A method, device, equipment and storage medium for generating management decisions

Abstract

The present invention discloses a method, apparatus, device, and storage medium for generating management decisions. The method comprises: collecting relevant data from a property management system to obtain a long text dataset; using RAG technology to clean and preprocess the long text dataset to obtain preprocessed property management data; constructing a property manager agent; and using the property manager agent to perform reasoning and analysis on the property management data to generate management decisions. The steps of constructing the property manager agent include: setting a task; invoking a large language model to process the task; executing instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and execution tool sequences; and using a cache processing mechanism to store the historical processing results. The property manager agent of the present invention can not only improve the efficiency and quality of property management, but also enhance the living experience of residents.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and property information management technology, and in particular to a method, device, equipment and storage medium for generating management decisions. Background Art

[0002] In recent years, with technological advancements and improved living standards, property management has been undergoing a transformation from traditional manual services to intelligent management. This process has also led to significant changes in the internal management of property management companies. In the past, property management companies relied heavily on manual labor for daily management, resulting in low efficiency and inconsistent service quality. Property managers face an ever-increasing amount of information and face the daily burden of processing large amounts of redundant and complex data, making it difficult to make effective decisions quickly. Consequently, property managers often struggle to quickly extract key community information from this complex information, leading to irrational work arrangements and, in turn, compromising service quality. Summary of the Invention

[0003] The embodiments of the present invention provide a method, apparatus, device and storage medium for generating management decisions, aiming to solve the problems of low information processing efficiency and uneven service quality in traditional property management.

[0004] In a first aspect, an embodiment of the present invention provides a method for generating management decisions based on a property manager agent, comprising:

[0005] Collect relevant data from the property management system and obtain long text data sets;

[0006] Using RAG technology to clean and preprocess the long text data set to obtain preprocessed property management data;

[0007] Build a property manager agent;

[0008] Performing reasoning and analysis on the property management data by the property manager agent to generate management decisions;

[0009] The steps of constructing the property manager agent include:

[0010] Setting tasks, wherein the tasks include defining objectives, setting output constraints and resource lists;

[0011] Calling a large language model to process the task;

[0012] Executing the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and executing tool sequences;

[0013] A cache processing mechanism is used to store the historical processing results.

[0014] In a second aspect, an embodiment of the present invention provides a device for generating management decisions based on a property manager agent, comprising:

[0015] The collection unit is used to collect relevant data from the property management system and obtain long text data sets;

[0016] A preprocessing unit, configured to clean and preprocess the long text data set using RAG technology to obtain preprocessed property management data;

[0017] Construction unit, used to build property manager agent;

[0018] a reasoning and analysis unit, configured to perform reasoning and analysis on the property management data through the property manager agent to generate management decisions;

[0019] The steps of constructing the property manager agent include:

[0020] Setting tasks, wherein the tasks include defining objectives, setting output constraints and resource lists;

[0021] Calling a large language model to process the task;

[0022] Executing the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and executing tool sequences;

[0023] A cache processing mechanism is used to store the historical processing results.

[0024] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method for generating management decisions based on a property manager agent as described in the first aspect is implemented.

[0025] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the method for generating management decisions based on a property manager agent as described in the first aspect.

[0026] Embodiments of the present invention provide a method, apparatus, device, and storage medium for generating management decisions. The present invention collects relevant data from a property management system to obtain a large amount of long text data sets. The long text data sets are cleaned and preprocessed using RAG (Retrieval-Augmented Generation) technology to obtain property management data. A property manager agent is then constructed, which performs reasoning and analysis on the property management data to generate management decisions. In constructing the property manager agent, the present invention sets specific tasks and calls a large language model to handle these tasks. The large language model is capable of understanding and processing complex instructions, including parsing instructions, selecting execution tools, and executing tool sequences. By executing the output instructions of the large language model, historical processing results are obtained, and finally, a cache processing mechanism is used to store historical processing results. The embodiments of the present invention significantly improve the efficiency and quality of property management, optimize the accuracy and timeliness of decisions, improve service efficiency, reduce labor costs, and alleviate workloads. At the same time, a more humane service model is implemented, driving the property management industry toward intelligent and efficient development. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 A flowchart of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a sub-process of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0030] Figure 3 A schematic diagram of another sub-flow of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0031] Figure 4 A schematic diagram of the RAG technology architecture provided by an embodiment of the present invention;

[0032] Figure 5 A schematic diagram of another sub-flow of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0033] Figure 6 A schematic diagram of another sub-flow of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0034] Figure 7 A schematic diagram of another sub-flow of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0035] Figure 8 A schematic block diagram of the overall process of generating management decisions based on a property manager agent provided by an embodiment of the present invention;

