;\n        } else {\n          return `${chatMessage.content}`;\n        }\n      })\n      .join(\"\\n\");\n  }\n\n  removeDuplicateStrings(strings: Array<string>): Array<string> {\n    return [...new Set(strings)];\n  }\n\n  /**\n   * This method breaks down the chat history into chunks of messages.\n   * Each chunk consists of a sequence of messages ending with an AI message and the subsequent user response, if any.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history to be chunked.\n   *\n   * @returns {MessageChunkResult[]} An array of message chunks. Each chunk includes a sequence of messages and the subsequent user response.\n   *\n   * @description\n   * The method iterates over the chat history and pushes each message into a temporary array.\n   * When it encounters an AI message, it checks for a subsequent user message.\n   * If a user message is found, it is considered as the user response to the AI message.\n   * If no user message is found after the AI message, the user response is undefined.\n   * The method then pushes the chunk (sequence of messages and user response) into the result array.\n   * This process continues until all messages in the chat history have been processed.\n   */\n  chunkMessagesByAIResponse(chatHistory: BaseMessage[]): MessageChunkResult[] {\n    const newArray: MessageChunkResult[] = [];\n    const tempArray: BaseMessage[] = [];\n\n    chatHistory.forEach((item, index) => {\n      tempArray.push(item);\n      if (item._getType() === \"ai\") {\n        let userResponse: BaseMessage | undefined = chatHistory[index + 1];\n        if (!userResponse || userResponse._getType() !== \"human\") {\n          userResponse = undefined;\n        }\n\n        newArray.push({\n          chunkedMessages: tempArray,\n          userResponse: userResponse\n            ? new HumanMessage(userResponse)\n            : undefined,\n        });\n      }\n    });\n\n    return newArray;\n  }\n\n  /**\n   * This method processes a chat history to generate insights about the user.\n   *\n   * @param {ChainValues} values - The input values for the chain. It should contain a key for chat history.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<ChainValues>} A promise that resolves to a list of insights about the user.\n   *\n   * @throws {Error} If the chat history key is not found in the input values or if the chat history is not an array of BaseMessages.\n   *\n   * @description\n   * The method performs the following steps:\n   * 1. Checks if the chat history key is present in the input values and if the chat history is an array of BaseMessages.\n   * 2. Breaks the chat history into chunks of messages.\n   * 3. For each chunk, it generates an initial prediction for the user's next message.\n   * 4. For each prediction, it generates insights and prediction violations, and regenerates the prediction based on the violations.\n   * 5. For each set of messages, it generates a fact/insight about the user.\n   * The method returns a list of these insights.\n   */\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    if (!(this.chatHistoryKey in values)) {\n      throw new Error(`Chat history key ${this.chatHistoryKey} not found`);\n    }\n\n    const chatHistory: unknown[] = values[this.chatHistoryKey];\n\n    const isEveryMessageBaseMessage = chatHistory.every((message) =>\n      isBaseMessage(message)\n    );\n    if (!isEveryMessageBaseMessage) {\n      throw new Error(\"Chat history must be an array of BaseMessages\");\n    }\n\n    const messageChunks = this.chunkMessagesByAIResponse(\n      chatHistory as BaseMessage[]\n    );\n\n    // Generate the initial prediction for every user message.\n    const userPredictions = await Promise.all(\n      messageChunks.map(async (chatHistoryChunk) => ({\n        userPredictions: await this.predictNextUserMessage(\n          chatHistoryChunk.chunkedMessages\n        ),\n        userResponse: chatHistoryChunk.userResponse,\n        runManager,\n      }))\n    );\n\n    // Generate insights, and prediction violations for every user message.\n    // This call also regenerates the prediction based on the violations.\n    const predictionViolations = await Promise.all(\n      userPredictions.map((prediction) =>\n        this.getPredictionViolations({\n          userPredictions: prediction.userPredictions,\n          userResponse: prediction.userResponse,\n          runManager,\n        })\n      )\n    );\n\n    // Generate a fact/insight about the user for every set of messages.\n    const insights = await Promise.all(\n      predictionViolations.map((violation) =>\n        this.generateFacts({\n          userResponse: violation.userResponse,\n          predictions: {\n            revisedPrediction: violation.revisedPrediction,\n            explainedPredictionErrors: violation.explainedPredictionErrors,\n          },\n        })\n      )\n    );\n\n    return {\n      insights,\n    };\n  }\n\n  /**\n   * This method predicts the next user message based on the chat history.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history based on which the next user message is predicted.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<PredictNextUserMessageResponse>} A promise that resolves to the predicted next user message, the user state, and any insights.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'userState', 'predictedUserMessage', and 'insights'.\n   */\n  private async predictNextUserMessage(\n    chatHistory: BaseMessage[],\n    runManager?: CallbackManagerForChainRun\n  ): Promise<PredictNextUserMessageResponse> {\n    const messageString = this.getChatHistoryString(chatHistory);\n\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICT_NEXT_USER_MESSAGE_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICT_NEXT_USER_MESSAGE_FUNCTION.name },\n      });\n\n    const chain = PREDICT_NEXT_USER_MESSAGE_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    const res = await chain.invoke(\n      {\n        chat_history: messageString,\n      },\n      runManager?.getChild(\"prediction\")\n    );\n\n    if (\n      !(\n        \"userState\" in res &&\n        \"predictedUserMessage\" in res &&\n        \"insights\" in res\n      )\n    ) {\n      throw new Error(`Invalid response from LLM: ${JSON.stringify(res)}`);\n    }\n\n    const predictionResponse = res as PredictNextUserMessageResponse;\n\n    // Query the retriever for relevant insights. Use the generates insights as a query.\n    const retrievedDocs = await this.retrieveRelevantInsights(\n      predictionResponse.insights\n    );\n    const relevantDocs = this.removeDuplicateStrings([\n      ...predictionResponse.insights,\n      ...retrievedDocs,\n    ]);\n\n    return {\n      ...predictionResponse,\n      insights: relevantDocs,\n    };\n  }\n\n  /**\n   * Retrieves relevant insights based on the provided insights.\n   *\n   * @param {Array<string>} insights - An array of insights to be used for retrieving relevant documents.\n   *\n   * @returns {Promise<Array<string>>} A promise that resolves to an array of relevant insights content.\n   */\n  private async retrieveRelevantInsights(\n    insights: Array<string>\n  ): Promise<Array<string>> {\n    // Only extract the first relevant doc from the retriever. We don't need more than one.\n    const relevantInsightsDocuments = await Promise.all(\n      insights.map(async (insight) => {\n        const relevantInsight = await this.retriever.invoke(insight);\n        return relevantInsight[0];\n      })\n    );\n\n    const relevantInsightsContent = relevantInsightsDocuments.map(\n      (document) => document.pageContent\n    );\n\n    return relevantInsightsContent;\n  }\n\n  /**\n   * This method generates prediction violations based on the predicted and actual user responses.\n   * It also generates a revised prediction based on the identified violations.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {PredictNextUserMessageResponse} params.userPredictions - The predicted user message, user state, and insights.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<{ userResponse: BaseMessage | undefined; revisedPrediction: string; explainedPredictionErrors: Array<string>; }>} A promise that resolves to an object containing the actual user response, the revised prediction, and the explained prediction errors.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'violationExplanation', 'explainedPredictionErrors', and 'accuratePrediction'.\n   */\n  private async getPredictionViolations({\n    userPredictions,\n    userResponse,\n    runManager,\n  }: {\n    userPredictions: PredictNextUserMessageResponse;\n    userResponse?: BaseMessage;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<GetPredictionViolationsResponse> {\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICTION_VIOLATIONS_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICTION_VIOLATIONS_FUNCTION.name },\n      });\n\n    const chain = PREDICTION_VIOLATIONS_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = (await chain.invoke(\n      {\n        predicted_output: userPredictions.predictedUserMessage,\n        actual_output: userResponse?.content ?? \"\",\n        user_insights: userPredictions.insights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_violations\")\n    )) as Awaited<{\n      violationExplanation: string;\n      explainedPredictionErrors: Array<string>;\n      accuratePrediction: boolean;\n    }>;\n\n    // Generate a revised prediction based on violations.\n    const revisedPrediction = await this.generateRevisedPrediction({\n      originalPrediction: userPredictions.predictedUserMessage,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n      userInsights: userPredictions.insights,\n      runManager,\n    });\n\n    return {\n      userResponse,\n      revisedPrediction,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n    };\n  }\n\n  /**\n   * This method generates a revised prediction based on the original prediction, explained prediction errors, and user insights.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {string} params.originalPrediction - The original prediction made by the model.\n   * @param {Array<string>} params.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {Array<string>} params.userInsights - An array of insights about the user.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a revised prediction.\n   */\n  private async generateRevisedPrediction({\n    originalPrediction,\n    explainedPredictionErrors,\n    userInsights,\n    runManager,\n  }: {\n    originalPrediction: string;\n    explainedPredictionErrors: Array<string>;\n    userInsights: Array<string>;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const revisedPredictionChain = GENERATE_REVISED_PREDICTION_PROMPT.pipe(\n      this.llm\n    ).pipe(this.stringOutputParser);\n\n    const revisedPredictionRes = await revisedPredictionChain.invoke(\n      {\n        prediction: originalPrediction,\n        explained_prediction_errors: explainedPredictionErrors.join(\"\\n\"),\n        user_insights: userInsights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_revision\")\n    );\n\n    return revisedPredictionRes;\n  }\n\n  /**\n   * This method generates facts or insights about the user based on the revised prediction, explained prediction errors, and the user's response.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {Object} params.predictions - The revised prediction and explained prediction errors.\n   * @param {string} params.predictions.revisedPrediction - The revised prediction made by the model.\n   * @param {Array<string>} params.predictions.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a string containing the generated facts or insights about the user.\n   */\n  private async generateFacts({\n    userResponse,\n    predictions,\n    runManager,\n  }: {\n    userResponse?: BaseMessage;\n    /**\n     * Optional if the prediction was accurate.\n     */\n    predictions: {\n      revisedPrediction: string;\n      explainedPredictionErrors: Array<string>;\n    };\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const chain = GENERATE_FACTS_PROMPT.pipe(this.llm).pipe(\n      this.stringOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = await chain.invoke(\n      {\n        prediction_violations: predictions.explainedPredictionErrors.join(\"\\n\"),\n        prediction: predictions.revisedPrediction,\n        user_message: userResponse?.content ?? \"\",\n      },\n      runManager?.getChild(\"generate_facts\")\n    );\n\n    return res;\n  }\n\n  /**\n   * Static method that creates a ViolationOfExpectationsChain instance from a\n   * ChatOpenAI and retriever. It also accepts optional options\n   * to customize the chain.\n   *\n   * @param llm The ChatOpenAI instance.\n   * @param retriever The retriever used for similarity search.\n   * @param options Optional options to customize the chain.\n   *\n   * @returns A new instance of ViolationOfExpectationsChain.\n   */\n  static fromLLM(\n    llm: ChatOpenAI,\n    retriever: BaseRetrieverInterface,\n    options?: Partial<\n      Omit<ViolationOfExpectationsChainInput, \"llm\" | \"retriever\">\n    >\n  ): ViolationOfExpectationsChain {\n    return new this({\n      retriever,\n      llm,\n      ...options,\n    });\n  }\n}\n"],"mappings":";;;;;;;;;;;;AA6CA,IAAa,+BAAb,cACU,UAEV;CACE,OAAO,UAAU;AACf,SAAO;;CAGT,aAAqB;AACnB,SAAO;;CAGT,iBAAiB;CAEjB,cAAc;CAEd,IAAI,YAAY;AACd,SAAO,CAAC,KAAK,eAAe;;CAG9B,IAAI,aAAa;AACf,SAAO,CAAC,KAAK,YAAY;;CAG3B;CAEA;CAEA;CAEA;CAEA,YAAY,QAA2C;AACrD,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,OAAO;AAClB,OAAK,mBAAmB,IAAI,2BAA2B;AACvD,OAAK,qBAAqB,IAAI,oBAAoB;;CAGpD,qBAAqB,aAAoC;AACvD,SAAO,YACJ,KAAK,gBAAgB;AACpB,OAAI,YAAY,UAAU,KAAK,QAC7B,QAAO,UAAU,YAAY;YACpB,YAAY,UAAU,KAAK,KACpC,QAAO,OAAO,YAAY;OAE1B,QAAO,GAAG,YAAY;IAExB,CACD,KAAK,KAAK;;CAGf,uBAAuB,SAAuC;AAC5D,SAAO,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC;;;;;;;;;;;;;;;;;;CAmB9B,0BAA0B,aAAkD;EAC1E,MAAM,WAAiC,EAAE;EACzC,MAAM,YAA2B,EAAE;AAEnC,cAAY,SAAS,MAAM,UAAU;AACnC,aAAU,KAAK,KAAK;AACpB,OAAI,KAAK,UAAU,KAAK,MAAM;IAC5B,IAAI,eAAwC,YAAY,QAAQ;AAChE,QAAI,CAAC,gBAAgB,aAAa,UAAU,KAAK,QAC/C,gBAAe;AAGjB,aAAS,KAAK;KACZ,iBAAiB;KACjB,cAAc,eACV,IAAI,aAAa,aAAa,GAC9B;KACL,CAAC;;IAEJ;AAEF,SAAO;;;;;;;;;;;;;;;;;;;;;CAsBT,MAAM,MACJ,QACA,YACsB;AACtB,MAAI,EAAE,KAAK,kBAAkB,QAC3B,OAAM,IAAI,MAAM,oBAAoB,KAAK,eAAe,YAAY;EAGtE,MAAM,cAAyB,OAAO,KAAK;AAK3C,MAAI,CAH8B,YAAY,OAAO,YACnD,cAAc,QAAQ,CACvB,CAEC,OAAM,IAAI,MAAM,gDAAgD;EAGlE,MAAM,gBAAgB,KAAK,0BACzB,YACD;EAGD,MAAM,kBAAkB,MAAM,QAAQ,IACpC,cAAc,IAAI,OAAO,sBAAsB;GAC7C,iBAAiB,MAAM,KAAK,uBAC1B,iBAAiB,gBAClB;GACD,cAAc,iBAAiB;GAC/B;GACD,EAAE,CACJ;EAID,MAAM,uBAAuB,MAAM,QAAQ,IACzC,gBAAgB,KAAK,eACnB,KAAK,wBAAwB;GAC3B,iBAAiB,WAAW;GAC5B,cAAc,WAAW;GACzB;GACD,CAAC,CACH,CACF;AAeD,SAAO,EACL,UAbe,MAAM,QAAQ,IAC7B,qBAAqB,KAAK,cACxB,KAAK,cAAc;GACjB,cAAc,UAAU;GACxB,aAAa;IACX,mBAAmB,UAAU;IAC7B,2BAA2B,UAAU;IACtC;GACF,CAAC,CACH,CACF,EAIA;;;;;;;;;;;;CAaH,MAAc,uBACZ,aACA,YACyC;EACzC,MAAM,gBAAgB,KAAK,qBAAqB,YAAY;EAE5D,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAAC,mCAAmC,CAAC,CAC/C,WAAW,EACV,eAAe,EAAE,MAAM,mCAAmC,MAAM,EACjE,CAAC;EAMJ,MAAM,MAAM,MAJE,iCAAiC,KAAK,iBAAiB,CAAC,KACpE,KAAK,iBACN,CAEuB,OACtB,EACE,cAAc,eACf,EACD,YAAY,SAAS,aAAa,CACnC;AAED,MACE,EACE,eAAe,OACf,0BAA0B,OAC1B,cAAc,KAGhB,OAAM,IAAI,MAAM,8BAA8B,KAAK,UAAU,IAAI,GAAG;EAGtE,MAAM,qBAAqB;EAG3B,MAAM,gBAAgB,MAAM,KAAK,yBAC/B,mBAAmB,SACpB;EACD,MAAM,eAAe,KAAK,uBAAuB,CAC/C,GAAG,mBAAmB,UACtB,GAAG,cACJ,CAAC;AAEF,SAAO;GACL,GAAG;GACH,UAAU;GACX;;;;;;;;;CAUH,MAAc,yBACZ,UACwB;AAaxB,UAXkC,MAAM,QAAQ,IAC9C,SAAS,IAAI,OAAO,YAAY;AAE9B,WADwB,MAAM,KAAK,UAAU,OAAO,QAAQ,EACrC;IACvB,CACH,EAEyD,KACvD,aAAa,SAAS,YACxB;;;;;;;;;;;;;;;CAkBH,MAAc,wBAAwB,EACpC,iBACA,cACA,cAK2C;EAC3C,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAAC,+BAA+B,CAAC,CAC3C,WAAW,EACV,eAAe,EAAE,MAAM,+BAA+B,MAAM,EAC7D,CAAC;EAEJ,MAAM,QAAQ,6BAA6B,KAAK,iBAAiB,CAAC,KAChE,KAAK,iBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;EAEzE,MAAM,MAAO,MAAM,MAAM,OACvB;GACE,kBAAkB,gBAAgB;GAClC,eAAe,cAAc,WAAW;GACxC,eAAe,gBAAgB,SAAS,KAAK,KAAK;GACnD,EACD,YAAY,SAAS,wBAAwB,CAC9C;AAcD,SAAO;GACL;GACA,mBATwB,MAAM,KAAK,0BAA0B;IAC7D,oBAAoB,gBAAgB;IACpC,2BAA2B,IAAI;IAC/B,cAAc,gBAAgB;IAC9B;IACD,CAAC;GAKA,2BAA2B,IAAI;GAChC;;;;;;;;;;;;;CAcH,MAAc,0BAA0B,EACtC,oBACA,2BACA,cACA,cAMkB;AAclB,SAT6B,MAJE,mCAAmC,KAChE,KAAK,IACN,CAAC,KAAK,KAAK,mBAAmB,CAE2B,OACxD;GACE,YAAY;GACZ,6BAA6B,0BAA0B,KAAK,KAAK;GACjE,eAAe,aAAa,KAAK,KAAK;GACvC,EACD,YAAY,SAAS,sBAAsB,CAC5C;;;;;;;;;;;;;;CAiBH,MAAc,cAAc,EAC1B,cACA,aACA,cAWkB;EAClB,MAAM,QAAQ,sBAAsB,KAAK,KAAK,IAAI,CAAC,KACjD,KAAK,mBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;AAWzE,SATY,MAAM,MAAM,OACtB;GACE,uBAAuB,YAAY,0BAA0B,KAAK,KAAK;GACvE,YAAY,YAAY;GACxB,cAAc,cAAc,WAAW;GACxC,EACD,YAAY,SAAS,iBAAiB,CACvC;;;;;;;;;;;;;CAgBH,OAAO,QACL,KACA,WACA,SAG8B;AAC9B,SAAO,IAAI,KAAK;GACd;GACA;GACA,GAAG;GACJ,CAAC"}
