Commercial real estate digital assistant system based on large model and intelligent agent

Abstract

The invention relates to the technical field of intelligent questions and answers, and discloses a commercial real estate digital assistant system based on a large model and an intelligent agent, and the system comprises an original data obtaining module which is used for obtaining the original data of a commercial real estate; the data processing and knowledge base construction module is used for processing the original data to construct different types of knowledge bases; the data and tool application module is used for querying a corresponding knowledge base or calling a tool according to the target question, generating an answer and displaying the answer; the task disassembling and process execution module is used for disassembling the target problem into a plurality of task processes, generating an execution link, and combining the task processes and the execution link to execute the task; a domain large model training module; and a field large model management module. According to related application scenes and industry knowledge in the commercial real estate field, the large model in the training field is finely adjusted, the large model and the agent technology are combined, the digital assistant system oriented to the commercial real estate field is constructed, and complex tasks and requirements in the commercial real estate field are better coped with.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question answering technology, and in particular to a commercial real estate digital assistant system based on a large model and an intelligent agent. Background Art

[0002] Commercial real estate companies have accumulated a wealth of knowledge and data assets in their daily operations. These data are distributed in multiple fields such as plaza site selection, store recruitment, business format planning, operation management and customer service answers. How to efficiently store, analyze and utilize these knowledge and data assets is a key factor in the success of the digital transformation of commercial real estate companies. With the help of intelligent models, algorithms, technologies, systems and products, we can integrate existing data resources and fully tap the potential of data, thereby providing knowledge retrieval, data query, question suggestions and decision-making references for the daily operations of enterprises. The digital question-answering assistant system (abbreviated as digital assistant system) is a common system form that connects data assets and end users. Users can ask questions through interactive methods such as voice input or natural language. The digital assistant system obtains answers from the deposited knowledge base or data assets through artificial intelligence (AI) algorithms or other technologies and returns them to users, thereby providing users with accurate, efficient and visual knowledge query methods and data value mining methods.

[0003] On the one hand, existing technologies focus on applying intelligent question-answering technology to assist digital assistant systems in answering professional questions in various fields, including real estate marketing. On the other hand, they use intelligent agent technology to introduce tool capabilities to assist in the decomposition and execution of complex tasks. However, this combination is limited to the direct application of technology and is not suitable for the commercial real estate field and cannot handle complex commercial real estate questions and answers. Summary of the invention

[0004] In view of this, the present invention provides a commercial real estate digital assistant system based on a large model and an intelligent agent to solve the problem that the existing digital assistant system is not compatible with the commercial real estate field and is unable to handle complex commercial real estate questions and answers.

[0005] In a first aspect, the present invention provides a commercial real estate digital assistant system based on a large model and an intelligent agent, the system comprising: an original data acquisition module, a data processing and knowledge base construction module, a data and tool application module, a task decomposition and process execution module, a domain large model training module, and a domain large model management module, wherein:

[0006] The original data acquisition module is used to obtain the original data of commercial real estate, including: commercial real estate database, company internal document data, and industry knowledge data;

[0007] The data processing and knowledge base construction module is used to process the original data in different ways to build different types of knowledge bases, including: building an indicator knowledge base for commercial real estate databases through data governance, building a scenario knowledge base based on the company's internal document data, and building a model training sample library based on industry knowledge data;

[0008] The data and tool application module is used to query the corresponding knowledge base or call the tool according to the target question, generate the answer to the target question and display it;

[0009] The task decomposition and process execution module is used to decompose the target problem into multiple task processes, generate execution links, and execute tasks in combination with task processes and execution links;

[0010] The domain large model training module is used to train multiple domain large models based on the original data;

[0011] The domain big model management module is used to store multiple domain big models and call different domain big models according to the target problem.

[0012] The commercial real estate digital assistant system based on big models and intelligent agents provided by the present invention fine-tunes the training domain big model according to the relevant application scenarios and industry knowledge in the commercial real estate field, and combines the big model and intelligent agent technology to build a digital assistant system for the commercial real estate field. The system can communicate with users through question-and-answer interaction to better cope with complex tasks and needs in the commercial real estate field.

[0013] In an optional implementation, querying a corresponding knowledge base or calling a tool according to the target question, generating an answer to the target question and displaying the answer include:

[0014] By converting natural language queries into structured query language, dynamic query and visual display of indicator knowledge base can be achieved;

[0015] Based on the internal knowledge retrieval and query generated by retrieval enhancement, the query of the scene knowledge base is realized;

[0016] By calling tools or interfaces, external resources are combined with the knowledge base to achieve intelligent question and answer.

[0017] The commercial real estate digital assistant system based on a large model and an intelligent agent provided by the present invention uses different retrieval methods to retrieve the corresponding knowledge base according to the target question, and visualizes the query results to meet the needs of basic commercial real estate queries, supports calling tools or interfaces, and improves the flexibility of the digital assistant system.