[0036] Figure 9 A schematic diagram of an apparatus for generating management decisions based on a property manager agent according to an embodiment of the present invention;

[0037] Figure 10 A schematic diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0040] It should also be understood that the terms used in the present specification are only for the purpose of describing particular embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0041] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0042] See also Figure 1 , Figure 1 A flowchart of a method for generating management decisions based on a property manager agent provided by an embodiment of the present invention includes steps S101 to S104:

[0043] S101. Collect relevant data from the property management system to obtain a long text data set;

[0044] In this embodiment, in order to realize intelligent property management, we first conduct in-depth research based on property domain knowledge, industry regulations and system documents to build a comprehensive long text data set. This long text data set focuses on the business situation of the property project, including key indicators such as operating data, customer satisfaction, work order operations, and emergencies, aiming to provide a data foundation for deep reasoning and analysis. On this basis, this embodiment summarizes the matters that property managers need to pay attention to and deal with on a daily basis into ten categories: building area and number of service households, weather conditions of the day, yesterday's operating data, yesterday's key arrears households, yesterday's emergency summary, yesterday's customer complaints, follow-up of long-term unclosed complaints, yesterday's work order completion status, customer satisfaction surveys, and today's work plan. In order to collect relevant data in the above categories, the internal management system resources of the property were integrated and collected to obtain long text data sets, realizing comprehensive data aggregation.

[0045] S102, using RAG technology to clean and preprocess the long text data set to obtain preprocessed property management data;

[0046] In this embodiment, the data extracted from the long text dataset obtained in step S101 often contains duplication, inaccuracy, invalidity, and errors. To address this, this embodiment performs rigorous data cleaning on the long text dataset, including removing useless, duplicate, and erroneous data to ensure data quality and accuracy. Furthermore, given the reliance of the large language model employed in the following steps on property domain knowledge, this embodiment employs RAG (Retrieval-Augmented Generation) technology for data cleaning and preprocessing.

[0047] RAG technology can retrieve property management-related information from large amounts of long text datasets, ensuring that pre-processed property management data is more focused and accurate. By combining expertise in the real estate sector, RAG technology can better understand the data content, provide industry-specific context for data pre-processing, and provide high-quality input for subsequent large language model processing.

[0048] In one embodiment, if Figure 2 As shown, the step S102 includes steps S201 to S204:

[0049] S201. Convert the property-related queries and documents in the long text dataset into vector representations according to the following formula:

[0050] E:Q×C→R n

[0051] q=E(Q)

[0052] ci =E(C i )

[0053] Among them, Q and C represent property-related queries and documents, respectively, and R n is an n-dimensional real vector space, c i represents the vector representation of each document in the document collection, and q represents the query vector;

[0054] S202: Calculate the vector search between the query vector and the document vector according to the following formula:

[0055]

[0056] Among them, q·c i Represents the query vector q and the document vector c i The vector dot product of , ||·|| represents the Euclidean norm of the vector;

[0057] S203: Combine the retrieved relevant documents and property-related queries according to the following formula to form context information:

[0058] I=concatenate(Q,C R )

[0059] Among them, C R Indicates related documents;

[0060] S204: Input the context information into the generation model to generate a text set according to the following formula:

[0061] T=G(I)

[0062] G:R m →T

[0063] Here, G is the generative model and m is the dimension of the context vector.

[0064] In this embodiment, in step S201, the architecture of RAG technology consists of two core parts: the retrieval module (Retriever) and the generation module (Generator). In the retrieval module, the property-related query Q and the document C need to be converted into vector representations. Let E:Q×C→R n is the embedding function, where Q and C represent property-related queries and documents respectively, and R n is an n-dimensional real vector space, then for each document c in the document collection i , the vector representations of query Q and document C are q = E(Q) and c respectively i =E(C i ).

[0065] In step S202, vector search is performed by calculating the query vector q and the document vector c i The calculation of cosine similarity not only takes into account the directional relationship between vectors, but also has high computational efficiency.

[0066] Specifically, the dot product is calculated between the query vector and each document vector in the dataset. The Euclidean norm of the query vector and each document vector is then calculated. The dot product result is divided by the product of the norms of the query and document vectors to obtain the cosine similarity between each document vector and the query vector. Documents are ranked based on the cosine similarity, with the document with the highest similarity considered the most relevant to the query.

[0067] Through this vector search method, the present embodiment can quickly and accurately retrieve the most relevant information to the query from a large number of documents, which is crucial for providing key information support for property manager agents when making management decisions.

[0068] In step S203, the generative module (generative model) has already learned the syntax, semantics, and context of natural language using a large amount of text data during the training phase. This pre-training can be based on a general corpus or fine-tuned for a specific domain (such as property management). The generative model requires contextual information as input. This contextual information consists of two parts: the retrieved relevant documents and the property-related query.