+{"version":3,"file":"violation_of_expectations_chain.js","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_chain.ts"],"sourcesContent":["import type { BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { ChatOpenAI } from \"@langchain/openai\";\nimport {\n  BaseMessage,\n  HumanMessage,\n  isBaseMessage,\n} from \"@langchain/core/messages\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport { StringOutputParser } from \"@langchain/core/output_parsers\";\nimport { CallbackManagerForChainRun } from \"@langchain/core/callbacks/manager\";\nimport { JsonOutputFunctionsParser } from \"../../../output_parsers/openai_functions.js\";\nimport { BaseChain, ChainInputs } from \"../../../chains/base.js\";\nimport {\n  GetPredictionViolationsResponse,\n  MessageChunkResult,\n  PREDICTION_VIOLATIONS_FUNCTION,\n  PREDICT_NEXT_USER_MESSAGE_FUNCTION,\n  PredictNextUserMessageResponse,\n} from \"./types.js\";\nimport {\n  GENERATE_FACTS_PROMPT,\n  GENERATE_REVISED_PREDICTION_PROMPT,\n  PREDICTION_VIOLATIONS_PROMPT,\n  PREDICT_NEXT_USER_MESSAGE_PROMPT,\n} from \"./violation_of_expectations_prompt.js\";\n\n/**\n * Interface for the input parameters of the ViolationOfExpectationsChain class.\n */\nexport interface ViolationOfExpectationsChainInput extends ChainInputs {\n  /**\n   * The retriever to use for retrieving stored\n   * thoughts and insights.\n   */\n  retriever: BaseRetrieverInterface;\n  /**\n   * The LLM to use\n   */\n  llm: ChatOpenAI;\n}\n\n/**\n * Chain that generates key insights/facts of a user based on a\n * a chat conversation with an AI.\n */\nexport class ViolationOfExpectationsChain\n  extends BaseChain\n  implements ViolationOfExpectationsChainInput\n{\n  static lc_name() {\n    return \"ViolationOfExpectationsChain\";\n  }\n\n  _chainType(): string {\n    return \"violation_of_expectation_chain\";\n  }\n\n  chatHistoryKey = \"chat_history\";\n\n  thoughtsKey = \"thoughts\";\n\n  get inputKeys() {\n    return [this.chatHistoryKey];\n  }\n\n  get outputKeys() {\n    return [this.thoughtsKey];\n  }\n\n  retriever: BaseRetrieverInterface;\n\n  llm: ChatOpenAI;\n\n  jsonOutputParser: JsonOutputFunctionsParser;\n\n  stringOutputParser: StringOutputParser;\n\n  constructor(fields: ViolationOfExpectationsChainInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llm = fields.llm;\n    this.jsonOutputParser = new JsonOutputFunctionsParser();\n    this.stringOutputParser = new StringOutputParser();\n  }\n\n  getChatHistoryString(chatHistory: BaseMessage[]): string {\n    return chatHistory\n      .map((chatMessage) => {\n        if (chatMessage._getType() === \"human\") {\n          return `Human: ${chatMessage.content}`;\n        } else if (chatMessage._getType() === \"ai\") {\n          return `AI: ${chatMessage.content}`;\n        } else {\n          return `${chatMessage.content}`;\n        }\n      })\n      .join(\"\\n\");\n  }\n\n  removeDuplicateStrings(strings: Array<string>): Array<string> {\n    return [...new Set(strings)];\n  }\n\n  /**\n   * This method breaks down the chat history into chunks of messages.\n   * Each chunk consists of a sequence of messages ending with an AI message and the subsequent user response, if any.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history to be chunked.\n   *\n   * @returns {MessageChunkResult[]} An array of message chunks. Each chunk includes a sequence of messages and the subsequent user response.\n   *\n   * @description\n   * The method iterates over the chat history and pushes each message into a temporary array.\n   * When it encounters an AI message, it checks for a subsequent user message.\n   * If a user message is found, it is considered as the user response to the AI message.\n   * If no user message is found after the AI message, the user response is undefined.\n   * The method then pushes the chunk (sequence of messages and user response) into the result array.\n   * This process continues until all messages in the chat history have been processed.\n   */\n  chunkMessagesByAIResponse(chatHistory: BaseMessage[]): MessageChunkResult[] {\n    const newArray: MessageChunkResult[] = [];\n    const tempArray: BaseMessage[] = [];\n\n    chatHistory.forEach((item, index) => {\n      tempArray.push(item);\n      if (item._getType() === \"ai\") {\n        let userResponse: BaseMessage | undefined = chatHistory[index + 1];\n        if (!userResponse || userResponse._getType() !== \"human\") {\n          userResponse = undefined;\n        }\n\n        newArray.push({\n          chunkedMessages: tempArray,\n          userResponse: userResponse\n            ? new HumanMessage(userResponse)\n            : undefined,\n        });\n      }\n    });\n\n    return newArray;\n  }\n\n  /**\n   * This method processes a chat history to generate insights about the user.\n   *\n   * @param {ChainValues} values - The input values for the chain. It should contain a key for chat history.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<ChainValues>} A promise that resolves to a list of insights about the user.\n   *\n   * @throws {Error} If the chat history key is not found in the input values or if the chat history is not an array of BaseMessages.\n   *\n   * @description\n   * The method performs the following steps:\n   * 1. Checks if the chat history key is present in the input values and if the chat history is an array of BaseMessages.\n   * 2. Breaks the chat history into chunks of messages.\n   * 3. For each chunk, it generates an initial prediction for the user's next message.\n   * 4. For each prediction, it generates insights and prediction violations, and regenerates the prediction based on the violations.\n   * 5. For each set of messages, it generates a fact/insight about the user.\n   * The method returns a list of these insights.\n   */\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    if (!(this.chatHistoryKey in values)) {\n      throw new Error(`Chat history key ${this.chatHistoryKey} not found`);\n    }\n\n    const chatHistory: unknown[] = values[this.chatHistoryKey];\n\n    const isEveryMessageBaseMessage = chatHistory.every((message) =>\n      isBaseMessage(message)\n    );\n    if (!isEveryMessageBaseMessage) {\n      throw new Error(\"Chat history must be an array of BaseMessages\");\n    }\n\n    const messageChunks = this.chunkMessagesByAIResponse(\n      chatHistory as BaseMessage[]\n    );\n\n    // Generate the initial prediction for every user message.\n    const userPredictions = await Promise.all(\n      messageChunks.map(async (chatHistoryChunk) => ({\n        userPredictions: await this.predictNextUserMessage(\n          chatHistoryChunk.chunkedMessages\n        ),\n        userResponse: chatHistoryChunk.userResponse,\n        runManager,\n      }))\n    );\n\n    // Generate insights, and prediction violations for every user message.\n    // This call also regenerates the prediction based on the violations.\n    const predictionViolations = await Promise.all(\n      userPredictions.map((prediction) =>\n        this.getPredictionViolations({\n          userPredictions: prediction.userPredictions,\n          userResponse: prediction.userResponse,\n          runManager,\n        })\n      )\n    );\n\n    // Generate a fact/insight about the user for every set of messages.\n    const insights = await Promise.all(\n      predictionViolations.map((violation) =>\n        this.generateFacts({\n          userResponse: violation.userResponse,\n          predictions: {\n            revisedPrediction: violation.revisedPrediction,\n            explainedPredictionErrors: violation.explainedPredictionErrors,\n          },\n        })\n      )\n    );\n\n    return {\n      insights,\n    };\n  }\n\n  /**\n   * This method predicts the next user message based on the chat history.\n   *\n   * @param {BaseMessage[]} chatHistory - The chat history based on which the next user message is predicted.\n   * @param {CallbackManagerForChainRun} [runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<PredictNextUserMessageResponse>} A promise that resolves to the predicted next user message, the user state, and any insights.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'userState', 'predictedUserMessage', and 'insights'.\n   */\n  private async predictNextUserMessage(\n    chatHistory: BaseMessage[],\n    runManager?: CallbackManagerForChainRun\n  ): Promise<PredictNextUserMessageResponse> {\n    const messageString = this.getChatHistoryString(chatHistory);\n\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICT_NEXT_USER_MESSAGE_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICT_NEXT_USER_MESSAGE_FUNCTION.name },\n      });\n\n    const chain = PREDICT_NEXT_USER_MESSAGE_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    const res = await chain.invoke(\n      {\n        chat_history: messageString,\n      },\n      runManager?.getChild(\"prediction\")\n    );\n\n    if (\n      !(\n        \"userState\" in res &&\n        \"predictedUserMessage\" in res &&\n        \"insights\" in res\n      )\n    ) {\n      throw new Error(`Invalid response from LLM: ${JSON.stringify(res)}`);\n    }\n\n    const predictionResponse = res as PredictNextUserMessageResponse;\n\n    // Query the retriever for relevant insights. Use the generates insights as a query.\n    const retrievedDocs = await this.retrieveRelevantInsights(\n      predictionResponse.insights\n    );\n    const relevantDocs = this.removeDuplicateStrings([\n      ...predictionResponse.insights,\n      ...retrievedDocs,\n    ]);\n\n    return {\n      ...predictionResponse,\n      insights: relevantDocs,\n    };\n  }\n\n  /**\n   * Retrieves relevant insights based on the provided insights.\n   *\n   * @param {Array<string>} insights - An array of insights to be used for retrieving relevant documents.\n   *\n   * @returns {Promise<Array<string>>} A promise that resolves to an array of relevant insights content.\n   */\n  private async retrieveRelevantInsights(\n    insights: Array<string>\n  ): Promise<Array<string>> {\n    // Only extract the first relevant doc from the retriever. We don't need more than one.\n    const relevantInsightsDocuments = await Promise.all(\n      insights.map(async (insight) => {\n        const relevantInsight = await this.retriever.invoke(insight);\n        return relevantInsight[0];\n      })\n    );\n\n    const relevantInsightsContent = relevantInsightsDocuments.map(\n      (document) => document.pageContent\n    );\n\n    return relevantInsightsContent;\n  }\n\n  /**\n   * This method generates prediction violations based on the predicted and actual user responses.\n   * It also generates a revised prediction based on the identified violations.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {PredictNextUserMessageResponse} params.userPredictions - The predicted user message, user state, and insights.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<{ userResponse: BaseMessage | undefined; revisedPrediction: string; explainedPredictionErrors: Array<string>; }>} A promise that resolves to an object containing the actual user response, the revised prediction, and the explained prediction errors.\n   *\n   * @throws {Error} If the response from the language model does not contain the expected keys: 'violationExplanation', 'explainedPredictionErrors', and 'accuratePrediction'.\n   */\n  private async getPredictionViolations({\n    userPredictions,\n    userResponse,\n    runManager,\n  }: {\n    userPredictions: PredictNextUserMessageResponse;\n    userResponse?: BaseMessage;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<GetPredictionViolationsResponse> {\n    const llmWithFunctions = this.llm\n      .bindTools([PREDICTION_VIOLATIONS_FUNCTION])\n      .withConfig({\n        function_call: { name: PREDICTION_VIOLATIONS_FUNCTION.name },\n      });\n\n    const chain = PREDICTION_VIOLATIONS_PROMPT.pipe(llmWithFunctions).pipe(\n      this.jsonOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = (await chain.invoke(\n      {\n        predicted_output: userPredictions.predictedUserMessage,\n        actual_output: userResponse?.content ?? \"\",\n        user_insights: userPredictions.insights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_violations\")\n    )) as Awaited<{\n      violationExplanation: string;\n      explainedPredictionErrors: Array<string>;\n      accuratePrediction: boolean;\n    }>;\n\n    // Generate a revised prediction based on violations.\n    const revisedPrediction = await this.generateRevisedPrediction({\n      originalPrediction: userPredictions.predictedUserMessage,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n      userInsights: userPredictions.insights,\n      runManager,\n    });\n\n    return {\n      userResponse,\n      revisedPrediction,\n      explainedPredictionErrors: res.explainedPredictionErrors,\n    };\n  }\n\n  /**\n   * This method generates a revised prediction based on the original prediction, explained prediction errors, and user insights.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {string} params.originalPrediction - The original prediction made by the model.\n   * @param {Array<string>} params.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {Array<string>} params.userInsights - An array of insights about the user.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a revised prediction.\n   */\n  private async generateRevisedPrediction({\n    originalPrediction,\n    explainedPredictionErrors,\n    userInsights,\n    runManager,\n  }: {\n    originalPrediction: string;\n    explainedPredictionErrors: Array<string>;\n    userInsights: Array<string>;\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const revisedPredictionChain = GENERATE_REVISED_PREDICTION_PROMPT.pipe(\n      this.llm\n    ).pipe(this.stringOutputParser);\n\n    const revisedPredictionRes = await revisedPredictionChain.invoke(\n      {\n        prediction: originalPrediction,\n        explained_prediction_errors: explainedPredictionErrors.join(\"\\n\"),\n        user_insights: userInsights.join(\"\\n\"),\n      },\n      runManager?.getChild(\"prediction_revision\")\n    );\n\n    return revisedPredictionRes;\n  }\n\n  /**\n   * This method generates facts or insights about the user based on the revised prediction, explained prediction errors, and the user's response.\n   *\n   * @param {Object} params - The parameters for the method.\n   * @param {BaseMessage} [params.userResponse] - The actual user response.\n   * @param {Object} params.predictions - The revised prediction and explained prediction errors.\n   * @param {string} params.predictions.revisedPrediction - The revised prediction made by the model.\n   * @param {Array<string>} params.predictions.explainedPredictionErrors - An array of explained prediction errors.\n   * @param {CallbackManagerForChainRun} [params.runManager] - Optional callback manager for the chain run.\n   *\n   * @returns {Promise<string>} A promise that resolves to a string containing the generated facts or insights about the user.\n   */\n  private async generateFacts({\n    userResponse,\n    predictions,\n    runManager,\n  }: {\n    userResponse?: BaseMessage;\n    /**\n     * Optional if the prediction was accurate.\n     */\n    predictions: {\n      revisedPrediction: string;\n      explainedPredictionErrors: Array<string>;\n    };\n    runManager?: CallbackManagerForChainRun;\n  }): Promise<string> {\n    const chain = GENERATE_FACTS_PROMPT.pipe(this.llm).pipe(\n      this.stringOutputParser\n    );\n\n    if (typeof userResponse?.content !== \"string\") {\n      throw new Error(\"This chain does not support non-string model output.\");\n    }\n    const res = await chain.invoke(\n      {\n        prediction_violations: predictions.explainedPredictionErrors.join(\"\\n\"),\n        prediction: predictions.revisedPrediction,\n        user_message: userResponse?.content ?? \"\",\n      },\n      runManager?.getChild(\"generate_facts\")\n    );\n\n    return res;\n  }\n\n  /**\n   * Static method that creates a ViolationOfExpectationsChain instance from a\n   * ChatOpenAI and retriever. It also accepts optional options\n   * to customize the chain.\n   *\n   * @param llm The ChatOpenAI instance.\n   * @param retriever The retriever used for similarity search.\n   * @param options Optional options to customize the chain.\n   *\n   * @returns A new instance of ViolationOfExpectationsChain.\n   */\n  static fromLLM(\n    llm: ChatOpenAI,\n    retriever: BaseRetrieverInterface,\n    options?: Partial<\n      Omit<ViolationOfExpectationsChainInput, \"llm\" | \"retriever\">\n    >\n  ): ViolationOfExpectationsChain {\n    return new this({\n      retriever,\n      llm,\n      ...options,\n    });\n  }\n}\n"],"mappings":";;;;;;;;;;;AA6CA,IAAa,+BAAb,cACU,UAEV;CACE,OAAO,UAAU;AACf,SAAO;;CAGT,aAAqB;AACnB,SAAO;;CAGT,iBAAiB;CAEjB,cAAc;CAEd,IAAI,YAAY;AACd,SAAO,CAAC,KAAK,eAAe;;CAG9B,IAAI,aAAa;AACf,SAAO,CAAC,KAAK,YAAY;;CAG3B;CAEA;CAEA;CAEA;CAEA,YAAY,QAA2C;AACrD,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,MAAM,OAAO;AAClB,OAAK,mBAAmB,IAAI,2BAA2B;AACvD,OAAK,qBAAqB,IAAI,oBAAoB;;CAGpD,qBAAqB,aAAoC;AACvD,SAAO,YACJ,KAAK,gBAAgB;AACpB,OAAI,YAAY,UAAU,KAAK,QAC7B,QAAO,UAAU,YAAY;YACpB,YAAY,UAAU,KAAK,KACpC,QAAO,OAAO,YAAY;OAE1B,QAAO,GAAG,YAAY;IAExB,CACD,KAAK,KAAK;;CAGf,uBAAuB,SAAuC;AAC5D,SAAO,CAAC,GAAG,IAAI,IAAI,QAAQ,CAAC;;;;;;;;;;;;;;;;;;CAmB9B,0BAA0B,aAAkD;EAC1E,MAAM,WAAiC,EAAE;EACzC,MAAM,YAA2B,EAAE;AAEnC,cAAY,SAAS,MAAM,UAAU;AACnC,aAAU,KAAK,KAAK;AACpB,OAAI,KAAK,UAAU,KAAK,MAAM;IAC5B,IAAI,eAAwC,YAAY,QAAQ;AAChE,QAAI,CAAC,gBAAgB,aAAa,UAAU,KAAK,QAC/C,gBAAe,KAAA;AAGjB,aAAS,KAAK;KACZ,iBAAiB;KACjB,cAAc,eACV,IAAI,aAAa,aAAa,GAC9B,KAAA;KACL,CAAC;;IAEJ;AAEF,SAAO;;;;;;;;;;;;;;;;;;;;;CAsBT,MAAM,MACJ,QACA,YACsB;AACtB,MAAI,EAAE,KAAK,kBAAkB,QAC3B,OAAM,IAAI,MAAM,oBAAoB,KAAK,eAAe,YAAY;EAGtE,MAAM,cAAyB,OAAO,KAAK;AAK3C,MAAI,CAH8B,YAAY,OAAO,YACnD,cAAc,QAAQ,CACvB,CAEC,OAAM,IAAI,MAAM,gDAAgD;EAGlE,MAAM,gBAAgB,KAAK,0BACzB,YACD;EAGD,MAAM,kBAAkB,MAAM,QAAQ,IACpC,cAAc,IAAI,OAAO,sBAAsB;GAC7C,iBAAiB,MAAM,KAAK,uBAC1B,iBAAiB,gBAClB;GACD,cAAc,iBAAiB;GAC/B;GACD,EAAE,CACJ;EAID,MAAM,uBAAuB,MAAM,QAAQ,IACzC,gBAAgB,KAAK,eACnB,KAAK,wBAAwB;GAC3B,iBAAiB,WAAW;GAC5B,cAAc,WAAW;GACzB;GACD,CAAC,CACH,CACF;AAeD,SAAO,EACL,UAbe,MAAM,QAAQ,IAC7B,qBAAqB,KAAK,cACxB,KAAK,cAAc;GACjB,cAAc,UAAU;GACxB,aAAa;IACX,mBAAmB,UAAU;IAC7B,2BAA2B,UAAU;IACtC;GACF,CAAC,CACH,CACF,EAIA;;;;;;;;;;;;CAaH,MAAc,uBACZ,aACA,YACyC;EACzC,MAAM,gBAAgB,KAAK,qBAAqB,YAAY;EAE5D,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAAC,mCAAmC,CAAC,CAC/C,WAAW,EACV,eAAe,EAAE,MAAM,mCAAmC,MAAM,EACjE,CAAC;EAMJ,MAAM,MAAM,MAJE,iCAAiC,KAAK,iBAAiB,CAAC,KACpE,KAAK,iBACN,CAEuB,OACtB,EACE,cAAc,eACf,EACD,YAAY,SAAS,aAAa,CACnC;AAED,MACE,EACE,eAAe,OACf,0BAA0B,OAC1B,cAAc,KAGhB,OAAM,IAAI,MAAM,8BAA8B,KAAK,UAAU,IAAI,GAAG;EAGtE,MAAM,qBAAqB;EAG3B,MAAM,gBAAgB,MAAM,KAAK,yBAC/B,mBAAmB,SACpB;EACD,MAAM,eAAe,KAAK,uBAAuB,CAC/C,GAAG,mBAAmB,UACtB,GAAG,cACJ,CAAC;AAEF,SAAO;GACL,GAAG;GACH,UAAU;GACX;;;;;;;;;CAUH,MAAc,yBACZ,UACwB;AAaxB,UAXkC,MAAM,QAAQ,IAC9C,SAAS,IAAI,OAAO,YAAY;AAE9B,WADwB,MAAM,KAAK,UAAU,OAAO,QAAQ,EACrC;IACvB,CACH,EAEyD,KACvD,aAAa,SAAS,YACxB;;;;;;;;;;;;;;;CAkBH,MAAc,wBAAwB,EACpC,iBACA,cACA,cAK2C;EAC3C,MAAM,mBAAmB,KAAK,IAC3B,UAAU,CAAC,+BAA+B,CAAC,CAC3C,WAAW,EACV,eAAe,EAAE,MAAM,+BAA+B,MAAM,EAC7D,CAAC;EAEJ,MAAM,QAAQ,6BAA6B,KAAK,iBAAiB,CAAC,KAChE,KAAK,iBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;EAEzE,MAAM,MAAO,MAAM,MAAM,OACvB;GACE,kBAAkB,gBAAgB;GAClC,eAAe,cAAc,WAAW;GACxC,eAAe,gBAAgB,SAAS,KAAK,KAAK;GACnD,EACD,YAAY,SAAS,wBAAwB,CAC9C;AAcD,SAAO;GACL;GACA,mBATwB,MAAM,KAAK,0BAA0B;IAC7D,oBAAoB,gBAAgB;IACpC,2BAA2B,IAAI;IAC/B,cAAc,gBAAgB;IAC9B;IACD,CAAC;GAKA,2BAA2B,IAAI;GAChC;;;;;;;;;;;;;CAcH,MAAc,0BAA0B,EACtC,oBACA,2BACA,cACA,cAMkB;AAclB,SAT6B,MAJE,mCAAmC,KAChE,KAAK,IACN,CAAC,KAAK,KAAK,mBAAmB,CAE2B,OACxD;GACE,YAAY;GACZ,6BAA6B,0BAA0B,KAAK,KAAK;GACjE,eAAe,aAAa,KAAK,KAAK;GACvC,EACD,YAAY,SAAS,sBAAsB,CAC5C;;;;;;;;;;;;;;CAiBH,MAAc,cAAc,EAC1B,cACA,aACA,cAWkB;EAClB,MAAM,QAAQ,sBAAsB,KAAK,KAAK,IAAI,CAAC,KACjD,KAAK,mBACN;AAED,MAAI,OAAO,cAAc,YAAY,SACnC,OAAM,IAAI,MAAM,uDAAuD;AAWzE,SATY,MAAM,MAAM,OACtB;GACE,uBAAuB,YAAY,0BAA0B,KAAK,KAAK;GACvE,YAAY,YAAY;GACxB,cAAc,cAAc,WAAW;GACxC,EACD,YAAY,SAAS,iBAAiB,CACvC;;;;;;;;;;;;;CAgBH,OAAO,QACL,KACA,WACA,SAG8B;AAC9B,SAAO,IAAI,KAAK;GACd;GACA;GACA,GAAG;GACJ,CAAC"}