[0069] Furthermore, the retrieved relevant documents and property-related queries are spliced ​​and merged to form a new, more comprehensive context. Specifically, the extracted key information is spliced ​​together with the property-related query text to form a continuous text sequence, ensuring that the spliced ​​text is coherent in format and easy for the generative model to understand and process.

[0070] In step S204, based on the input context information, the generative model generates a coherent, accurate, and informative text set T based on the input context information. Each text in text set T represents a different perspective on the property-related query. Specifically, during the generation process, the generative model continuously references the input context information to ensure that the text in the generated text set T is consistent with the query and the relevant document content.

[0071] To sum up, in steps S201 to S204, in the property data integration and preprocessing stage, that is, the preprocessing stage of the long text data set, not only efficient information retrieval is achieved, but also the quality of text generation is ensured, providing a solid data foundation for the property manager intelligent body.

[0072] In one embodiment, if Figure 3As shown, the step S102 further includes steps S301 to S304:

[0073] S301, using the preset rules generated by RAG technology to perform hash deduplication, correction and completion operations on the long text data set;

[0074] S302: Calculate the accuracy of the data in the long text dataset according to the following formula:

[0075]

[0076] Where D represents the data set, x i represents a single data item in the dataset D, and y represents the correct data item in the dataset D;

[0077] S303: Calculate the completeness rate of the data in the long text dataset according to the following formula:

[0078]

[0079] Where n represents the total number of data fields in the dataset D, F j Indicates the jth field. It is an indicator function that returns 0 or 1 according to the value of the input data field;

[0080] S304: Calculate the consistency of the data in the long text dataset according to the following formula:

[0081]

[0082] Where z represents the data item in the dataset D that meets the specifications.

[0083] In this embodiment, the preset rules R generated by RAG technology are used to perform deduplication, correction and completion operations on the data extracted from the long text data set. Specifically, the hash function H(X) is used to achieve data deduplication, where H is the hash function and X is the data item, ensuring the uniqueness of each data item. After data cleaning, data quality assessment is performed to ensure the validity and reliability of the data. Accuracy, completeness and consistency are key indicators for evaluating data quality. Through the above process, this embodiment effectively solves the quality problems encountered in the data integration process and ensures the high accuracy and integrity of the data. The application of RAG technology not only improves the speed of data processing, but also reduces the need for manual intervention, providing efficient and reliable data support for the intelligent property manager system, thereby laying a solid foundation for the intelligence and refinement of property management.

[0084] In summary, if Figure 4Figure 2 shows the RAG technology architecture, which is divided into the following stages: data sourcing, data preprocessing, data storage, vectorization and retrieval, prompt enhancement, and model response. First, the data source is divided into the following modules: operations: revenue, profit, and collection rate data; customers: satisfaction and complaint data; emergencies: fire and extreme weather data; work order data; equipment operation data; project data, historical experience and system documents, and industry regulations. Raw data (long text datasets) is collected from various channels within the property management system. This diverse data source provides a comprehensive information foundation for data collection. Next, this data is preprocessed to generate vector data, which is then stored for subsequent query and matching. Next, the user enters input through some method, which is vectorized and then retrieved from the vector database. The relevant information retrieved is then used to enhance the prompt to improve the generated results. Finally, based on the enhanced prompt, the model generates the final response or result. This entire process demonstrates how RAG technology can enhance the model's generation capabilities by combining business data and knowledge bases, thereby providing more accurate and useful responses.

[0085] S103, constructing a property manager intelligent agent;

[0086] In one embodiment, if Figure 5 As shown, step S103 includes steps S401 to S404:

[0087] S401, setting a task, wherein the task includes defining a goal, setting output limits and a resource list;

[0088] S402: Calling a large language model to process the task;

[0089] S403: Execute the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and execution tool sequences;

[0090] S404: Utilize a cache processing mechanism to store the historical processing results.

[0091] In this embodiment, there are four core links involved in the construction of the property manager intelligent agent: task setting, large language model calling, command execution and cache processing.

[0092] In step S401, the user sets a task for the intelligent property manager using a role description, either through a configuration file or interactive input. The user defines the intelligent property manager's target task and optimizes and integrates the task, including setting output constraints and resource lists. This standardizes the data format of the large language model interface to ensure the parsability of the call results. Task configuration requires the output to include two data components: process and command. The process describes the progress of the target completion and the agent's self-reflection, while the command contains the command and its parameters required to execute the task. After the task is set, the large language model M is invoked in step S402 to process the user-defined task.

[0093] Furthermore, in step S403, when executing a task, the property manager agent first parses the instructions output by the large language model M. These instructions will tell the property manager agent what to execute, thereby obtaining historical processing results, and then selects the optimal tool sequence from the tool set based on the reinforcement learning algorithm to maximize the expected return. Specifically, this embodiment allows users to customize the command tool set. Each tool Corresponding to the specific capabilities of the intelligent agent, such as text analysis, web search, and knowledge integration, when performing a task, the intelligent agent first parses the instructions output by the large model, and then selects the optimal tool sequence O from the tool set based on the reinforcement learning algorithm to maximize the expected return.

[0094] In order to achieve this goal, the agent is evaluated based on the evaluation function Z(S,f i ; A) to make decisions, the function takes into account the current state S, potential action f i And the parameter set A of the reinforcement learning model. The selection process can be formalized as:

[0095]

[0096] After the selected tool sequence O is executed, the result Result is the key output of the agent to complete the task. This execution process can be expressed by the following formula: Result = f i (O; A).

[0097] Furthermore, in step S404, to ensure the continuity of the process and the optimization of the results, this embodiment adopts a cache processing mechanism to save the historical processing results. The present invention will feed this historical information h into the model together with the current input, so as to ensure that the model can make more optimized decisions based on a more comprehensive set of information. First, through the embedding function, the current input x is fused with the historical information h extracted from the cache to generate a new input vector V x Then, this new input vector is fed into the generating function G to obtain the optimized output vector

[0098] V x =E(x,h)

[0099]

[0100] Among them, V x It is a new input vector combined with historical information, which contains the data of the current step and the relevant information of historical processing. It is an optimized output vector, which is based on the new input vector that incorporates historical information and aims to provide more accurate and task-oriented predictions or decisions.

[0101] To sum up, through the above four core links, this embodiment constructs an efficient and flexible property manager intelligent body, which can automatically execute commands according to the task objectives set by the user, and continuously optimize the processing flow to improve the intelligence level of property management.

[0102] In one embodiment, if Figure 6 As shown, step S402 includes steps S501 to S503:

[0103] S501. Encode the input sequence according to the following formula to obtain an input code:

[0104] V t =Embedding(t)+PositionalEncoding(t)

[0105] Among them, Embedding represents embedding, PositionalEncoding represents position encoding, and t represents the input sequence;

[0106] S502: Perform a linear transformation on the input code according to the following formula to obtain a query matrix, a key matrix, and a value matrix:

[0107] Q=V t W Q

[0108] K=V t W K

[0109] V=V t W V

[0110] Among them, W Q ,W K ,W V is a learnable weight matrix, Q represents the query matrix, K represents the key matrix, and V represents the value matrix;

[0111] S503. Calculate the self-attention weight using the query matrix, key matrix, and value matrix, and obtain the output result of the self-attention layer according to the following formula:

[0112]

[0113] Among them, d k is the vector dimension of the key matrix K, K T is the transposed matrix of the key matrix K.

[0114] In this embodiment, in step S501, the large language model M first needs to encode the input sequence, usually through an embedding layer, and then add positional encoding to retain the position information of the words in the sequence.

[0115] In step S502, the large language model M is processed. That is, the input code is linearly transformed. The core of the large language model M is the self-attention mechanism and the feed-forward neural network. Specifically, first, the correlation between each element in the input sequence is calculated through the self-attention layer. Specifically, the embedding representation V of the input sequence is t The input code is linearly transformed into three matrices: query, key, and value, denoted as Q, K, and V. Then, in step S503, the self-attention weight is calculated, and the output of the self-attention layer is obtained by the softmax function and the value matrix V.

[0116] In one embodiment, if Figure 7 As shown, the step S402 further includes steps S601 to S604:

[0117] S601: Use a multi-head attention mechanism to send the input code to multiple self-attention layers respectively, then splice the output results of the self-attention layers and perform a linear transformation through a weight matrix to obtain the linear transformation result of the multi-head attention mechanism according to the following formula:

[0118] V1=MultiHeadAttention(V t )=Concat(head1,…,head h )W O

[0119] Among them, W O Represents the weight matrix, head iRepresents the i-th output result of the self-attention layer;

[0120] S602: Perform a further nonlinear transformation on the linear transformation result of the multi-head attention mechanism through a feedforward neural network. The nonlinear transformation result is expressed as follows:

[0121] V Ffn =FFN(V1)=max(0,V1W1+b1)W2+b2

[0122] Among them, W1 and W2 are weight matrices in the feedforward neural network, b1 and b2 are bias terms in the feedforward neural network;

[0123] S603: Integrate the features of different layers of the large language model according to the following formula:

[0124] V Att =LayerNorm(Attention+V t )

[0125] V2=LayerNorm(V Ffn +V Att )

[0126] Among them, V Att It is the feature integration output result V2 of the attention layer, which represents the final feature obtained by integration;

[0127] S604: Generate a probability distribution through a linear layer and a softmax function, and obtain an output code according to the following formula:

[0128] P(W t |V1)=softmax(V2W e )

[0129] Among them, W e is the weight matrix of the output embedding layer, W t is the predicted next word.