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.cjs

@@ -1,6 +1,5 @@
-const require_runtime = require('../../../_virtual/_rolldown/runtime.cjs');
+require("../../../_virtual/_rolldown/runtime.cjs");
 let _langchain_core_prompts = require("@langchain/core/prompts");
-
 //#region src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts
 const PREDICT_NEXT_USER_MESSAGE_PROMPT = /* @__PURE__ */ _langchain_core_prompts.PromptTemplate.fromTemplate(`
 You have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.
@@ -43,10 +42,10 @@ Revised prediction: {prediction}
 Actual user message: {user_message}
 --------
 Relevant fact:`);
-
 //#endregion
 exports.GENERATE_FACTS_PROMPT = GENERATE_FACTS_PROMPT;
 exports.GENERATE_REVISED_PREDICTION_PROMPT = GENERATE_REVISED_PREDICTION_PROMPT;
 exports.PREDICTION_VIOLATIONS_PROMPT = PREDICTION_VIOLATIONS_PROMPT;
 exports.PREDICT_NEXT_USER_MESSAGE_PROMPT = PREDICT_NEXT_USER_MESSAGE_PROMPT;
+
 //# sourceMappingURL=violation_of_expectations_prompt.cjs.map

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"violation_of_expectations_prompt.cjs","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts"],"sourcesContent":["import { PromptTemplate } from \"@langchain/core/prompts\";\n\nexport const PREDICT_NEXT_USER_MESSAGE_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.\nGiven the user's chat history provide the following:\n- Concise reasoning about the users internal mental state.\n- Your prediction on how they will respond to the AI's most recent message.\n- A concise list of any additional insights that would be useful to improve prediction.\n--------\nChat History: {chat_history}`);\n\nexport const PREDICTION_VIOLATIONS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`You have been given a prediction and an actual message from a human and AI conversation.\nUsing the prediction, actual message, and additional user insights, generate the following:\n- How exactly was the original prediction violated? Which parts were wrong? State the exact differences.\n- If there were errors with the prediction, what were they and why?\n--------\nPredicted Output: {predicted_output}\n--------\nActual Output: {actual_output}\n--------\nUser Insights: {user_insights}\n--------\n`);\n\nexport const GENERATE_REVISED_PREDICTION_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with revising a prediction on what a user might say in a chat conversation.\n--------\nYour previous prediction: {prediction}\n--------\nWays in which your prediction was off: {explained_prediction_errors}\n--------\nKey insights to the user: {user_insights}\n--------\nGiven the above, revise your prediction to be more accurate.\nRevised Prediction:`);\n\nexport const GENERATE_FACTS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nGiven a user message, an LLM generated prediction of what that message might be, and a list of violations which the prediction made compared to the actual message, generate a fact about the user, relevant to the users message.\n--------\nPrediction violations: {prediction_violations}\n--------\nRevised prediction: {prediction}\n--------\nActual user message: {user_message}\n--------\nRelevant fact:`);\n"],"mappings":";;;;AAEA,MAAa,mCACK,uDAAe,aAAa;;;;;;;8BAOhB;AAE9B,MAAa,+BACK,uDAAe,aAAa;;;;;;;;;;;EAW5C;AAEF,MAAa,qCACK,uDAAe,aAAa;;;;;;;;;;qBAUzB;AAErB,MAAa,wBACK,uDAAe,aAAa;;;;;;;;;gBAS9B"}
+{"version":3,"file":"violation_of_expectations_prompt.cjs","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts"],"sourcesContent":["import { PromptTemplate } from \"@langchain/core/prompts\";\n\nexport const PREDICT_NEXT_USER_MESSAGE_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.\nGiven the user's chat history provide the following:\n- Concise reasoning about the users internal mental state.\n- Your prediction on how they will respond to the AI's most recent message.\n- A concise list of any additional insights that would be useful to improve prediction.\n--------\nChat History: {chat_history}`);\n\nexport const PREDICTION_VIOLATIONS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`You have been given a prediction and an actual message from a human and AI conversation.\nUsing the prediction, actual message, and additional user insights, generate the following:\n- How exactly was the original prediction violated? Which parts were wrong? State the exact differences.\n- If there were errors with the prediction, what were they and why?\n--------\nPredicted Output: {predicted_output}\n--------\nActual Output: {actual_output}\n--------\nUser Insights: {user_insights}\n--------\n`);\n\nexport const GENERATE_REVISED_PREDICTION_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with revising a prediction on what a user might say in a chat conversation.\n--------\nYour previous prediction: {prediction}\n--------\nWays in which your prediction was off: {explained_prediction_errors}\n--------\nKey insights to the user: {user_insights}\n--------\nGiven the above, revise your prediction to be more accurate.\nRevised Prediction:`);\n\nexport const GENERATE_FACTS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nGiven a user message, an LLM generated prediction of what that message might be, and a list of violations which the prediction made compared to the actual message, generate a fact about the user, relevant to the users message.\n--------\nPrediction violations: {prediction_violations}\n--------\nRevised prediction: {prediction}\n--------\nActual user message: {user_message}\n--------\nRelevant fact:`);\n"],"mappings":";;;AAEA,MAAa,mCACK,wCAAA,eAAe,aAAa;;;;;;;8BAOhB;AAE9B,MAAa,+BACK,wCAAA,eAAe,aAAa;;;;;;;;;;;EAW5C;AAEF,MAAa,qCACK,wCAAA,eAAe,aAAa;;;;;;;;;;qBAUzB;AAErB,MAAa,wBACK,wCAAA,eAAe,aAAa;;;;;;;;;gBAS9B"}

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.js

@@ -1,5 +1,4 @@
 import { PromptTemplate } from "@langchain/core/prompts";
-
 //#region src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts
 const PREDICT_NEXT_USER_MESSAGE_PROMPT = /* @__PURE__ */ PromptTemplate.fromTemplate(`
 You have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.
@@ -42,7 +41,7 @@ Revised prediction: {prediction}
 Actual user message: {user_message}
 --------
 Relevant fact:`);
-
 //#endregion
 export { GENERATE_FACTS_PROMPT, GENERATE_REVISED_PREDICTION_PROMPT, PREDICTION_VIOLATIONS_PROMPT, PREDICT_NEXT_USER_MESSAGE_PROMPT };
+
 //# sourceMappingURL=violation_of_expectations_prompt.js.map

package/dist/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.js.map

@@ -1 +1 @@
-{"version":3,"file":"violation_of_expectations_prompt.js","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts"],"sourcesContent":["import { PromptTemplate } from \"@langchain/core/prompts\";\n\nexport const PREDICT_NEXT_USER_MESSAGE_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.\nGiven the user's chat history provide the following:\n- Concise reasoning about the users internal mental state.\n- Your prediction on how they will respond to the AI's most recent message.\n- A concise list of any additional insights that would be useful to improve prediction.\n--------\nChat History: {chat_history}`);\n\nexport const PREDICTION_VIOLATIONS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`You have been given a prediction and an actual message from a human and AI conversation.\nUsing the prediction, actual message, and additional user insights, generate the following:\n- How exactly was the original prediction violated? Which parts were wrong? State the exact differences.\n- If there were errors with the prediction, what were they and why?\n--------\nPredicted Output: {predicted_output}\n--------\nActual Output: {actual_output}\n--------\nUser Insights: {user_insights}\n--------\n`);\n\nexport const GENERATE_REVISED_PREDICTION_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with revising a prediction on what a user might say in a chat conversation.\n--------\nYour previous prediction: {prediction}\n--------\nWays in which your prediction was off: {explained_prediction_errors}\n--------\nKey insights to the user: {user_insights}\n--------\nGiven the above, revise your prediction to be more accurate.\nRevised Prediction:`);\n\nexport const GENERATE_FACTS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nGiven a user message, an LLM generated prediction of what that message might be, and a list of violations which the prediction made compared to the actual message, generate a fact about the user, relevant to the users message.\n--------\nPrediction violations: {prediction_violations}\n--------\nRevised prediction: {prediction}\n--------\nActual user message: {user_message}\n--------\nRelevant fact:`);\n"],"mappings":";;;AAEA,MAAa,mCACK,+BAAe,aAAa;;;;;;;8BAOhB;AAE9B,MAAa,+BACK,+BAAe,aAAa;;;;;;;;;;;EAW5C;AAEF,MAAa,qCACK,+BAAe,aAAa;;;;;;;;;;qBAUzB;AAErB,MAAa,wBACK,+BAAe,aAAa;;;;;;;;;gBAS9B"}
+{"version":3,"file":"violation_of_expectations_prompt.js","names":[],"sources":["../../../../src/experimental/chains/violation_of_expectations/violation_of_expectations_prompt.ts"],"sourcesContent":["import { PromptTemplate } from \"@langchain/core/prompts\";\n\nexport const PREDICT_NEXT_USER_MESSAGE_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with coming up with insights and data-points based on a chat history between a human and an AI.\nGiven the user's chat history provide the following:\n- Concise reasoning about the users internal mental state.\n- Your prediction on how they will respond to the AI's most recent message.\n- A concise list of any additional insights that would be useful to improve prediction.\n--------\nChat History: {chat_history}`);\n\nexport const PREDICTION_VIOLATIONS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`You have been given a prediction and an actual message from a human and AI conversation.\nUsing the prediction, actual message, and additional user insights, generate the following:\n- How exactly was the original prediction violated? Which parts were wrong? State the exact differences.\n- If there were errors with the prediction, what were they and why?\n--------\nPredicted Output: {predicted_output}\n--------\nActual Output: {actual_output}\n--------\nUser Insights: {user_insights}\n--------\n`);\n\nexport const GENERATE_REVISED_PREDICTION_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nYou have been tasked with revising a prediction on what a user might say in a chat conversation.\n--------\nYour previous prediction: {prediction}\n--------\nWays in which your prediction was off: {explained_prediction_errors}\n--------\nKey insights to the user: {user_insights}\n--------\nGiven the above, revise your prediction to be more accurate.\nRevised Prediction:`);\n\nexport const GENERATE_FACTS_PROMPT =\n  /* #__PURE__ */ PromptTemplate.fromTemplate(`\nGiven a user message, an LLM generated prediction of what that message might be, and a list of violations which the prediction made compared to the actual message, generate a fact about the user, relevant to the users message.\n--------\nPrediction violations: {prediction_violations}\n--------\nRevised prediction: {prediction}\n--------\nActual user message: {user_message}\n--------\nRelevant fact:`);\n"],"mappings":";;AAEA,MAAa,mCACK,+BAAe,aAAa;;;;;;;8BAOhB;AAE9B,MAAa,+BACK,+BAAe,aAAa;;;;;;;;;;;EAW5C;AAEF,MAAa,qCACK,+BAAe,aAAa;;;;;;;;;;qBAUzB;AAErB,MAAa,wBACK,+BAAe,aAAa;;;;;;;;;gBAS9B"}

package/dist/experimental/generative_agents/generative_agent.cjs

@@ -1,8 +1,7 @@
-const require_runtime = require('../../_virtual/_rolldown/runtime.cjs');
-const require_base = require('../../chains/base.cjs');
-const require_llm_chain = require('../../chains/llm_chain.cjs');
+require("../../_virtual/_rolldown/runtime.cjs");
+const require_base = require("../../chains/base.cjs");
+const require_llm_chain = require("../../chains/llm_chain.cjs");
 let _langchain_core_prompts = require("@langchain/core/prompts");
-
 //#region src/experimental/generative_agents/generative_agent.ts
 /**
 * Implementation of a generative agent that can learn and form new memories over
@@ -300,7 +299,7 @@ ${this.summary}`;
 		return this.longTermMemory.addMemory(memoryContent, now, metadata, callbacks);
 	}
 };
-
 //#endregion
 exports.GenerativeAgent = GenerativeAgent;
+
 //# sourceMappingURL=generative_agent.cjs.map

package/dist/experimental/generative_agents/generative_agent.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"generative_agent.cjs","names":["BaseChain","LLMChain","PromptTemplate"],"sources":["../../../src/experimental/generative_agents/generative_agent.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { PromptTemplate } from \"@langchain/core/prompts\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport {\n  CallbackManagerForChainRun,\n  Callbacks,\n} from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../../chains/llm_chain.js\";\nimport { GenerativeAgentMemory } from \"./generative_agent_memory.js\";\nimport { BaseChain } from \"../../chains/base.js\";\n\n/**\n * Configuration for the GenerativeAgent class. Defines the character's\n * name, optional age, permanent traits, status, verbosity, and summary\n * refresh seconds.\n */\nexport type GenerativeAgentConfig = {\n  name: string;\n  age?: number;\n  traits: string;\n  status: string;\n  verbose?: boolean;\n  summaryRefreshSeconds?: number;\n  // dailySummaries?: string[];\n};\n\n/**\n * Implementation of a generative agent that can learn and form new memories over\n * time. It extends the BaseChain class, which is a generic\n * sequence of calls to components, including other chains.\n * @example\n * ```typescript\n * const tommie: GenerativeAgent = new GenerativeAgent(\n *   new OpenAI({ temperature: 0.9, maxTokens: 1500 }),\n *   new GenerativeAgentMemory(\n *     new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *     new TimeWeightedVectorStoreRetriever({\n *       vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *       otherScoreKeys: [\"importance\"],\n *       k: 15,\n *     }),\n *     { reflectionThreshold: 8 },\n *   ),\n *   {\n *     name: \"Tommie\",\n *     age: 25,\n *     traits: \"anxious, likes design, talkative\",\n *     status: \"looking for a job\",\n *   },\n * );\n *\n * await tommie.addMemory(\n *   \"Tommie remembers his dog, Bruno, from when he was a kid\",\n *   new Date(),\n * );\n * const summary = await tommie.getSummary({ forceRefresh: true });\n * const response = await tommie.generateDialogueResponse(\n *   \"USER says Hello Tommie, how are you today?\",\n * );\n * ```\n */\nexport class GenerativeAgent extends BaseChain {\n  static lc_name() {\n    return \"GenerativeAgent\";\n  }\n\n  // a character with memory and innate characterisitics\n  name: string; // the character's name\n\n  age?: number; // the optional age of the character\n\n  traits: string; // permanent traits to ascribe to the character\n\n  status: string; // the traits of the character you wish not to change\n\n  longTermMemory: GenerativeAgentMemory;\n\n  llm: BaseLanguageModelInterface; // the underlying language model\n\n  verbose: boolean; // false\n\n  private summary: string; // stateful self-summary generated via reflection on the character's memory.\n\n  private summaryRefreshSeconds = 3600;\n\n  private lastRefreshed: Date; // the last time the character's summary was regenerated\n\n  // TODO: Add support for daily summaries\n  // private dailySummaries: string[] = []; // summary of the events in the plan that the agent took.\n\n  _chainType(): string {\n    return \"generative_agent_executor\";\n  }\n\n  get inputKeys(): string[] {\n    return [\"observation\", \"suffix\", \"now\"];\n  }\n\n  get outputKeys(): string[] {\n    return [\"output\", \"continue_dialogue\"];\n  }\n\n  constructor(\n    llm: BaseLanguageModelInterface,\n    longTermMemory: GenerativeAgentMemory,\n    config: GenerativeAgentConfig\n  ) {\n    super();\n    this.llm = llm;\n    this.longTermMemory = longTermMemory;\n    this.name = config.name;\n    this.age = config.age;\n    this.traits = config.traits;\n    this.status = config.status;\n    this.verbose = config.verbose ?? this.verbose;\n    this.summary = \"\";\n    this.summaryRefreshSeconds =\n      config.summaryRefreshSeconds ?? this.summaryRefreshSeconds;\n    this.lastRefreshed = new Date();\n    // this.dailySummaries = config.dailySummaries ?? this.dailySummaries;\n  }\n\n  // LLM methods\n  /**\n   * Parses a newline-separated string into a list of strings.\n   * @param text The string to parse.\n   * @returns An array of strings parsed from the input text.\n   */\n  parseList(text: string): string[] {\n    // parse a newline-seperated string into a list of strings\n    const lines: string[] = text.trim().split(\"\\n\");\n    const result: string[] = lines.map((line: string) =>\n      line.replace(/^\\s*\\d+\\.\\s*/, \"\").trim()\n    );\n    return result;\n  }\n\n  /**\n   * Creates a new LLMChain with the given prompt and the agent's language\n   * model, verbosity, output key, and memory.\n   * @param prompt The prompt to use for the LLMChain.\n   * @returns A new LLMChain instance.\n   */\n  chain(prompt: PromptTemplate): LLMChain {\n    const chain = new LLMChain({\n      llm: this.llm,\n      prompt,\n      verbose: this.verbose,\n      outputKey: \"output\", // new\n      memory: this.longTermMemory,\n    });\n    return chain;\n  }\n\n  /**\n   * Extracts the observed entity from the given observation.\n   * @param observation The observation to extract the entity from.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted entity as a string.\n   */\n  async getEntityFromObservations(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the observed entity in the following observation? {observation}\" +\n        \"\\nEntity=\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        observation,\n      },\n      runManager?.getChild(\"entity_extractor\")\n    );\n\n    return result.output;\n  }\n\n  /**\n   * Extracts the action of the given entity from the given observation.\n   * @param observation The observation to extract the action from.\n   * @param entityName The name of the entity to extract the action for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted action as a string.\n   */\n  async getEntityAction(\n    observation: string,\n    entityName: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the {entity} doing in the following observation? {observation}\" +\n        \"\\nThe {entity} is\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        entity: entityName,\n        observation,\n      },\n      runManager?.getChild(\"entity_action_extractor\")\n    );\n    const trimmedResult = result.output.trim();\n    return trimmedResult;\n  }\n\n  /**\n   * Summarizes memories that are most relevant to an observation.\n   * @param observation The observation to summarize related memories for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The summarized memories as a string.\n   */\n  async summarizeRelatedMemories(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    // summarize memories that are most relevant to an observation\n    const prompt = PromptTemplate.fromTemplate(\n      `\n{q1}?\nContext from memory:\n{relevant_memories}\nRelevant context:`\n    );\n    const entityName = await this.getEntityFromObservations(\n      observation,\n      runManager\n    );\n    const entityAction = await this.getEntityAction(\n      observation,\n      entityName,\n      runManager\n    );\n    const q1 = `What is the relationship between ${this.name} and ${entityName}`;\n    const q2 = `${entityName} is ${entityAction}`;\n    const response = await this.chain(prompt).call(\n      {\n        q1,\n        queries: [q1, q2],\n      },\n      runManager?.getChild(\"entity_relationships\")\n    );\n\n    return response.output.trim(); // added output\n  }\n\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    const { observation, suffix, now } = values;\n    // react to a given observation or dialogue act\n    const prompt = PromptTemplate.fromTemplate(\n      `{agent_summary_description}` +\n        `\\nIt is {current_time}.` +\n        `\\n{agent_name}'s status: {agent_status}` +\n        `\\nSummary of relevant context from {agent_name}'s memory:` +\n        \"\\n{relevant_memories}\" +\n        `\\nMost recent observations: {most_recent_memories}` +\n        `\\nObservation: {observation}` +\n        `\\n\\n${suffix}`\n    );\n\n    const agentSummaryDescription = await this.getSummary({}, runManager); // now = now in param\n    const relevantMemoriesStr = await this.summarizeRelatedMemories(\n      observation,\n      runManager\n    );\n    const currentTime = (now || new Date()).toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    const chainInputs: ChainValues = {\n      agent_summary_description: agentSummaryDescription,\n      current_time: currentTime,\n      agent_name: this.name,\n      observation,\n      agent_status: this.status,\n      most_recent_memories: \"\",\n    };\n\n    chainInputs[this.longTermMemory.getRelevantMemoriesKey()] =\n      relevantMemoriesStr;\n\n    const consumedTokens = await this.llm.getNumTokens(\n      await prompt.format({ ...chainInputs })\n    );\n\n    chainInputs[this.longTermMemory.getMostRecentMemoriesTokenKey()] =\n      consumedTokens;\n    const response = await this.chain(prompt).call(\n      chainInputs,\n      runManager?.getChild(\"reaction_from_summary\")\n    );\n\n    const rawOutput = response.output;\n    let output = rawOutput;\n    let continue_dialogue = false;\n\n    if (rawOutput.includes(\"REACT:\")) {\n      const reaction = this._cleanResponse(rawOutput.split(\"REACT:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and reacted by ${reaction}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${reaction}`;\n      continue_dialogue = false;\n    } else if (rawOutput.includes(\"SAY:\")) {\n      const saidValue = this._cleanResponse(rawOutput.split(\"SAY:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${saidValue}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${saidValue}`;\n      continue_dialogue = true;\n    } else if (rawOutput.includes(\"GOODBYE:\")) {\n      const farewell = this._cleanResponse(\n        rawOutput.split(\"GOODBYE:\").pop() ?? \"\"\n      );\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${farewell}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${farewell}`;\n      continue_dialogue = false;\n    }\n\n    return { output, continue_dialogue };\n  }\n\n  private _cleanResponse(text: string | undefined): string {\n    if (text === undefined) {\n      return \"\";\n    }\n    const regex = new RegExp(`^${this.name} `);\n    return text.replace(regex, \"\").trim();\n  }\n\n  /**\n   * Generates a reaction to the given observation.\n   * @param observation The observation to generate a reaction for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateReaction(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate: string =\n      `Should {agent_name} react to the observation, and if so,` +\n      ` what would be an appropriate reaction? Respond in one line.` +\n      ` If the action is to engage in dialogue, write:\\nSAY: \"what to say\"` +\n      ` \\notherwise, write:\\nREACT: {agent_name}'s reaction (if anything).` +\n      ` \\nEither do nothing, react, or say something but not both.\\n\\n`;\n\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  /**\n   * Generates a dialogue response to the given observation.\n   * @param observation The observation to generate a dialogue response for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateDialogueResponse(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate = `What would ${this.name} say? To end the conversation, write: GOODBYE: \"what to say\". Otherwise to continue the conversation, write: SAY: \"what to say next\"\\n\\n`;\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  // Agent stateful' summary methods\n  // Each dialog or response prompt includes a header\n  // summarizing the agent's self-description. This is\n  // updated periodically through probing it's memories\n  /**\n   * Gets the agent's summary, which includes the agent's name, age, traits,\n   * and a summary of the agent's core characteristics. The summary is\n   * updated periodically through probing the agent's memories.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The agent's summary as a string.\n   */\n  async getSummary(\n    config?: {\n      now?: Date;\n      forceRefresh?: boolean;\n    },\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const { now = new Date(), forceRefresh = false } = config ?? {};\n\n    const sinceRefresh = Math.floor(\n      (now.getTime() - this.lastRefreshed.getTime()) / 1000\n    );\n\n    if (\n      !this.summary ||\n      sinceRefresh >= this.summaryRefreshSeconds ||\n      forceRefresh\n    ) {\n      this.summary = await this.computeAgentSummary(runManager);\n      this.lastRefreshed = now;\n    }\n\n    let age;\n    if (this.age) {\n      age = this.age;\n    } else {\n      age = \"N/A\";\n    }\n\n    return `Name: ${this.name} (age: ${age})\nInnate traits: ${this.traits}\n${this.summary}`;\n  }\n\n  /**\n   * Computes the agent's summary by summarizing the agent's core\n   * characteristics given the agent's relevant memories.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The computed summary as a string.\n   */\n  async computeAgentSummary(\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"How would you summarize {name}'s core characteristics given the following statements:\\n\" +\n        \"----------\" +\n        \"{relevant_memories}\" +\n        \"----------\" +\n        \"Do not embellish.\" +\n        \"\\n\\nSummary: \"\n    );\n    // the agent seeks to think about their core characterisitics\n    const result = await this.chain(prompt).call(\n      {\n        name: this.name,\n        queries: [`${this.name}'s core characteristics`],\n      },\n      runManager?.getChild(\"compute_agent_summary\")\n    );\n    return result.output.trim();\n  }\n\n  /**\n   * Returns a full header of the agent's status, summary, and current time.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @returns The full header as a string.\n   */\n  getFullHeader(\n    config: {\n      now?: Date;\n      forceRefresh?: boolean;\n    } = {}\n  ): string {\n    const { now = new Date(), forceRefresh = false } = config;\n    // return a full header of the agent's status, summary, and current time.\n    const summary = this.getSummary({ now, forceRefresh });\n    const currentTimeString = now.toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    return `${summary}\\nIt is ${currentTimeString}.\\n${this.name}'s status: ${this.status}`;\n  }\n\n  /**\n   * Adds a memory to the agent's long-term memory.\n   * @param memoryContent The content of the memory to add.\n   * @param now Optional current date.\n   * @param metadata Optional metadata for the memory.\n   * @param callbacks Optional Callbacks instance.\n   * @returns The result of adding the memory to the agent's long-term memory.\n   */\n  async addMemory(\n    memoryContent: string,\n    now?: Date,\n    metadata?: Record<string, unknown>,\n    callbacks?: Callbacks\n  ) {\n    return this.longTermMemory.addMemory(\n      memoryContent,\n      now,\n      metadata,\n      callbacks\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6DA,IAAa,kBAAb,cAAqCA,uBAAU;CAC7C,OAAO,UAAU;AACf,SAAO;;CAIT;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA,AAAQ;CAER,AAAQ,wBAAwB;CAEhC,AAAQ;CAKR,aAAqB;AACnB,SAAO;;CAGT,IAAI,YAAsB;AACxB,SAAO;GAAC;GAAe;GAAU;GAAM;;CAGzC,IAAI,aAAuB;AACzB,SAAO,CAAC,UAAU,oBAAoB;;CAGxC,YACE,KACA,gBACA,QACA;AACA,SAAO;AACP,OAAK,MAAM;AACX,OAAK,iBAAiB;AACtB,OAAK,OAAO,OAAO;AACnB,OAAK,MAAM,OAAO;AAClB,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,OAAO;AACrB,OAAK,UAAU,OAAO,WAAW,KAAK;AACtC,OAAK,UAAU;AACf,OAAK,wBACH,OAAO,yBAAyB,KAAK;AACvC,OAAK,gCAAgB,IAAI,MAAM;;;;;;;CAUjC,UAAU,MAAwB;AAMhC,SAJwB,KAAK,MAAM,CAAC,MAAM,KAAK,CAChB,KAAK,SAClC,KAAK,QAAQ,gBAAgB,GAAG,CAAC,MAAM,CACxC;;;;;;;;CAUH,MAAM,QAAkC;AAQtC,SAPc,IAAIC,2BAAS;GACzB,KAAK,KAAK;GACV;GACA,SAAS,KAAK;GACd,WAAW;GACX,QAAQ,KAAK;GACd,CAAC;;;;;;;;CAUJ,MAAM,0BACJ,aACA,YACiB;EACjB,MAAM,SAASC,uCAAe,aAC5B,mFAED;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC,EACE,aACD,EACD,YAAY,SAAS,mBAAmB,CACzC,EAEa;;;;;;;;;CAUhB,MAAM,gBACJ,aACA,YACA,YACiB;EACjB,MAAM,SAASA,uCAAe,aAC5B,0FAED;AAUD,UARe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,QAAQ;GACR;GACD,EACD,YAAY,SAAS,0BAA0B,CAChD,EAC4B,OAAO,MAAM;;;;;;;;CAU5C,MAAM,yBACJ,aACA,YACiB;EAEjB,MAAM,SAASA,uCAAe,aAC5B;;;;mBAKD;EACD,MAAM,aAAa,MAAM,KAAK,0BAC5B,aACA,WACD;EACD,MAAM,eAAe,MAAM,KAAK,gBAC9B,aACA,YACA,WACD;EACD,MAAM,KAAK,oCAAoC,KAAK,KAAK,OAAO;EAChE,MAAM,KAAK,GAAG,WAAW,MAAM;AAS/B,UARiB,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC;GACE;GACA,SAAS,CAAC,IAAI,GAAG;GAClB,EACD,YAAY,SAAS,uBAAuB,CAC7C,EAEe,OAAO,MAAM;;CAG/B,MAAM,MACJ,QACA,YACsB;EACtB,MAAM,EAAE,aAAa,QAAQ,QAAQ;EAErC,MAAM,SAASA,uCAAe,aAC5B;;;;;;gCAOS,SACV;EAED,MAAM,0BAA0B,MAAM,KAAK,WAAW,EAAE,EAAE,WAAW;EACrE,MAAM,sBAAsB,MAAM,KAAK,yBACrC,aACA,WACD;EASD,MAAM,cAA2B;GAC/B,2BAA2B;GAC3B,eAVmB,uBAAO,IAAI,MAAM,EAAE,eAAe,SAAS;IAC9D,OAAO;IACP,KAAK;IACL,MAAM;IACN,MAAM;IACN,QAAQ;IACR,QAAQ;IACT,CAAC;GAIA,YAAY,KAAK;GACjB;GACA,cAAc,KAAK;GACnB,sBAAsB;GACvB;AAED,cAAY,KAAK,eAAe,wBAAwB,IACtD;EAEF,MAAM,iBAAiB,MAAM,KAAK,IAAI,aACpC,MAAM,OAAO,OAAO,EAAE,GAAG,aAAa,CAAC,CACxC;AAED,cAAY,KAAK,eAAe,+BAA+B,IAC7D;EAMF,MAAM,aALW,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC,aACA,YAAY,SAAS,wBAAwB,CAC9C,EAE0B;EAC3B,IAAI,SAAS;EACb,IAAI,oBAAoB;AAExB,MAAI,UAAU,SAAS,SAAS,EAAE;GAChC,MAAM,WAAW,KAAK,eAAe,UAAU,MAAM,SAAS,CAAC,KAAK,CAAC;AACrE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,kBAAkB,YACvD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG;AACZ,uBAAoB;aACX,UAAU,SAAS,OAAO,EAAE;GACrC,MAAM,YAAY,KAAK,eAAe,UAAU,MAAM,OAAO,CAAC,KAAK,CAAC;AACpE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,aACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;aACX,UAAU,SAAS,WAAW,EAAE;GACzC,MAAM,WAAW,KAAK,eACpB,UAAU,MAAM,WAAW,CAAC,KAAK,IAAI,GACtC;AACD,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,YACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;;AAGtB,SAAO;GAAE;GAAQ;GAAmB;;CAGtC,AAAQ,eAAe,MAAkC;AACvD,MAAI,SAAS,OACX,QAAO;EAET,MAAM,QAAQ,IAAI,OAAO,IAAI,KAAK,KAAK,GAAG;AAC1C,SAAO,KAAK,QAAQ,OAAO,GAAG,CAAC,MAAM;;;;;;;;CASvC,MAAM,iBACJ,aACA,KAC4B;EAQ5B,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QARA;GASA;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;CASpC,MAAM,yBACJ,aACA,KAC4B;EAC5B,MAAM,uBAAuB,cAAc,KAAK,KAAK;EACrD,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QAAQ;GACR;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;;;CAepC,MAAM,WACJ,QAIA,YACiB;EACjB,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU,UAAU,EAAE;EAE/D,MAAM,eAAe,KAAK,OACvB,IAAI,SAAS,GAAG,KAAK,cAAc,SAAS,IAAI,IAClD;AAED,MACE,CAAC,KAAK,WACN,gBAAgB,KAAK,yBACrB,cACA;AACA,QAAK,UAAU,MAAM,KAAK,oBAAoB,WAAW;AACzD,QAAK,gBAAgB;;EAGvB,IAAI;AACJ,MAAI,KAAK,IACP,OAAM,KAAK;MAEX,OAAM;AAGR,SAAO,SAAS,KAAK,KAAK,SAAS,IAAI;iBAC1B,KAAK,OAAO;EAC3B,KAAK;;;;;;;;CASL,MAAM,oBACJ,YACiB;EACjB,MAAM,SAASA,uCAAe,aAC5B,+JAMD;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,MAAM,KAAK;GACX,SAAS,CAAC,GAAG,KAAK,KAAK,yBAAyB;GACjD,EACD,YAAY,SAAS,wBAAwB,CAC9C,EACa,OAAO,MAAM;;;;;;;CAQ7B,cACE,SAGI,EAAE,EACE;EACR,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU;AAWnD,SAAO,GATS,KAAK,WAAW;GAAE;GAAK;GAAc,CAAC,CASpC,UARQ,IAAI,eAAe,SAAS;GACpD,OAAO;GACP,KAAK;GACL,MAAM;GACN,MAAM;GACN,QAAQ;GACR,QAAQ;GACT,CAAC,CAC4C,KAAK,KAAK,KAAK,aAAa,KAAK;;;;;;;;;;CAWjF,MAAM,UACJ,eACA,KACA,UACA,WACA;AACA,SAAO,KAAK,eAAe,UACzB,eACA,KACA,UACA,UACD"}