[0130] In this embodiment, in step S601, in order to enhance the expressive power of the large language model M, a multi-head attention mechanism is adopted to feed the input sequence into multiple self-attention layers respectively, and then concatenate the outputs of these layers and pass them through an additional weight matrix W. OPerform a linear transformation. In step S602, after the self-attention layer, the output of each position is further nonlinearly transformed through the feedforward neural network. V1 in step S602 is the output of step S601 (i.e., the linear transformation result of the multi-head attention mechanism). In step S603, to improve the stability of model training, layer normalization is applied after the self-attention layer and the feedforward neural network, and residual connections are added to ensure that the large language model M can effectively integrate features from different layers.

[0131] Furthermore, in step S604, the large language model M generates a probability distribution for each word in the vocabulary by passing the output of the linear layer to the Softmax function. This probability distribution represents the probability that each word in the vocabulary is the next word in the current context.

[0132] The final output encoding is a vector where each element corresponds to the probability of a word in the vocabulary. This vector can be used as a tool to predict the next word. Typically, the word with the highest probability will be selected as the predicted output of the large language model M.

[0133] The embodiment of the present invention also proposes an agent planning algorithm based on reverse thinking chain, which aims to optimize the traditional forward thinking strategy. In complex decision-making problems, it is directly from the initial state S0 to the target state S g The strategy of advancement (denoted as S0→S g ), however, this approach is often inefficient, especially in the presence of decision bottlenecks. Forward planning can easily lead to exponential growth in path exploration and increase computational redundancy. This embodiment adopts a reverse planning method, denoted as S g →S0, the core idea is to reduce the search space by constraining the intermediate steps through the target state.

[0134] Specifically, the planned path P is defined as a sequence of state transitions, that is, P = {S0, S1, ..., S g}. The reverse planning algorithm starts from S g Starting from S0, we deduce the reverse direction and express the reversibility of state transition through the following formula:

[0135] S i-1 =f -1 (S i )

[0136] Among them, f -1 Represents the inverse operation of state transition, S i-1 It's S i In addition, a "self-verification" mechanism, denoted as V(P), is introduced to ensure the validity of the planned path. This mechanism verifies each state transition through the following logic:

[0137]

[0138] Among them, assertV is a verification function used to assert that the state transition is valid. This process involves for each step S in P i →S i+1 Logical reasoning and verification ensure that each step correctly achieves state transitions, thereby improving the overall accuracy and efficiency of planning. This approach significantly enhances the property manager agent’s planning capabilities in complex environments.

[0139] S104, performing reasoning analysis on the property management data by the property manager agent to generate a management decision;

[0140] In this embodiment, the property manager agent conducts in-depth analysis of collected property management data, including but not limited to resident complaints, facility maintenance records, and security monitoring data, to identify issues and potential trends. Based on the results of this inference analysis, the property manager agent generates a series of management decision recommendations. These decisions may involve resource allocation, service improvements, and emergency response plans, aiming to optimize property management processes and improve service quality.

[0141] In one embodiment, if Figure 8 As shown, the step S104 includes steps S701 to S704:

[0142] S701. Output the inference result according to the following formula:

[0143] R o =a(D;R)

[0144] Among them, R represents the preset rules generated by RAG technology, and a represents the analysis algorithm;

[0145] S702. Based on the inference result, the property behavior is divided into standardized behavior and non-standardized behavior according to the following formula:

[0146] B,N=J(R o )

[0147] Among them, J is the classification function, B is the standardized behavior, and N is the unstandardized behavior;

[0148] S703: Automatically generate work orders through the property manager agent according to the standardized behavior, and assign tasks through integration with the work order system;

[0149] S704: For non-standardized behaviors, the property manager agent sends daily reports to management personnel according to a preset schedule.

[0150] In this embodiment, the property manager agent conducts an in-depth analysis of the project business status, obtains reasoning results, and divides the reasoning results into two categories: standardized behavior and non-standardized behavior.

[0151] For standardized actions, the Property Manager Agent has direct decision-making capabilities and seamlessly integrates with the work order system. Based on pre-set decision logic, the Property Manager Agent automatically dispatches work orders to responsible personnel, ensuring that standardized actions are executed promptly and accurately.

[0152] For non-standardized behaviors, the property manager agent adopts a notification and reminder mechanism. This invention has designed a timed push function. The property manager agent regularly sends detailed situation descriptions and necessary reminders to relevant personnel based on the specific circumstances of non-standardized actions, ensuring that key information can be promptly addressed and processed.

[0153] In one embodiment, the method for generating management decisions based on a property manager agent further includes steps S801 to S803:

[0154] S801: Construct a dictionary including key output results and corresponding timestamps. The dictionary is represented by the following formula:

[0155] tm={t1:result1,t2:result2,…,t n :result n}

[0156] Among them, result i Indicates the output result, t i Indicates the corresponding timestamp;

[0157] S802. Determine the relevance of the behavior information based on the current goal and historical information according to the following formula:

[0158] relevant_info=F S (target,history)

[0159] Among them, target represents the current target, history represents historical information, and F S is the filter function;

[0160] In this embodiment, in order to ensure the relevance of behavior information, a screening function F is first defined. S , and then select relevant behavior information based on the current target target and historical information history.

[0161] S803: Update the cache of the dictionary, add the new key output result to the dictionary, and remove irrelevant behavior information according to the following formula:

[0162] tm update =F U (tm,new_result,target)

[0163] Among them, F U It is a cache update function, and new_result represents the new key output result.

[0164] In this implementation, in each update iteration, the cache update function F U Add the new key output result new_result to tm and remove the no longer relevant behavior information.

[0165] In summary, in steps S801 to S803, when processing large amounts of data, directly inputting it into the large language model M will inevitably lead to increased token consumption (the smallest unit of text) and increased costs. In addition, the input of irrelevant information may reduce the accuracy of the output of the large language model M. To improve the efficiency of token use and ensure the reliability of the large language model M, this embodiment proposes a cache management strategy.

[0166] This strategy involves aggregating the output of each step into a tool message dictionary, tools_message(tm), which acts as an information intermediary, caching only progress information directly relevant to achieving the current plan's goals. This approach significantly reduces the token consumption of the large language model M when the property manager agent performs complex tasks. It enables the agent to effectively review and utilize historical action data without missing important steps due to information overload, while maintaining the relevance of information and the accuracy of model output. This not only reduces operating costs but also improves the efficiency and effectiveness of intelligent property managers when handling complex business scenarios.

[0167] like Figure 9 As shown, an embodiment of the present invention further provides a device 900 for generating management decisions based on a property manager agent, including a collection unit 901, a preprocessing unit 902, a construction unit 903, and a reasoning and analysis unit 904.

[0168] The collection unit 901 is used to collect relevant data of the property management system and obtain a long text data set;

[0169] A preprocessing unit 902 is configured to clean and preprocess the long text data set using RAG technology to obtain preprocessed property management data;

[0170] A construction unit 903 is used to construct a property manager agent;

[0171] The reasoning and analysis unit 904 is configured to perform reasoning and analysis on the property management data through the property manager agent to generate management decisions;

[0172] Wherein, the construction unit 903 includes:

[0173] Setting tasks, wherein the tasks include defining objectives, setting output constraints and resource lists;

[0174] Calling a large language model to process the task;

[0175] Executing the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and executing tool sequences;

[0176] A cache processing mechanism is used to store the historical processing results.

[0177] The device for generating management decisions based on a property manager agent provided by an embodiment of the present invention collects relevant data from the property management system to obtain a large amount of long text data sets. It uses RAG (Retrieval-Augmented Generation) technology to clean and preprocess the long text data sets to obtain property management data. Then, a property manager agent is constructed. The property manager agent performs reasoning and analysis on the property management data to generate management decisions. In terms of constructing the property manager agent, the present invention sets specific tasks and calls a large language model to handle these tasks. The large language model can understand and process complex instructions, including parsing instructions, selecting execution tools, and executing tool sequences. By executing the output instructions of the large language model, historical processing results are obtained, and finally a cache processing mechanism is used to store historical processing results. The embodiment of the present invention significantly improves the efficiency and quality of property management, optimizes the accuracy and timeliness of decision-making, improves service efficiency, reduces labor costs, and alleviates the workload. At the same time, it realizes a more humane service model and promotes the development of the property management industry towards intelligence and efficiency.

[0178] The embodiment of the present invention also provides a device for generating management decisions based on a property manager agent, which can be implemented in the form of a computer program. The computer program can be used in Figure 10 Runs on the computer equipment shown.

[0179] See also Figure 10 , Figure 10 1 is a schematic block diagram of a computer device provided by an embodiment of the present invention. The computer device 1000 is a server, which can be an independent server or a server cluster composed of multiple servers.

[0180] See Figure 10The computer device 1000 includes a processor 1002 , a memory, and a network interface 105 connected via a system bus 1001 , wherein the memory may include a non-volatile storage medium 1003 and an internal memory 1004 .