+{"version":3,"file":"generative_agent.cjs","names":["BaseChain","LLMChain","PromptTemplate"],"sources":["../../../src/experimental/generative_agents/generative_agent.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { PromptTemplate } from \"@langchain/core/prompts\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport {\n  CallbackManagerForChainRun,\n  Callbacks,\n} from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../../chains/llm_chain.js\";\nimport { GenerativeAgentMemory } from \"./generative_agent_memory.js\";\nimport { BaseChain } from \"../../chains/base.js\";\n\n/**\n * Configuration for the GenerativeAgent class. Defines the character's\n * name, optional age, permanent traits, status, verbosity, and summary\n * refresh seconds.\n */\nexport type GenerativeAgentConfig = {\n  name: string;\n  age?: number;\n  traits: string;\n  status: string;\n  verbose?: boolean;\n  summaryRefreshSeconds?: number;\n  // dailySummaries?: string[];\n};\n\n/**\n * Implementation of a generative agent that can learn and form new memories over\n * time. It extends the BaseChain class, which is a generic\n * sequence of calls to components, including other chains.\n * @example\n * ```typescript\n * const tommie: GenerativeAgent = new GenerativeAgent(\n *   new OpenAI({ temperature: 0.9, maxTokens: 1500 }),\n *   new GenerativeAgentMemory(\n *     new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *     new TimeWeightedVectorStoreRetriever({\n *       vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *       otherScoreKeys: [\"importance\"],\n *       k: 15,\n *     }),\n *     { reflectionThreshold: 8 },\n *   ),\n *   {\n *     name: \"Tommie\",\n *     age: 25,\n *     traits: \"anxious, likes design, talkative\",\n *     status: \"looking for a job\",\n *   },\n * );\n *\n * await tommie.addMemory(\n *   \"Tommie remembers his dog, Bruno, from when he was a kid\",\n *   new Date(),\n * );\n * const summary = await tommie.getSummary({ forceRefresh: true });\n * const response = await tommie.generateDialogueResponse(\n *   \"USER says Hello Tommie, how are you today?\",\n * );\n * ```\n */\nexport class GenerativeAgent extends BaseChain {\n  static lc_name() {\n    return \"GenerativeAgent\";\n  }\n\n  // a character with memory and innate characterisitics\n  name: string; // the character's name\n\n  age?: number; // the optional age of the character\n\n  traits: string; // permanent traits to ascribe to the character\n\n  status: string; // the traits of the character you wish not to change\n\n  longTermMemory: GenerativeAgentMemory;\n\n  llm: BaseLanguageModelInterface; // the underlying language model\n\n  verbose: boolean; // false\n\n  private summary: string; // stateful self-summary generated via reflection on the character's memory.\n\n  private summaryRefreshSeconds = 3600;\n\n  private lastRefreshed: Date; // the last time the character's summary was regenerated\n\n  // TODO: Add support for daily summaries\n  // private dailySummaries: string[] = []; // summary of the events in the plan that the agent took.\n\n  _chainType(): string {\n    return \"generative_agent_executor\";\n  }\n\n  get inputKeys(): string[] {\n    return [\"observation\", \"suffix\", \"now\"];\n  }\n\n  get outputKeys(): string[] {\n    return [\"output\", \"continue_dialogue\"];\n  }\n\n  constructor(\n    llm: BaseLanguageModelInterface,\n    longTermMemory: GenerativeAgentMemory,\n    config: GenerativeAgentConfig\n  ) {\n    super();\n    this.llm = llm;\n    this.longTermMemory = longTermMemory;\n    this.name = config.name;\n    this.age = config.age;\n    this.traits = config.traits;\n    this.status = config.status;\n    this.verbose = config.verbose ?? this.verbose;\n    this.summary = \"\";\n    this.summaryRefreshSeconds =\n      config.summaryRefreshSeconds ?? this.summaryRefreshSeconds;\n    this.lastRefreshed = new Date();\n    // this.dailySummaries = config.dailySummaries ?? this.dailySummaries;\n  }\n\n  // LLM methods\n  /**\n   * Parses a newline-separated string into a list of strings.\n   * @param text The string to parse.\n   * @returns An array of strings parsed from the input text.\n   */\n  parseList(text: string): string[] {\n    // parse a newline-seperated string into a list of strings\n    const lines: string[] = text.trim().split(\"\\n\");\n    const result: string[] = lines.map((line: string) =>\n      line.replace(/^\\s*\\d+\\.\\s*/, \"\").trim()\n    );\n    return result;\n  }\n\n  /**\n   * Creates a new LLMChain with the given prompt and the agent's language\n   * model, verbosity, output key, and memory.\n   * @param prompt The prompt to use for the LLMChain.\n   * @returns A new LLMChain instance.\n   */\n  chain(prompt: PromptTemplate): LLMChain {\n    const chain = new LLMChain({\n      llm: this.llm,\n      prompt,\n      verbose: this.verbose,\n      outputKey: \"output\", // new\n      memory: this.longTermMemory,\n    });\n    return chain;\n  }\n\n  /**\n   * Extracts the observed entity from the given observation.\n   * @param observation The observation to extract the entity from.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted entity as a string.\n   */\n  async getEntityFromObservations(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the observed entity in the following observation? {observation}\" +\n        \"\\nEntity=\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        observation,\n      },\n      runManager?.getChild(\"entity_extractor\")\n    );\n\n    return result.output;\n  }\n\n  /**\n   * Extracts the action of the given entity from the given observation.\n   * @param observation The observation to extract the action from.\n   * @param entityName The name of the entity to extract the action for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted action as a string.\n   */\n  async getEntityAction(\n    observation: string,\n    entityName: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the {entity} doing in the following observation? {observation}\" +\n        \"\\nThe {entity} is\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        entity: entityName,\n        observation,\n      },\n      runManager?.getChild(\"entity_action_extractor\")\n    );\n    const trimmedResult = result.output.trim();\n    return trimmedResult;\n  }\n\n  /**\n   * Summarizes memories that are most relevant to an observation.\n   * @param observation The observation to summarize related memories for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The summarized memories as a string.\n   */\n  async summarizeRelatedMemories(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    // summarize memories that are most relevant to an observation\n    const prompt = PromptTemplate.fromTemplate(\n      `\n{q1}?\nContext from memory:\n{relevant_memories}\nRelevant context:`\n    );\n    const entityName = await this.getEntityFromObservations(\n      observation,\n      runManager\n    );\n    const entityAction = await this.getEntityAction(\n      observation,\n      entityName,\n      runManager\n    );\n    const q1 = `What is the relationship between ${this.name} and ${entityName}`;\n    const q2 = `${entityName} is ${entityAction}`;\n    const response = await this.chain(prompt).call(\n      {\n        q1,\n        queries: [q1, q2],\n      },\n      runManager?.getChild(\"entity_relationships\")\n    );\n\n    return response.output.trim(); // added output\n  }\n\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    const { observation, suffix, now } = values;\n    // react to a given observation or dialogue act\n    const prompt = PromptTemplate.fromTemplate(\n      `{agent_summary_description}` +\n        `\\nIt is {current_time}.` +\n        `\\n{agent_name}'s status: {agent_status}` +\n        `\\nSummary of relevant context from {agent_name}'s memory:` +\n        \"\\n{relevant_memories}\" +\n        `\\nMost recent observations: {most_recent_memories}` +\n        `\\nObservation: {observation}` +\n        `\\n\\n${suffix}`\n    );\n\n    const agentSummaryDescription = await this.getSummary({}, runManager); // now = now in param\n    const relevantMemoriesStr = await this.summarizeRelatedMemories(\n      observation,\n      runManager\n    );\n    const currentTime = (now || new Date()).toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    const chainInputs: ChainValues = {\n      agent_summary_description: agentSummaryDescription,\n      current_time: currentTime,\n      agent_name: this.name,\n      observation,\n      agent_status: this.status,\n      most_recent_memories: \"\",\n    };\n\n    chainInputs[this.longTermMemory.getRelevantMemoriesKey()] =\n      relevantMemoriesStr;\n\n    const consumedTokens = await this.llm.getNumTokens(\n      await prompt.format({ ...chainInputs })\n    );\n\n    chainInputs[this.longTermMemory.getMostRecentMemoriesTokenKey()] =\n      consumedTokens;\n    const response = await this.chain(prompt).call(\n      chainInputs,\n      runManager?.getChild(\"reaction_from_summary\")\n    );\n\n    const rawOutput = response.output;\n    let output = rawOutput;\n    let continue_dialogue = false;\n\n    if (rawOutput.includes(\"REACT:\")) {\n      const reaction = this._cleanResponse(rawOutput.split(\"REACT:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and reacted by ${reaction}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${reaction}`;\n      continue_dialogue = false;\n    } else if (rawOutput.includes(\"SAY:\")) {\n      const saidValue = this._cleanResponse(rawOutput.split(\"SAY:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${saidValue}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${saidValue}`;\n      continue_dialogue = true;\n    } else if (rawOutput.includes(\"GOODBYE:\")) {\n      const farewell = this._cleanResponse(\n        rawOutput.split(\"GOODBYE:\").pop() ?? \"\"\n      );\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${farewell}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${farewell}`;\n      continue_dialogue = false;\n    }\n\n    return { output, continue_dialogue };\n  }\n\n  private _cleanResponse(text: string | undefined): string {\n    if (text === undefined) {\n      return \"\";\n    }\n    const regex = new RegExp(`^${this.name} `);\n    return text.replace(regex, \"\").trim();\n  }\n\n  /**\n   * Generates a reaction to the given observation.\n   * @param observation The observation to generate a reaction for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateReaction(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate: string =\n      `Should {agent_name} react to the observation, and if so,` +\n      ` what would be an appropriate reaction? Respond in one line.` +\n      ` If the action is to engage in dialogue, write:\\nSAY: \"what to say\"` +\n      ` \\notherwise, write:\\nREACT: {agent_name}'s reaction (if anything).` +\n      ` \\nEither do nothing, react, or say something but not both.\\n\\n`;\n\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  /**\n   * Generates a dialogue response to the given observation.\n   * @param observation The observation to generate a dialogue response for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateDialogueResponse(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate = `What would ${this.name} say? To end the conversation, write: GOODBYE: \"what to say\". Otherwise to continue the conversation, write: SAY: \"what to say next\"\\n\\n`;\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  // Agent stateful' summary methods\n  // Each dialog or response prompt includes a header\n  // summarizing the agent's self-description. This is\n  // updated periodically through probing it's memories\n  /**\n   * Gets the agent's summary, which includes the agent's name, age, traits,\n   * and a summary of the agent's core characteristics. The summary is\n   * updated periodically through probing the agent's memories.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The agent's summary as a string.\n   */\n  async getSummary(\n    config?: {\n      now?: Date;\n      forceRefresh?: boolean;\n    },\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const { now = new Date(), forceRefresh = false } = config ?? {};\n\n    const sinceRefresh = Math.floor(\n      (now.getTime() - this.lastRefreshed.getTime()) / 1000\n    );\n\n    if (\n      !this.summary ||\n      sinceRefresh >= this.summaryRefreshSeconds ||\n      forceRefresh\n    ) {\n      this.summary = await this.computeAgentSummary(runManager);\n      this.lastRefreshed = now;\n    }\n\n    let age;\n    if (this.age) {\n      age = this.age;\n    } else {\n      age = \"N/A\";\n    }\n\n    return `Name: ${this.name} (age: ${age})\nInnate traits: ${this.traits}\n${this.summary}`;\n  }\n\n  /**\n   * Computes the agent's summary by summarizing the agent's core\n   * characteristics given the agent's relevant memories.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The computed summary as a string.\n   */\n  async computeAgentSummary(\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"How would you summarize {name}'s core characteristics given the following statements:\\n\" +\n        \"----------\" +\n        \"{relevant_memories}\" +\n        \"----------\" +\n        \"Do not embellish.\" +\n        \"\\n\\nSummary: \"\n    );\n    // the agent seeks to think about their core characterisitics\n    const result = await this.chain(prompt).call(\n      {\n        name: this.name,\n        queries: [`${this.name}'s core characteristics`],\n      },\n      runManager?.getChild(\"compute_agent_summary\")\n    );\n    return result.output.trim();\n  }\n\n  /**\n   * Returns a full header of the agent's status, summary, and current time.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @returns The full header as a string.\n   */\n  getFullHeader(\n    config: {\n      now?: Date;\n      forceRefresh?: boolean;\n    } = {}\n  ): string {\n    const { now = new Date(), forceRefresh = false } = config;\n    // return a full header of the agent's status, summary, and current time.\n    const summary = this.getSummary({ now, forceRefresh });\n    const currentTimeString = now.toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    return `${summary}\\nIt is ${currentTimeString}.\\n${this.name}'s status: ${this.status}`;\n  }\n\n  /**\n   * Adds a memory to the agent's long-term memory.\n   * @param memoryContent The content of the memory to add.\n   * @param now Optional current date.\n   * @param metadata Optional metadata for the memory.\n   * @param callbacks Optional Callbacks instance.\n   * @returns The result of adding the memory to the agent's long-term memory.\n   */\n  async addMemory(\n    memoryContent: string,\n    now?: Date,\n    metadata?: Record<string, unknown>,\n    callbacks?: Callbacks\n  ) {\n    return this.longTermMemory.addMemory(\n      memoryContent,\n      now,\n      metadata,\n      callbacks\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6DA,IAAa,kBAAb,cAAqCA,aAAAA,UAAU;CAC7C,OAAO,UAAU;AACf,SAAO;;CAIT;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA,wBAAgC;CAEhC;CAKA,aAAqB;AACnB,SAAO;;CAGT,IAAI,YAAsB;AACxB,SAAO;GAAC;GAAe;GAAU;GAAM;;CAGzC,IAAI,aAAuB;AACzB,SAAO,CAAC,UAAU,oBAAoB;;CAGxC,YACE,KACA,gBACA,QACA;AACA,SAAO;AACP,OAAK,MAAM;AACX,OAAK,iBAAiB;AACtB,OAAK,OAAO,OAAO;AACnB,OAAK,MAAM,OAAO;AAClB,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,OAAO;AACrB,OAAK,UAAU,OAAO,WAAW,KAAK;AACtC,OAAK,UAAU;AACf,OAAK,wBACH,OAAO,yBAAyB,KAAK;AACvC,OAAK,gCAAgB,IAAI,MAAM;;;;;;;CAUjC,UAAU,MAAwB;AAMhC,SAJwB,KAAK,MAAM,CAAC,MAAM,KAAK,CAChB,KAAK,SAClC,KAAK,QAAQ,gBAAgB,GAAG,CAAC,MAAM,CACxC;;;;;;;;CAUH,MAAM,QAAkC;AAQtC,SAPc,IAAIC,kBAAAA,SAAS;GACzB,KAAK,KAAK;GACV;GACA,SAAS,KAAK;GACd,WAAW;GACX,QAAQ,KAAK;GACd,CAAC;;;;;;;;CAUJ,MAAM,0BACJ,aACA,YACiB;EACjB,MAAM,SAASC,wBAAAA,eAAe,aAC5B,mFAED;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC,EACE,aACD,EACD,YAAY,SAAS,mBAAmB,CACzC,EAEa;;;;;;;;;CAUhB,MAAM,gBACJ,aACA,YACA,YACiB;EACjB,MAAM,SAASA,wBAAAA,eAAe,aAC5B,0FAED;AAUD,UARe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,QAAQ;GACR;GACD,EACD,YAAY,SAAS,0BAA0B,CAChD,EAC4B,OAAO,MAAM;;;;;;;;CAU5C,MAAM,yBACJ,aACA,YACiB;EAEjB,MAAM,SAASA,wBAAAA,eAAe,aAC5B;;;;mBAKD;EACD,MAAM,aAAa,MAAM,KAAK,0BAC5B,aACA,WACD;EACD,MAAM,eAAe,MAAM,KAAK,gBAC9B,aACA,YACA,WACD;EACD,MAAM,KAAK,oCAAoC,KAAK,KAAK,OAAO;EAChE,MAAM,KAAK,GAAG,WAAW,MAAM;AAS/B,UARiB,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC;GACE;GACA,SAAS,CAAC,IAAI,GAAG;GAClB,EACD,YAAY,SAAS,uBAAuB,CAC7C,EAEe,OAAO,MAAM;;CAG/B,MAAM,MACJ,QACA,YACsB;EACtB,MAAM,EAAE,aAAa,QAAQ,QAAQ;EAErC,MAAM,SAASA,wBAAAA,eAAe,aAC5B;;;;;;gCAOS,SACV;EAED,MAAM,0BAA0B,MAAM,KAAK,WAAW,EAAE,EAAE,WAAW;EACrE,MAAM,sBAAsB,MAAM,KAAK,yBACrC,aACA,WACD;EASD,MAAM,cAA2B;GAC/B,2BAA2B;GAC3B,eAVmB,uBAAO,IAAI,MAAM,EAAE,eAAe,SAAS;IAC9D,OAAO;IACP,KAAK;IACL,MAAM;IACN,MAAM;IACN,QAAQ;IACR,QAAQ;IACT,CAAC;GAIA,YAAY,KAAK;GACjB;GACA,cAAc,KAAK;GACnB,sBAAsB;GACvB;AAED,cAAY,KAAK,eAAe,wBAAwB,IACtD;EAEF,MAAM,iBAAiB,MAAM,KAAK,IAAI,aACpC,MAAM,OAAO,OAAO,EAAE,GAAG,aAAa,CAAC,CACxC;AAED,cAAY,KAAK,eAAe,+BAA+B,IAC7D;EAMF,MAAM,aALW,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC,aACA,YAAY,SAAS,wBAAwB,CAC9C,EAE0B;EAC3B,IAAI,SAAS;EACb,IAAI,oBAAoB;AAExB,MAAI,UAAU,SAAS,SAAS,EAAE;GAChC,MAAM,WAAW,KAAK,eAAe,UAAU,MAAM,SAAS,CAAC,KAAK,CAAC;AACrE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,kBAAkB,YACvD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG;AACZ,uBAAoB;aACX,UAAU,SAAS,OAAO,EAAE;GACrC,MAAM,YAAY,KAAK,eAAe,UAAU,MAAM,OAAO,CAAC,KAAK,CAAC;AACpE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,aACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;aACX,UAAU,SAAS,WAAW,EAAE;GACzC,MAAM,WAAW,KAAK,eACpB,UAAU,MAAM,WAAW,CAAC,KAAK,IAAI,GACtC;AACD,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,YACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;;AAGtB,SAAO;GAAE;GAAQ;GAAmB;;CAGtC,eAAuB,MAAkC;AACvD,MAAI,SAAS,KAAA,EACX,QAAO;EAET,MAAM,QAAQ,IAAI,OAAO,IAAI,KAAK,KAAK,GAAG;AAC1C,SAAO,KAAK,QAAQ,OAAO,GAAG,CAAC,MAAM;;;;;;;;CASvC,MAAM,iBACJ,aACA,KAC4B;EAQ5B,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QARA;GASA;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;CASpC,MAAM,yBACJ,aACA,KAC4B;EAC5B,MAAM,uBAAuB,cAAc,KAAK,KAAK;EACrD,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QAAQ;GACR;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;;;CAepC,MAAM,WACJ,QAIA,YACiB;EACjB,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU,UAAU,EAAE;EAE/D,MAAM,eAAe,KAAK,OACvB,IAAI,SAAS,GAAG,KAAK,cAAc,SAAS,IAAI,IAClD;AAED,MACE,CAAC,KAAK,WACN,gBAAgB,KAAK,yBACrB,cACA;AACA,QAAK,UAAU,MAAM,KAAK,oBAAoB,WAAW;AACzD,QAAK,gBAAgB;;EAGvB,IAAI;AACJ,MAAI,KAAK,IACP,OAAM,KAAK;MAEX,OAAM;AAGR,SAAO,SAAS,KAAK,KAAK,SAAS,IAAI;iBAC1B,KAAK,OAAO;EAC3B,KAAK;;;;;;;;CASL,MAAM,oBACJ,YACiB;EACjB,MAAM,SAASA,wBAAAA,eAAe,aAC5B,+JAMD;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,MAAM,KAAK;GACX,SAAS,CAAC,GAAG,KAAK,KAAK,yBAAyB;GACjD,EACD,YAAY,SAAS,wBAAwB,CAC9C,EACa,OAAO,MAAM;;;;;;;CAQ7B,cACE,SAGI,EAAE,EACE;EACR,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU;AAWnD,SAAO,GATS,KAAK,WAAW;GAAE;GAAK;GAAc,CAAC,CASpC,UARQ,IAAI,eAAe,SAAS;GACpD,OAAO;GACP,KAAK;GACL,MAAM;GACN,MAAM;GACN,QAAQ;GACR,QAAQ;GACT,CAAC,CAC4C,KAAK,KAAK,KAAK,aAAa,KAAK;;;;;;;;;;CAWjF,MAAM,UACJ,eACA,KACA,UACA,WACA;AACA,SAAO,KAAK,eAAe,UACzB,eACA,KACA,UACA,UACD"}