[0181] The non-volatile storage medium 1003 can store an operating system 10031 and a computer program 10032. When the computer program 10032 is executed, the processor 1002 can execute a method for generating management decisions based on a property manager agent.

[0182] The processor 1002 is used to provide computing and control capabilities to support the operation of the entire computer device 1000.

[0183] The internal memory 1004 provides an environment for the operation of the computer program 10032 in the non-volatile storage medium 1003. When the computer program 10032 is executed by the processor 1002, the processor 1002 can execute a method for generating management decisions based on a property manager intelligent agent.

[0184] The network interface 1005 is used for network communication, such as providing data information transmission. Those skilled in the art will understand that Figure 10 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device 1000 to which the solution of the present invention is applied. The specific computer device 1000 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0185] Those skilled in the art will understand that Figure 10 The embodiment of the computer device shown in the figure does not constitute a limitation on the specific composition of the computer device. In other embodiments, the computer device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. For example, in some embodiments, the computer device may only include a memory and a processor. In such an embodiment, the structure and function of the memory and processor are the same as those in the figure. Figure 1 The embodiments shown are consistent and will not be described again here.

[0186] It should be understood that in the embodiment of the present invention, the processor 1002 may be a central processing unit (CPU), and the processor 1002 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0187] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0188] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, other division methods may be used. Units with the same function may also be combined into a single unit. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not implemented.

[0189] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the objectives of the embodiments of the present invention.

[0190] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0191] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk.

[0192] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for generating management decisions based on a property manager agent, characterized in that: include: Collect relevant data from the property management system and obtain long text data sets; Using RAG technology to clean and preprocess the long text data set to obtain preprocessed property management data; Build a property manager agent; Performing reasoning and analysis on the property management data by the property manager agent to generate management decisions; The steps of constructing the property manager agent include: Setting tasks, wherein the tasks include defining objectives, setting output constraints and resource lists; Calling a large language model to process the task; Executing the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and executing tool sequences; Utilizing a cache processing mechanism to store the historical processing results; The method for generating management decisions based on a property manager agent further includes: Construct a dictionary including key output results and corresponding timestamps, which is expressed as follows: tm={t1:result1,t2:result2,…,t n :result n } Among them, result i Indicates the output result, t i Indicates the corresponding timestamp; Based on the current goal and historical information, the relevance of the behavioral information is determined as follows: relevant_info=F S (target,history) Among them, target represents the current target, history represents historical information, and F S is the filter function; Update the cache of the dictionary, add the new key output results to the dictionary, and remove irrelevant behavior information according to the following formula: tm update =F U (tm,new_result,target) Among them, F U It is a cache update function, new_result represents the new key output result; The process of performing reasoning and analysis on the property management data by the property manager agent to generate management decisions includes: Output the inference results according to the following formula: R o =a(D;R) Among them, R represents the preset rules generated by RAG technology, a represents the analysis algorithm, and D represents the data set; Based on the above reasoning results, property behaviors are divided into standardized behaviors and non-standardized behaviors according to the following formula: B,N=J(R o ) Among them, J is the classification function, B is the standardized behavior, and N is the unstandardized behavior; Automatically generate work orders through the property manager agent according to the standardized behavior, and assign tasks through integration with the work order system; For non-standardized behaviors, the property manager agent sends daily reports to management personnel according to a preset schedule.

2. The method for generating management decisions based on a property manager agent according to claim 1, characterized in that: Calling the large language model to process the task includes: Encode the input sequence according to the following formula to obtain the input code: V t =Embedding(t)+PositionalEncoding(t) Among them, Embedding represents embedding, PositionalEncoding represents position encoding, and t represents the input sequence; The input code is linearly transformed according to the following formula to obtain the query matrix, key matrix and value matrix: Q=V t W Q K=V t W K V=V t W V Among them, W Q ,W K ,W V is a learnable weight matrix, Q represents the query matrix, K represents the key matrix, and V represents the value matrix; The self-attention weight is calculated by the query matrix, key matrix and value matrix, and the output result of the self-attention layer is obtained according to the following formula: Among them, d k is the vector dimension of the key matrix K, K T is the transposed matrix of the key matrix K.

3. The method for generating management decisions based on a property manager agent according to claim 2, characterized in that: The calling of the large language model to process the task further includes: The multi-head attention mechanism is used to feed the input code into multiple self-attention layers respectively, and then the output results of the self-attention layers are spliced ​​and linearly transformed through the weight matrix. The linear transformation result of the multi-head attention mechanism is obtained according to the following formula: V1=MultiHeadAttention(V t )=Concat(head1,…,head h )W O Among them, W O Represents the weight matrix, head i Represents the i-th output result of the self-attention layer; The linear transformation result of the multi-head attention mechanism is further transformed nonlinearly by a feedforward neural network. The result of the nonlinear transformation is expressed as follows: V Ffn =FFN(V1)=max(0,V1W1+b1)W2+b2 Among them, W1 and W2 are weight matrices in the feedforward neural network, b1 and b2 are bias terms in the feedforward neural network; The features of different layers of the large language model are integrated as follows: V Att =LayerNorm(Attention+V t ) V2=LayerNorm(V Ffn +V Att ) Among them, V Att It is the feature integration output of the attention layer, and V2 represents the final feature obtained by integration; The probability distribution is generated by the linear layer and the softmax function, and the output encoding is obtained according to the following formula: P(W t |V1)=softmax(V2W e ) Among them, W e is the weight matrix of the output embedding layer, W t is the predicted next word.

4. The method for generating management decisions based on a property manager agent according to claim 1, characterized in that: The RAG technology is used to clean and preprocess the long text data set to obtain preprocessed property management data, including: The property-related queries and documents in the long text dataset are converted into vector representations according to the following formula: E:Q×C→R n q=E(Q) c i =E(C i ) Among them, Q and C represent property-related queries and documents, respectively, and R n is an n-dimensional real vector space, C i Represents each document in the document collection, c i represents the vector representation of each document in the document collection, and q represents the query vector; The vector search between the query vector and the document vector is calculated as follows: Among them, q·c i Represents the query vector q and the document vector c i The vector dot product of , ‖·‖ represents the Euclidean norm of the vector; The retrieved relevant documents and property-related queries are concatenated and merged to form context information according to the following formula: I=concatenate(Q,C R ) Among them, C R Indicates related documents; The context information is input into the generation model to generate a text set according to the following formula: T=G(I) G:R m →T Among them, G is the generative model, m is the dimension of the context vector, and R m is an m-dimensional real vector space.

5. The method for generating management decisions based on a property manager agent according to claim 4, characterized in that: The method of using RAG technology to clean and preprocess the long text data set to obtain preprocessed property management data also includes: Using the preset rules generated by RAG technology, the long text data set is hashed for deduplication, correction, and completion; The accuracy of the data in the long text dataset is calculated according to the following formula: Where D represents the data set, x i represents a single data item in the dataset D, and y represents the correct data item in the dataset D; The completeness rate of the data in the long text dataset is calculated according to the following formula: Where n represents the total number of data fields in the dataset D, F j represents the jth field, It is an indicator function that returns 0 or 1 according to the value of the input data field; The consistency of the data in the long text dataset is calculated according to the following formula: Where z represents the data item in the dataset D that meets the specifications.

6. A device for generating management decisions based on a property manager agent, characterized in that: include: The collection unit is used to collect relevant data from the property management system and obtain long text data sets; A preprocessing unit, configured to clean and preprocess the long text data set using RAG technology to obtain preprocessed property management data; Construction unit, used to build property manager agent; a reasoning and analysis unit, configured to perform reasoning and analysis on the property management data through the property manager agent to generate management decisions; The steps of constructing the property manager agent include: Setting tasks, wherein the tasks include defining objectives, setting output constraints and resource lists; Calling a large language model to process the task; Executing the instructions output by the large language model to obtain historical processing results; wherein the instructions include parsing instructions, selecting execution tools, and executing tool sequences; Utilizing a cache processing mechanism to store the historical processing results; The process of performing reasoning and analysis on the property management data by the property manager agent to generate management decisions includes: Output the inference results according to the following formula: R o =a(D;R) Among them, R represents the preset rules generated by RAG technology, a represents the analysis algorithm, and D represents the data set; Based on the above reasoning results, property behaviors are divided into standardized behaviors and non-standardized behaviors according to the following formula: B,N=J(R o ) Among them, J is the classification function, B is the standardized behavior, and N is the unstandardized behavior; Automatically generate work orders through the property manager agent according to the standardized behavior, and assign tasks through integration with the work order system; For non-standardized behaviors, the property manager agent sends daily reports to management personnel according to a preset schedule; The device for generating management decisions based on a property manager agent includes: Construct a dictionary including key output results and corresponding timestamps, which is expressed as follows: tm={t1:result1,t2:result2,…,t n :result n } Among them, result i Indicates the output result, t i Indicates the corresponding timestamp; Based on the current goal and historical information, the relevance of the behavioral information is determined as follows: relevant_info=F S (target,history) Among them, target represents the current target, history represents historical information, and F S is the filter function; Update the cache of the dictionary, add the new key output results to the dictionary, and remove irrelevant behavior information according to the following formula: tm update =F U (tm,new_result,target) Among them, F U It is a cache update function, and new_result represents the new key output result.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the method for generating management decisions based on a property manager agent as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to execute the method for generating management decisions based on a property manager agent as described in any one of claims 1 to 5.

Images