package/dist/experimental/generative_agents/generative_agent.js

@@ -1,7 +1,6 @@
 import { BaseChain } from "../../chains/base.js";
 import { LLMChain } from "../../chains/llm_chain.js";
 import { PromptTemplate } from "@langchain/core/prompts";
-
 //#region src/experimental/generative_agents/generative_agent.ts
 /**
 * Implementation of a generative agent that can learn and form new memories over
@@ -299,7 +298,7 @@ ${this.summary}`;
 		return this.longTermMemory.addMemory(memoryContent, now, metadata, callbacks);
 	}
 };
-
 //#endregion
 export { GenerativeAgent };
+
 //# sourceMappingURL=generative_agent.js.map

package/dist/experimental/generative_agents/generative_agent.js.map

@@ -1 +1 @@
-{"version":3,"file":"generative_agent.js","names":[],"sources":["../../../src/experimental/generative_agents/generative_agent.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { PromptTemplate } from \"@langchain/core/prompts\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport {\n  CallbackManagerForChainRun,\n  Callbacks,\n} from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../../chains/llm_chain.js\";\nimport { GenerativeAgentMemory } from \"./generative_agent_memory.js\";\nimport { BaseChain } from \"../../chains/base.js\";\n\n/**\n * Configuration for the GenerativeAgent class. Defines the character's\n * name, optional age, permanent traits, status, verbosity, and summary\n * refresh seconds.\n */\nexport type GenerativeAgentConfig = {\n  name: string;\n  age?: number;\n  traits: string;\n  status: string;\n  verbose?: boolean;\n  summaryRefreshSeconds?: number;\n  // dailySummaries?: string[];\n};\n\n/**\n * Implementation of a generative agent that can learn and form new memories over\n * time. It extends the BaseChain class, which is a generic\n * sequence of calls to components, including other chains.\n * @example\n * ```typescript\n * const tommie: GenerativeAgent = new GenerativeAgent(\n *   new OpenAI({ temperature: 0.9, maxTokens: 1500 }),\n *   new GenerativeAgentMemory(\n *     new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *     new TimeWeightedVectorStoreRetriever({\n *       vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *       otherScoreKeys: [\"importance\"],\n *       k: 15,\n *     }),\n *     { reflectionThreshold: 8 },\n *   ),\n *   {\n *     name: \"Tommie\",\n *     age: 25,\n *     traits: \"anxious, likes design, talkative\",\n *     status: \"looking for a job\",\n *   },\n * );\n *\n * await tommie.addMemory(\n *   \"Tommie remembers his dog, Bruno, from when he was a kid\",\n *   new Date(),\n * );\n * const summary = await tommie.getSummary({ forceRefresh: true });\n * const response = await tommie.generateDialogueResponse(\n *   \"USER says Hello Tommie, how are you today?\",\n * );\n * ```\n */\nexport class GenerativeAgent extends BaseChain {\n  static lc_name() {\n    return \"GenerativeAgent\";\n  }\n\n  // a character with memory and innate characterisitics\n  name: string; // the character's name\n\n  age?: number; // the optional age of the character\n\n  traits: string; // permanent traits to ascribe to the character\n\n  status: string; // the traits of the character you wish not to change\n\n  longTermMemory: GenerativeAgentMemory;\n\n  llm: BaseLanguageModelInterface; // the underlying language model\n\n  verbose: boolean; // false\n\n  private summary: string; // stateful self-summary generated via reflection on the character's memory.\n\n  private summaryRefreshSeconds = 3600;\n\n  private lastRefreshed: Date; // the last time the character's summary was regenerated\n\n  // TODO: Add support for daily summaries\n  // private dailySummaries: string[] = []; // summary of the events in the plan that the agent took.\n\n  _chainType(): string {\n    return \"generative_agent_executor\";\n  }\n\n  get inputKeys(): string[] {\n    return [\"observation\", \"suffix\", \"now\"];\n  }\n\n  get outputKeys(): string[] {\n    return [\"output\", \"continue_dialogue\"];\n  }\n\n  constructor(\n    llm: BaseLanguageModelInterface,\n    longTermMemory: GenerativeAgentMemory,\n    config: GenerativeAgentConfig\n  ) {\n    super();\n    this.llm = llm;\n    this.longTermMemory = longTermMemory;\n    this.name = config.name;\n    this.age = config.age;\n    this.traits = config.traits;\n    this.status = config.status;\n    this.verbose = config.verbose ?? this.verbose;\n    this.summary = \"\";\n    this.summaryRefreshSeconds =\n      config.summaryRefreshSeconds ?? this.summaryRefreshSeconds;\n    this.lastRefreshed = new Date();\n    // this.dailySummaries = config.dailySummaries ?? this.dailySummaries;\n  }\n\n  // LLM methods\n  /**\n   * Parses a newline-separated string into a list of strings.\n   * @param text The string to parse.\n   * @returns An array of strings parsed from the input text.\n   */\n  parseList(text: string): string[] {\n    // parse a newline-seperated string into a list of strings\n    const lines: string[] = text.trim().split(\"\\n\");\n    const result: string[] = lines.map((line: string) =>\n      line.replace(/^\\s*\\d+\\.\\s*/, \"\").trim()\n    );\n    return result;\n  }\n\n  /**\n   * Creates a new LLMChain with the given prompt and the agent's language\n   * model, verbosity, output key, and memory.\n   * @param prompt The prompt to use for the LLMChain.\n   * @returns A new LLMChain instance.\n   */\n  chain(prompt: PromptTemplate): LLMChain {\n    const chain = new LLMChain({\n      llm: this.llm,\n      prompt,\n      verbose: this.verbose,\n      outputKey: \"output\", // new\n      memory: this.longTermMemory,\n    });\n    return chain;\n  }\n\n  /**\n   * Extracts the observed entity from the given observation.\n   * @param observation The observation to extract the entity from.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted entity as a string.\n   */\n  async getEntityFromObservations(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the observed entity in the following observation? {observation}\" +\n        \"\\nEntity=\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        observation,\n      },\n      runManager?.getChild(\"entity_extractor\")\n    );\n\n    return result.output;\n  }\n\n  /**\n   * Extracts the action of the given entity from the given observation.\n   * @param observation The observation to extract the action from.\n   * @param entityName The name of the entity to extract the action for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted action as a string.\n   */\n  async getEntityAction(\n    observation: string,\n    entityName: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the {entity} doing in the following observation? {observation}\" +\n        \"\\nThe {entity} is\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        entity: entityName,\n        observation,\n      },\n      runManager?.getChild(\"entity_action_extractor\")\n    );\n    const trimmedResult = result.output.trim();\n    return trimmedResult;\n  }\n\n  /**\n   * Summarizes memories that are most relevant to an observation.\n   * @param observation The observation to summarize related memories for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The summarized memories as a string.\n   */\n  async summarizeRelatedMemories(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    // summarize memories that are most relevant to an observation\n    const prompt = PromptTemplate.fromTemplate(\n      `\n{q1}?\nContext from memory:\n{relevant_memories}\nRelevant context:`\n    );\n    const entityName = await this.getEntityFromObservations(\n      observation,\n      runManager\n    );\n    const entityAction = await this.getEntityAction(\n      observation,\n      entityName,\n      runManager\n    );\n    const q1 = `What is the relationship between ${this.name} and ${entityName}`;\n    const q2 = `${entityName} is ${entityAction}`;\n    const response = await this.chain(prompt).call(\n      {\n        q1,\n        queries: [q1, q2],\n      },\n      runManager?.getChild(\"entity_relationships\")\n    );\n\n    return response.output.trim(); // added output\n  }\n\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    const { observation, suffix, now } = values;\n    // react to a given observation or dialogue act\n    const prompt = PromptTemplate.fromTemplate(\n      `{agent_summary_description}` +\n        `\\nIt is {current_time}.` +\n        `\\n{agent_name}'s status: {agent_status}` +\n        `\\nSummary of relevant context from {agent_name}'s memory:` +\n        \"\\n{relevant_memories}\" +\n        `\\nMost recent observations: {most_recent_memories}` +\n        `\\nObservation: {observation}` +\n        `\\n\\n${suffix}`\n    );\n\n    const agentSummaryDescription = await this.getSummary({}, runManager); // now = now in param\n    const relevantMemoriesStr = await this.summarizeRelatedMemories(\n      observation,\n      runManager\n    );\n    const currentTime = (now || new Date()).toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    const chainInputs: ChainValues = {\n      agent_summary_description: agentSummaryDescription,\n      current_time: currentTime,\n      agent_name: this.name,\n      observation,\n      agent_status: this.status,\n      most_recent_memories: \"\",\n    };\n\n    chainInputs[this.longTermMemory.getRelevantMemoriesKey()] =\n      relevantMemoriesStr;\n\n    const consumedTokens = await this.llm.getNumTokens(\n      await prompt.format({ ...chainInputs })\n    );\n\n    chainInputs[this.longTermMemory.getMostRecentMemoriesTokenKey()] =\n      consumedTokens;\n    const response = await this.chain(prompt).call(\n      chainInputs,\n      runManager?.getChild(\"reaction_from_summary\")\n    );\n\n    const rawOutput = response.output;\n    let output = rawOutput;\n    let continue_dialogue = false;\n\n    if (rawOutput.includes(\"REACT:\")) {\n      const reaction = this._cleanResponse(rawOutput.split(\"REACT:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and reacted by ${reaction}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${reaction}`;\n      continue_dialogue = false;\n    } else if (rawOutput.includes(\"SAY:\")) {\n      const saidValue = this._cleanResponse(rawOutput.split(\"SAY:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${saidValue}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${saidValue}`;\n      continue_dialogue = true;\n    } else if (rawOutput.includes(\"GOODBYE:\")) {\n      const farewell = this._cleanResponse(\n        rawOutput.split(\"GOODBYE:\").pop() ?? \"\"\n      );\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${farewell}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${farewell}`;\n      continue_dialogue = false;\n    }\n\n    return { output, continue_dialogue };\n  }\n\n  private _cleanResponse(text: string | undefined): string {\n    if (text === undefined) {\n      return \"\";\n    }\n    const regex = new RegExp(`^${this.name} `);\n    return text.replace(regex, \"\").trim();\n  }\n\n  /**\n   * Generates a reaction to the given observation.\n   * @param observation The observation to generate a reaction for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateReaction(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate: string =\n      `Should {agent_name} react to the observation, and if so,` +\n      ` what would be an appropriate reaction? Respond in one line.` +\n      ` If the action is to engage in dialogue, write:\\nSAY: \"what to say\"` +\n      ` \\notherwise, write:\\nREACT: {agent_name}'s reaction (if anything).` +\n      ` \\nEither do nothing, react, or say something but not both.\\n\\n`;\n\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  /**\n   * Generates a dialogue response to the given observation.\n   * @param observation The observation to generate a dialogue response for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateDialogueResponse(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate = `What would ${this.name} say? To end the conversation, write: GOODBYE: \"what to say\". Otherwise to continue the conversation, write: SAY: \"what to say next\"\\n\\n`;\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  // Agent stateful' summary methods\n  // Each dialog or response prompt includes a header\n  // summarizing the agent's self-description. This is\n  // updated periodically through probing it's memories\n  /**\n   * Gets the agent's summary, which includes the agent's name, age, traits,\n   * and a summary of the agent's core characteristics. The summary is\n   * updated periodically through probing the agent's memories.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The agent's summary as a string.\n   */\n  async getSummary(\n    config?: {\n      now?: Date;\n      forceRefresh?: boolean;\n    },\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const { now = new Date(), forceRefresh = false } = config ?? {};\n\n    const sinceRefresh = Math.floor(\n      (now.getTime() - this.lastRefreshed.getTime()) / 1000\n    );\n\n    if (\n      !this.summary ||\n      sinceRefresh >= this.summaryRefreshSeconds ||\n      forceRefresh\n    ) {\n      this.summary = await this.computeAgentSummary(runManager);\n      this.lastRefreshed = now;\n    }\n\n    let age;\n    if (this.age) {\n      age = this.age;\n    } else {\n      age = \"N/A\";\n    }\n\n    return `Name: ${this.name} (age: ${age})\nInnate traits: ${this.traits}\n${this.summary}`;\n  }\n\n  /**\n   * Computes the agent's summary by summarizing the agent's core\n   * characteristics given the agent's relevant memories.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The computed summary as a string.\n   */\n  async computeAgentSummary(\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"How would you summarize {name}'s core characteristics given the following statements:\\n\" +\n        \"----------\" +\n        \"{relevant_memories}\" +\n        \"----------\" +\n        \"Do not embellish.\" +\n        \"\\n\\nSummary: \"\n    );\n    // the agent seeks to think about their core characterisitics\n    const result = await this.chain(prompt).call(\n      {\n        name: this.name,\n        queries: [`${this.name}'s core characteristics`],\n      },\n      runManager?.getChild(\"compute_agent_summary\")\n    );\n    return result.output.trim();\n  }\n\n  /**\n   * Returns a full header of the agent's status, summary, and current time.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @returns The full header as a string.\n   */\n  getFullHeader(\n    config: {\n      now?: Date;\n      forceRefresh?: boolean;\n    } = {}\n  ): string {\n    const { now = new Date(), forceRefresh = false } = config;\n    // return a full header of the agent's status, summary, and current time.\n    const summary = this.getSummary({ now, forceRefresh });\n    const currentTimeString = now.toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    return `${summary}\\nIt is ${currentTimeString}.\\n${this.name}'s status: ${this.status}`;\n  }\n\n  /**\n   * Adds a memory to the agent's long-term memory.\n   * @param memoryContent The content of the memory to add.\n   * @param now Optional current date.\n   * @param metadata Optional metadata for the memory.\n   * @param callbacks Optional Callbacks instance.\n   * @returns The result of adding the memory to the agent's long-term memory.\n   */\n  async addMemory(\n    memoryContent: string,\n    now?: Date,\n    metadata?: Record<string, unknown>,\n    callbacks?: Callbacks\n  ) {\n    return this.longTermMemory.addMemory(\n      memoryContent,\n      now,\n      metadata,\n      callbacks\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6DA,IAAa,kBAAb,cAAqC,UAAU;CAC7C,OAAO,UAAU;AACf,SAAO;;CAIT;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA,AAAQ;CAER,AAAQ,wBAAwB;CAEhC,AAAQ;CAKR,aAAqB;AACnB,SAAO;;CAGT,IAAI,YAAsB;AACxB,SAAO;GAAC;GAAe;GAAU;GAAM;;CAGzC,IAAI,aAAuB;AACzB,SAAO,CAAC,UAAU,oBAAoB;;CAGxC,YACE,KACA,gBACA,QACA;AACA,SAAO;AACP,OAAK,MAAM;AACX,OAAK,iBAAiB;AACtB,OAAK,OAAO,OAAO;AACnB,OAAK,MAAM,OAAO;AAClB,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,OAAO;AACrB,OAAK,UAAU,OAAO,WAAW,KAAK;AACtC,OAAK,UAAU;AACf,OAAK,wBACH,OAAO,yBAAyB,KAAK;AACvC,OAAK,gCAAgB,IAAI,MAAM;;;;;;;CAUjC,UAAU,MAAwB;AAMhC,SAJwB,KAAK,MAAM,CAAC,MAAM,KAAK,CAChB,KAAK,SAClC,KAAK,QAAQ,gBAAgB,GAAG,CAAC,MAAM,CACxC;;;;;;;;CAUH,MAAM,QAAkC;AAQtC,SAPc,IAAI,SAAS;GACzB,KAAK,KAAK;GACV;GACA,SAAS,KAAK;GACd,WAAW;GACX,QAAQ,KAAK;GACd,CAAC;;;;;;;;CAUJ,MAAM,0BACJ,aACA,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,mFAED;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC,EACE,aACD,EACD,YAAY,SAAS,mBAAmB,CACzC,EAEa;;;;;;;;;CAUhB,MAAM,gBACJ,aACA,YACA,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,0FAED;AAUD,UARe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,QAAQ;GACR;GACD,EACD,YAAY,SAAS,0BAA0B,CAChD,EAC4B,OAAO,MAAM;;;;;;;;CAU5C,MAAM,yBACJ,aACA,YACiB;EAEjB,MAAM,SAAS,eAAe,aAC5B;;;;mBAKD;EACD,MAAM,aAAa,MAAM,KAAK,0BAC5B,aACA,WACD;EACD,MAAM,eAAe,MAAM,KAAK,gBAC9B,aACA,YACA,WACD;EACD,MAAM,KAAK,oCAAoC,KAAK,KAAK,OAAO;EAChE,MAAM,KAAK,GAAG,WAAW,MAAM;AAS/B,UARiB,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC;GACE;GACA,SAAS,CAAC,IAAI,GAAG;GAClB,EACD,YAAY,SAAS,uBAAuB,CAC7C,EAEe,OAAO,MAAM;;CAG/B,MAAM,MACJ,QACA,YACsB;EACtB,MAAM,EAAE,aAAa,QAAQ,QAAQ;EAErC,MAAM,SAAS,eAAe,aAC5B;;;;;;gCAOS,SACV;EAED,MAAM,0BAA0B,MAAM,KAAK,WAAW,EAAE,EAAE,WAAW;EACrE,MAAM,sBAAsB,MAAM,KAAK,yBACrC,aACA,WACD;EASD,MAAM,cAA2B;GAC/B,2BAA2B;GAC3B,eAVmB,uBAAO,IAAI,MAAM,EAAE,eAAe,SAAS;IAC9D,OAAO;IACP,KAAK;IACL,MAAM;IACN,MAAM;IACN,QAAQ;IACR,QAAQ;IACT,CAAC;GAIA,YAAY,KAAK;GACjB;GACA,cAAc,KAAK;GACnB,sBAAsB;GACvB;AAED,cAAY,KAAK,eAAe,wBAAwB,IACtD;EAEF,MAAM,iBAAiB,MAAM,KAAK,IAAI,aACpC,MAAM,OAAO,OAAO,EAAE,GAAG,aAAa,CAAC,CACxC;AAED,cAAY,KAAK,eAAe,+BAA+B,IAC7D;EAMF,MAAM,aALW,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC,aACA,YAAY,SAAS,wBAAwB,CAC9C,EAE0B;EAC3B,IAAI,SAAS;EACb,IAAI,oBAAoB;AAExB,MAAI,UAAU,SAAS,SAAS,EAAE;GAChC,MAAM,WAAW,KAAK,eAAe,UAAU,MAAM,SAAS,CAAC,KAAK,CAAC;AACrE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,kBAAkB,YACvD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG;AACZ,uBAAoB;aACX,UAAU,SAAS,OAAO,EAAE;GACrC,MAAM,YAAY,KAAK,eAAe,UAAU,MAAM,OAAO,CAAC,KAAK,CAAC;AACpE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,aACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;aACX,UAAU,SAAS,WAAW,EAAE;GACzC,MAAM,WAAW,KAAK,eACpB,UAAU,MAAM,WAAW,CAAC,KAAK,IAAI,GACtC;AACD,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,YACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;;AAGtB,SAAO;GAAE;GAAQ;GAAmB;;CAGtC,AAAQ,eAAe,MAAkC;AACvD,MAAI,SAAS,OACX,QAAO;EAET,MAAM,QAAQ,IAAI,OAAO,IAAI,KAAK,KAAK,GAAG;AAC1C,SAAO,KAAK,QAAQ,OAAO,GAAG,CAAC,MAAM;;;;;;;;CASvC,MAAM,iBACJ,aACA,KAC4B;EAQ5B,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QARA;GASA;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;CASpC,MAAM,yBACJ,aACA,KAC4B;EAC5B,MAAM,uBAAuB,cAAc,KAAK,KAAK;EACrD,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QAAQ;GACR;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;;;CAepC,MAAM,WACJ,QAIA,YACiB;EACjB,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU,UAAU,EAAE;EAE/D,MAAM,eAAe,KAAK,OACvB,IAAI,SAAS,GAAG,KAAK,cAAc,SAAS,IAAI,IAClD;AAED,MACE,CAAC,KAAK,WACN,gBAAgB,KAAK,yBACrB,cACA;AACA,QAAK,UAAU,MAAM,KAAK,oBAAoB,WAAW;AACzD,QAAK,gBAAgB;;EAGvB,IAAI;AACJ,MAAI,KAAK,IACP,OAAM,KAAK;MAEX,OAAM;AAGR,SAAO,SAAS,KAAK,KAAK,SAAS,IAAI;iBAC1B,KAAK,OAAO;EAC3B,KAAK;;;;;;;;CASL,MAAM,oBACJ,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,+JAMD;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,MAAM,KAAK;GACX,SAAS,CAAC,GAAG,KAAK,KAAK,yBAAyB;GACjD,EACD,YAAY,SAAS,wBAAwB,CAC9C,EACa,OAAO,MAAM;;;;;;;CAQ7B,cACE,SAGI,EAAE,EACE;EACR,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU;AAWnD,SAAO,GATS,KAAK,WAAW;GAAE;GAAK;GAAc,CAAC,CASpC,UARQ,IAAI,eAAe,SAAS;GACpD,OAAO;GACP,KAAK;GACL,MAAM;GACN,MAAM;GACN,QAAQ;GACR,QAAQ;GACT,CAAC,CAC4C,KAAK,KAAK,KAAK,aAAa,KAAK;;;;;;;;;;CAWjF,MAAM,UACJ,eACA,KACA,UACA,WACA;AACA,SAAO,KAAK,eAAe,UACzB,eACA,KACA,UACA,UACD"}
+{"version":3,"file":"generative_agent.js","names":[],"sources":["../../../src/experimental/generative_agents/generative_agent.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { PromptTemplate } from \"@langchain/core/prompts\";\nimport { ChainValues } from \"@langchain/core/utils/types\";\nimport {\n  CallbackManagerForChainRun,\n  Callbacks,\n} from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../../chains/llm_chain.js\";\nimport { GenerativeAgentMemory } from \"./generative_agent_memory.js\";\nimport { BaseChain } from \"../../chains/base.js\";\n\n/**\n * Configuration for the GenerativeAgent class. Defines the character's\n * name, optional age, permanent traits, status, verbosity, and summary\n * refresh seconds.\n */\nexport type GenerativeAgentConfig = {\n  name: string;\n  age?: number;\n  traits: string;\n  status: string;\n  verbose?: boolean;\n  summaryRefreshSeconds?: number;\n  // dailySummaries?: string[];\n};\n\n/**\n * Implementation of a generative agent that can learn and form new memories over\n * time. It extends the BaseChain class, which is a generic\n * sequence of calls to components, including other chains.\n * @example\n * ```typescript\n * const tommie: GenerativeAgent = new GenerativeAgent(\n *   new OpenAI({ temperature: 0.9, maxTokens: 1500 }),\n *   new GenerativeAgentMemory(\n *     new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *     new TimeWeightedVectorStoreRetriever({\n *       vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *       otherScoreKeys: [\"importance\"],\n *       k: 15,\n *     }),\n *     { reflectionThreshold: 8 },\n *   ),\n *   {\n *     name: \"Tommie\",\n *     age: 25,\n *     traits: \"anxious, likes design, talkative\",\n *     status: \"looking for a job\",\n *   },\n * );\n *\n * await tommie.addMemory(\n *   \"Tommie remembers his dog, Bruno, from when he was a kid\",\n *   new Date(),\n * );\n * const summary = await tommie.getSummary({ forceRefresh: true });\n * const response = await tommie.generateDialogueResponse(\n *   \"USER says Hello Tommie, how are you today?\",\n * );\n * ```\n */\nexport class GenerativeAgent extends BaseChain {\n  static lc_name() {\n    return \"GenerativeAgent\";\n  }\n\n  // a character with memory and innate characterisitics\n  name: string; // the character's name\n\n  age?: number; // the optional age of the character\n\n  traits: string; // permanent traits to ascribe to the character\n\n  status: string; // the traits of the character you wish not to change\n\n  longTermMemory: GenerativeAgentMemory;\n\n  llm: BaseLanguageModelInterface; // the underlying language model\n\n  verbose: boolean; // false\n\n  private summary: string; // stateful self-summary generated via reflection on the character's memory.\n\n  private summaryRefreshSeconds = 3600;\n\n  private lastRefreshed: Date; // the last time the character's summary was regenerated\n\n  // TODO: Add support for daily summaries\n  // private dailySummaries: string[] = []; // summary of the events in the plan that the agent took.\n\n  _chainType(): string {\n    return \"generative_agent_executor\";\n  }\n\n  get inputKeys(): string[] {\n    return [\"observation\", \"suffix\", \"now\"];\n  }\n\n  get outputKeys(): string[] {\n    return [\"output\", \"continue_dialogue\"];\n  }\n\n  constructor(\n    llm: BaseLanguageModelInterface,\n    longTermMemory: GenerativeAgentMemory,\n    config: GenerativeAgentConfig\n  ) {\n    super();\n    this.llm = llm;\n    this.longTermMemory = longTermMemory;\n    this.name = config.name;\n    this.age = config.age;\n    this.traits = config.traits;\n    this.status = config.status;\n    this.verbose = config.verbose ?? this.verbose;\n    this.summary = \"\";\n    this.summaryRefreshSeconds =\n      config.summaryRefreshSeconds ?? this.summaryRefreshSeconds;\n    this.lastRefreshed = new Date();\n    // this.dailySummaries = config.dailySummaries ?? this.dailySummaries;\n  }\n\n  // LLM methods\n  /**\n   * Parses a newline-separated string into a list of strings.\n   * @param text The string to parse.\n   * @returns An array of strings parsed from the input text.\n   */\n  parseList(text: string): string[] {\n    // parse a newline-seperated string into a list of strings\n    const lines: string[] = text.trim().split(\"\\n\");\n    const result: string[] = lines.map((line: string) =>\n      line.replace(/^\\s*\\d+\\.\\s*/, \"\").trim()\n    );\n    return result;\n  }\n\n  /**\n   * Creates a new LLMChain with the given prompt and the agent's language\n   * model, verbosity, output key, and memory.\n   * @param prompt The prompt to use for the LLMChain.\n   * @returns A new LLMChain instance.\n   */\n  chain(prompt: PromptTemplate): LLMChain {\n    const chain = new LLMChain({\n      llm: this.llm,\n      prompt,\n      verbose: this.verbose,\n      outputKey: \"output\", // new\n      memory: this.longTermMemory,\n    });\n    return chain;\n  }\n\n  /**\n   * Extracts the observed entity from the given observation.\n   * @param observation The observation to extract the entity from.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted entity as a string.\n   */\n  async getEntityFromObservations(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the observed entity in the following observation? {observation}\" +\n        \"\\nEntity=\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        observation,\n      },\n      runManager?.getChild(\"entity_extractor\")\n    );\n\n    return result.output;\n  }\n\n  /**\n   * Extracts the action of the given entity from the given observation.\n   * @param observation The observation to extract the action from.\n   * @param entityName The name of the entity to extract the action for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The extracted action as a string.\n   */\n  async getEntityAction(\n    observation: string,\n    entityName: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"What is the {entity} doing in the following observation? {observation}\" +\n        \"\\nThe {entity} is\"\n    );\n\n    const result = await this.chain(prompt).call(\n      {\n        entity: entityName,\n        observation,\n      },\n      runManager?.getChild(\"entity_action_extractor\")\n    );\n    const trimmedResult = result.output.trim();\n    return trimmedResult;\n  }\n\n  /**\n   * Summarizes memories that are most relevant to an observation.\n   * @param observation The observation to summarize related memories for.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The summarized memories as a string.\n   */\n  async summarizeRelatedMemories(\n    observation: string,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    // summarize memories that are most relevant to an observation\n    const prompt = PromptTemplate.fromTemplate(\n      `\n{q1}?\nContext from memory:\n{relevant_memories}\nRelevant context:`\n    );\n    const entityName = await this.getEntityFromObservations(\n      observation,\n      runManager\n    );\n    const entityAction = await this.getEntityAction(\n      observation,\n      entityName,\n      runManager\n    );\n    const q1 = `What is the relationship between ${this.name} and ${entityName}`;\n    const q2 = `${entityName} is ${entityAction}`;\n    const response = await this.chain(prompt).call(\n      {\n        q1,\n        queries: [q1, q2],\n      },\n      runManager?.getChild(\"entity_relationships\")\n    );\n\n    return response.output.trim(); // added output\n  }\n\n  async _call(\n    values: ChainValues,\n    runManager?: CallbackManagerForChainRun\n  ): Promise<ChainValues> {\n    const { observation, suffix, now } = values;\n    // react to a given observation or dialogue act\n    const prompt = PromptTemplate.fromTemplate(\n      `{agent_summary_description}` +\n        `\\nIt is {current_time}.` +\n        `\\n{agent_name}'s status: {agent_status}` +\n        `\\nSummary of relevant context from {agent_name}'s memory:` +\n        \"\\n{relevant_memories}\" +\n        `\\nMost recent observations: {most_recent_memories}` +\n        `\\nObservation: {observation}` +\n        `\\n\\n${suffix}`\n    );\n\n    const agentSummaryDescription = await this.getSummary({}, runManager); // now = now in param\n    const relevantMemoriesStr = await this.summarizeRelatedMemories(\n      observation,\n      runManager\n    );\n    const currentTime = (now || new Date()).toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    const chainInputs: ChainValues = {\n      agent_summary_description: agentSummaryDescription,\n      current_time: currentTime,\n      agent_name: this.name,\n      observation,\n      agent_status: this.status,\n      most_recent_memories: \"\",\n    };\n\n    chainInputs[this.longTermMemory.getRelevantMemoriesKey()] =\n      relevantMemoriesStr;\n\n    const consumedTokens = await this.llm.getNumTokens(\n      await prompt.format({ ...chainInputs })\n    );\n\n    chainInputs[this.longTermMemory.getMostRecentMemoriesTokenKey()] =\n      consumedTokens;\n    const response = await this.chain(prompt).call(\n      chainInputs,\n      runManager?.getChild(\"reaction_from_summary\")\n    );\n\n    const rawOutput = response.output;\n    let output = rawOutput;\n    let continue_dialogue = false;\n\n    if (rawOutput.includes(\"REACT:\")) {\n      const reaction = this._cleanResponse(rawOutput.split(\"REACT:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and reacted by ${reaction}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${reaction}`;\n      continue_dialogue = false;\n    } else if (rawOutput.includes(\"SAY:\")) {\n      const saidValue = this._cleanResponse(rawOutput.split(\"SAY:\").pop());\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${saidValue}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${saidValue}`;\n      continue_dialogue = true;\n    } else if (rawOutput.includes(\"GOODBYE:\")) {\n      const farewell = this._cleanResponse(\n        rawOutput.split(\"GOODBYE:\").pop() ?? \"\"\n      );\n      await this.addMemory(\n        `${this.name} observed ${observation} and said ${farewell}`,\n        now,\n        {},\n        runManager?.getChild(\"memory\")\n      );\n      output = `${this.name} said ${farewell}`;\n      continue_dialogue = false;\n    }\n\n    return { output, continue_dialogue };\n  }\n\n  private _cleanResponse(text: string | undefined): string {\n    if (text === undefined) {\n      return \"\";\n    }\n    const regex = new RegExp(`^${this.name} `);\n    return text.replace(regex, \"\").trim();\n  }\n\n  /**\n   * Generates a reaction to the given observation.\n   * @param observation The observation to generate a reaction for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateReaction(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate: string =\n      `Should {agent_name} react to the observation, and if so,` +\n      ` what would be an appropriate reaction? Respond in one line.` +\n      ` If the action is to engage in dialogue, write:\\nSAY: \"what to say\"` +\n      ` \\notherwise, write:\\nREACT: {agent_name}'s reaction (if anything).` +\n      ` \\nEither do nothing, react, or say something but not both.\\n\\n`;\n\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  /**\n   * Generates a dialogue response to the given observation.\n   * @param observation The observation to generate a dialogue response for.\n   * @param now Optional current date.\n   * @returns A boolean indicating whether to continue the dialogue and the output string.\n   */\n  async generateDialogueResponse(\n    observation: string,\n    now?: Date\n  ): Promise<[boolean, string]> {\n    const callToActionTemplate = `What would ${this.name} say? To end the conversation, write: GOODBYE: \"what to say\". Otherwise to continue the conversation, write: SAY: \"what to say next\"\\n\\n`;\n    const { output, continue_dialogue } = await this.call({\n      observation,\n      suffix: callToActionTemplate,\n      now,\n    });\n    return [continue_dialogue, output];\n  }\n\n  // Agent stateful' summary methods\n  // Each dialog or response prompt includes a header\n  // summarizing the agent's self-description. This is\n  // updated periodically through probing it's memories\n  /**\n   * Gets the agent's summary, which includes the agent's name, age, traits,\n   * and a summary of the agent's core characteristics. The summary is\n   * updated periodically through probing the agent's memories.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The agent's summary as a string.\n   */\n  async getSummary(\n    config?: {\n      now?: Date;\n      forceRefresh?: boolean;\n    },\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const { now = new Date(), forceRefresh = false } = config ?? {};\n\n    const sinceRefresh = Math.floor(\n      (now.getTime() - this.lastRefreshed.getTime()) / 1000\n    );\n\n    if (\n      !this.summary ||\n      sinceRefresh >= this.summaryRefreshSeconds ||\n      forceRefresh\n    ) {\n      this.summary = await this.computeAgentSummary(runManager);\n      this.lastRefreshed = now;\n    }\n\n    let age;\n    if (this.age) {\n      age = this.age;\n    } else {\n      age = \"N/A\";\n    }\n\n    return `Name: ${this.name} (age: ${age})\nInnate traits: ${this.traits}\n${this.summary}`;\n  }\n\n  /**\n   * Computes the agent's summary by summarizing the agent's core\n   * characteristics given the agent's relevant memories.\n   * @param runManager Optional CallbackManagerForChainRun instance.\n   * @returns The computed summary as a string.\n   */\n  async computeAgentSummary(\n    runManager?: CallbackManagerForChainRun\n  ): Promise<string> {\n    const prompt = PromptTemplate.fromTemplate(\n      \"How would you summarize {name}'s core characteristics given the following statements:\\n\" +\n        \"----------\" +\n        \"{relevant_memories}\" +\n        \"----------\" +\n        \"Do not embellish.\" +\n        \"\\n\\nSummary: \"\n    );\n    // the agent seeks to think about their core characterisitics\n    const result = await this.chain(prompt).call(\n      {\n        name: this.name,\n        queries: [`${this.name}'s core characteristics`],\n      },\n      runManager?.getChild(\"compute_agent_summary\")\n    );\n    return result.output.trim();\n  }\n\n  /**\n   * Returns a full header of the agent's status, summary, and current time.\n   * @param config Optional configuration object with current date and a boolean to force refresh.\n   * @returns The full header as a string.\n   */\n  getFullHeader(\n    config: {\n      now?: Date;\n      forceRefresh?: boolean;\n    } = {}\n  ): string {\n    const { now = new Date(), forceRefresh = false } = config;\n    // return a full header of the agent's status, summary, and current time.\n    const summary = this.getSummary({ now, forceRefresh });\n    const currentTimeString = now.toLocaleString(\"en-US\", {\n      month: \"long\",\n      day: \"numeric\",\n      year: \"numeric\",\n      hour: \"numeric\",\n      minute: \"numeric\",\n      hour12: true,\n    });\n    return `${summary}\\nIt is ${currentTimeString}.\\n${this.name}'s status: ${this.status}`;\n  }\n\n  /**\n   * Adds a memory to the agent's long-term memory.\n   * @param memoryContent The content of the memory to add.\n   * @param now Optional current date.\n   * @param metadata Optional metadata for the memory.\n   * @param callbacks Optional Callbacks instance.\n   * @returns The result of adding the memory to the agent's long-term memory.\n   */\n  async addMemory(\n    memoryContent: string,\n    now?: Date,\n    metadata?: Record<string, unknown>,\n    callbacks?: Callbacks\n  ) {\n    return this.longTermMemory.addMemory(\n      memoryContent,\n      now,\n      metadata,\n      callbacks\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA6DA,IAAa,kBAAb,cAAqC,UAAU;CAC7C,OAAO,UAAU;AACf,SAAO;;CAIT;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA;CAEA,wBAAgC;CAEhC;CAKA,aAAqB;AACnB,SAAO;;CAGT,IAAI,YAAsB;AACxB,SAAO;GAAC;GAAe;GAAU;GAAM;;CAGzC,IAAI,aAAuB;AACzB,SAAO,CAAC,UAAU,oBAAoB;;CAGxC,YACE,KACA,gBACA,QACA;AACA,SAAO;AACP,OAAK,MAAM;AACX,OAAK,iBAAiB;AACtB,OAAK,OAAO,OAAO;AACnB,OAAK,MAAM,OAAO;AAClB,OAAK,SAAS,OAAO;AACrB,OAAK,SAAS,OAAO;AACrB,OAAK,UAAU,OAAO,WAAW,KAAK;AACtC,OAAK,UAAU;AACf,OAAK,wBACH,OAAO,yBAAyB,KAAK;AACvC,OAAK,gCAAgB,IAAI,MAAM;;;;;;;CAUjC,UAAU,MAAwB;AAMhC,SAJwB,KAAK,MAAM,CAAC,MAAM,KAAK,CAChB,KAAK,SAClC,KAAK,QAAQ,gBAAgB,GAAG,CAAC,MAAM,CACxC;;;;;;;;CAUH,MAAM,QAAkC;AAQtC,SAPc,IAAI,SAAS;GACzB,KAAK,KAAK;GACV;GACA,SAAS,KAAK;GACd,WAAW;GACX,QAAQ,KAAK;GACd,CAAC;;;;;;;;CAUJ,MAAM,0BACJ,aACA,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,mFAED;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC,EACE,aACD,EACD,YAAY,SAAS,mBAAmB,CACzC,EAEa;;;;;;;;;CAUhB,MAAM,gBACJ,aACA,YACA,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,0FAED;AAUD,UARe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,QAAQ;GACR;GACD,EACD,YAAY,SAAS,0BAA0B,CAChD,EAC4B,OAAO,MAAM;;;;;;;;CAU5C,MAAM,yBACJ,aACA,YACiB;EAEjB,MAAM,SAAS,eAAe,aAC5B;;;;mBAKD;EACD,MAAM,aAAa,MAAM,KAAK,0BAC5B,aACA,WACD;EACD,MAAM,eAAe,MAAM,KAAK,gBAC9B,aACA,YACA,WACD;EACD,MAAM,KAAK,oCAAoC,KAAK,KAAK,OAAO;EAChE,MAAM,KAAK,GAAG,WAAW,MAAM;AAS/B,UARiB,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC;GACE;GACA,SAAS,CAAC,IAAI,GAAG;GAClB,EACD,YAAY,SAAS,uBAAuB,CAC7C,EAEe,OAAO,MAAM;;CAG/B,MAAM,MACJ,QACA,YACsB;EACtB,MAAM,EAAE,aAAa,QAAQ,QAAQ;EAErC,MAAM,SAAS,eAAe,aAC5B;;;;;;gCAOS,SACV;EAED,MAAM,0BAA0B,MAAM,KAAK,WAAW,EAAE,EAAE,WAAW;EACrE,MAAM,sBAAsB,MAAM,KAAK,yBACrC,aACA,WACD;EASD,MAAM,cAA2B;GAC/B,2BAA2B;GAC3B,eAVmB,uBAAO,IAAI,MAAM,EAAE,eAAe,SAAS;IAC9D,OAAO;IACP,KAAK;IACL,MAAM;IACN,MAAM;IACN,QAAQ;IACR,QAAQ;IACT,CAAC;GAIA,YAAY,KAAK;GACjB;GACA,cAAc,KAAK;GACnB,sBAAsB;GACvB;AAED,cAAY,KAAK,eAAe,wBAAwB,IACtD;EAEF,MAAM,iBAAiB,MAAM,KAAK,IAAI,aACpC,MAAM,OAAO,OAAO,EAAE,GAAG,aAAa,CAAC,CACxC;AAED,cAAY,KAAK,eAAe,+BAA+B,IAC7D;EAMF,MAAM,aALW,MAAM,KAAK,MAAM,OAAO,CAAC,KACxC,aACA,YAAY,SAAS,wBAAwB,CAC9C,EAE0B;EAC3B,IAAI,SAAS;EACb,IAAI,oBAAoB;AAExB,MAAI,UAAU,SAAS,SAAS,EAAE;GAChC,MAAM,WAAW,KAAK,eAAe,UAAU,MAAM,SAAS,CAAC,KAAK,CAAC;AACrE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,kBAAkB,YACvD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG;AACZ,uBAAoB;aACX,UAAU,SAAS,OAAO,EAAE;GACrC,MAAM,YAAY,KAAK,eAAe,UAAU,MAAM,OAAO,CAAC,KAAK,CAAC;AACpE,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,aACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;aACX,UAAU,SAAS,WAAW,EAAE;GACzC,MAAM,WAAW,KAAK,eACpB,UAAU,MAAM,WAAW,CAAC,KAAK,IAAI,GACtC;AACD,SAAM,KAAK,UACT,GAAG,KAAK,KAAK,YAAY,YAAY,YAAY,YACjD,KACA,EAAE,EACF,YAAY,SAAS,SAAS,CAC/B;AACD,YAAS,GAAG,KAAK,KAAK,QAAQ;AAC9B,uBAAoB;;AAGtB,SAAO;GAAE;GAAQ;GAAmB;;CAGtC,eAAuB,MAAkC;AACvD,MAAI,SAAS,KAAA,EACX,QAAO;EAET,MAAM,QAAQ,IAAI,OAAO,IAAI,KAAK,KAAK,GAAG;AAC1C,SAAO,KAAK,QAAQ,OAAO,GAAG,CAAC,MAAM;;;;;;;;CASvC,MAAM,iBACJ,aACA,KAC4B;EAQ5B,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QARA;GASA;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;CASpC,MAAM,yBACJ,aACA,KAC4B;EAC5B,MAAM,uBAAuB,cAAc,KAAK,KAAK;EACrD,MAAM,EAAE,QAAQ,sBAAsB,MAAM,KAAK,KAAK;GACpD;GACA,QAAQ;GACR;GACD,CAAC;AACF,SAAO,CAAC,mBAAmB,OAAO;;;;;;;;;;CAepC,MAAM,WACJ,QAIA,YACiB;EACjB,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU,UAAU,EAAE;EAE/D,MAAM,eAAe,KAAK,OACvB,IAAI,SAAS,GAAG,KAAK,cAAc,SAAS,IAAI,IAClD;AAED,MACE,CAAC,KAAK,WACN,gBAAgB,KAAK,yBACrB,cACA;AACA,QAAK,UAAU,MAAM,KAAK,oBAAoB,WAAW;AACzD,QAAK,gBAAgB;;EAGvB,IAAI;AACJ,MAAI,KAAK,IACP,OAAM,KAAK;MAEX,OAAM;AAGR,SAAO,SAAS,KAAK,KAAK,SAAS,IAAI;iBAC1B,KAAK,OAAO;EAC3B,KAAK;;;;;;;;CASL,MAAM,oBACJ,YACiB;EACjB,MAAM,SAAS,eAAe,aAC5B,+JAMD;AASD,UAPe,MAAM,KAAK,MAAM,OAAO,CAAC,KACtC;GACE,MAAM,KAAK;GACX,SAAS,CAAC,GAAG,KAAK,KAAK,yBAAyB;GACjD,EACD,YAAY,SAAS,wBAAwB,CAC9C,EACa,OAAO,MAAM;;;;;;;CAQ7B,cACE,SAGI,EAAE,EACE;EACR,MAAM,EAAE,sBAAM,IAAI,MAAM,EAAE,eAAe,UAAU;AAWnD,SAAO,GATS,KAAK,WAAW;GAAE;GAAK;GAAc,CAAC,CASpC,UARQ,IAAI,eAAe,SAAS;GACpD,OAAO;GACP,KAAK;GACL,MAAM;GACN,MAAM;GACN,QAAQ;GACR,QAAQ;GACT,CAAC,CAC4C,KAAK,KAAK,KAAK,aAAa,KAAK;;;;;;;;;;CAWjF,MAAM,UACJ,eACA,KACA,UACA,WACA;AACA,SAAO,KAAK,eAAe,UACzB,eACA,KACA,UACA,UACD"}

package/dist/experimental/generative_agents/generative_agent_memory.cjs

@@ -1,10 +1,9 @@
-const require_runtime = require('../../_virtual/_rolldown/runtime.cjs');
-const require_base = require('../../chains/base.cjs');
-const require_llm_chain = require('../../chains/llm_chain.cjs');
+require("../../_virtual/_rolldown/runtime.cjs");
+const require_base = require("../../chains/base.cjs");
+const require_llm_chain = require("../../chains/llm_chain.cjs");
 let _langchain_core_prompts = require("@langchain/core/prompts");
 let _langchain_core_memory = require("@langchain/core/memory");
 let _langchain_core_documents = require("@langchain/core/documents");
-
 //#region src/experimental/generative_agents/generative_agent_memory.ts
 /**
 * Class that manages the memory of a generative agent in LangChain. It
@@ -349,7 +348,7 @@ var GenerativeAgentMemory = class extends _langchain_core_memory.BaseMemory {
 	*/
 	clear() {}
 };
-
 //#endregion
 exports.GenerativeAgentMemory = GenerativeAgentMemory;
+
 //# sourceMappingURL=generative_agent_memory.cjs.map