[0018] In an optional implementation, the target problem is decomposed into multiple task flows, an execution link is generated, and the task is executed in combination with the task flow and the execution link, including:

[0019] Combined with historical conversation data, the target question is rewritten and filtered to obtain complete semantic content;

[0020] Divide and identify the target question's intent based on the complete semantic content, determine the user's intent, and judge the task type of the target question based on the user's intent;

[0021] If the task type of the target problem is a multi-node task, the target problem is decomposed and links are generated according to the preset algorithm. The formula of the preset algorithm is:

[0022] D r =F t (q,rule|m,history)

[0023] Among them, D r represents the generated analysis link graph, F t Indicates the selected big model thinking framework, q indicates the target question input by the user, rule indicates the business rule that matches the target question, m indicates the selected domain big model, and history indicates the historical conversation record;

[0024] After the tasks are decomposed and the links are generated, the operations corresponding to each task are executed in the order of the tasks in the link.

[0025] The commercial real estate digital assistant system based on a large model and an intelligent agent provided by the present invention realizes the understanding and generation of complex parallel tasks through task decomposition and link generation, thereby improving the decomposition accuracy and task completion rate of the digital assistant system for various complex tasks, making the digital assistant system suitable for real application scenarios in commercial real estate.

[0026] In an optional implementation, after the tasks are decomposed and the links are generated, operations corresponding to the tasks are performed in the order of the tasks in the links, including:

[0027] The task decomposition and link generation are in the form of a task description configuration file, which is parsed and formatted according to the preset method of the test script language to obtain a parsed task file;

[0028] Extract the parameters of the parsed task file and use the task flow execution engine to execute the operations corresponding to each task according to the task sequence in the link.

[0029] The commercial real estate digital assistant system based on large models and intelligent agents provided by the present invention unifies the formats of task decomposition and link generation, thereby providing clear guidance for the execution of subsequent tasks, cooperating with the process execution engine, and improving the accuracy of task decomposition and link generation and the efficiency of task execution.

[0030] In an optional implementation, combining the task flow and the execution link to execute the task includes:

[0031] Manually determine whether there are errors in the task flow and execution link. If there are errors in the task flow and execution link, manually supplement the information and regenerate the task flow and execution link until it is manually determined that there are no errors in the task flow and execution link;

[0032] For each task, determine whether human intervention is required. If human intervention is required, wait for human intervention to supplement and confirm the information, and then start executing the corresponding task after human confirmation; if human intervention is not required, execute the corresponding task directly.

[0033] The commercial real estate digital assistant system based on large models and intelligent agents provided by the present invention introduces a manual intervention mechanism in the task disassembly and link generation process to achieve retry of task disassembly and manual assistance in the execution process, thereby alleviating the logical errors or execution deviations caused by the large model illusion, avoiding the gradual accumulation of errors in the previous step of execution, and improving the accuracy of digital assistant disassembly and execution of tasks.

[0034] In an optional implementation, multiple domain models are obtained by training the original data, including:

[0035] Construct various forms of training data based on the original data, and preprocess the training data to obtain training samples;

[0036] Use the open source fine-tuning framework to fine-tune the general large model to obtain the domain large model;

[0037] The training samples are used to train the domain big model to obtain a trained domain big model.

[0038] The commercial real estate digital assistant system based on a large model and an intelligent agent provided by the present invention, through fine-tuning training of a general large model, enables the trained domain large model to better handle problems in the commercial real estate field and enhance the ability to understand knowledge in the commercial real estate field, while maintaining the ability to understand general knowledge and avoiding the phenomenon of "catastrophic forgetting".

[0039] In an optional implementation, different domain models are called according to the target problem, including:

[0040] If the target question type is knowledge retrieval, intelligent question answering, or text generation, the large text model is called to perform knowledge retrieval, answer generation, and task decomposition;

[0041] If the type of the target problem is data query or indicator display, the code-based large model is called to convert the target problem into code and execute it to achieve data or indicator display;

[0042] If the type of the target question is image material in a document, the multimodal large model is called to perform image parsing and format conversion.

[0043] The commercial real estate digital assistant system based on a large model and an intelligent agent provided by the present invention selects different domain large models according to the type of target question, answers questions in a targeted manner, and improves answering efficiency and accuracy.

[0044] In a second aspect, the present invention provides a commercial real estate question-answering method of a commercial real estate digital assistant system based on any one of the first aspects, the method comprising:

[0045] Obtain the target question, and use the preset domain model to divert and identify the intent of the target question to obtain the intent recognition result;

[0046] Based on the intention recognition results, the domain model is called to realize task decomposition and initial link generation. The initial link is optimized by combining the manual intervention mechanism to obtain the execution link of multiple tasks.

[0047] Tasks are executed according to the execution links of multiple tasks, and combined with the manual intervention mechanism, the execution results of each task are obtained and displayed.

[0048] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the method of the above-mentioned second aspect or any corresponding embodiment thereof by executing the computer instructions.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to cause a computer to execute the method of the second aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 is a schematic diagram of the structure of a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of a process of constructing a document knowledge base in a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of the process of disassembling and generating execution links of complex tasks in a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0054] Figure 4 is a flow chart of a manual intervention mechanism in a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0055] Figure 5 is a flow chart of a commercial real estate question-answering method of a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0056] Figure 6 is a schematic diagram of a specific embodiment of a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention;

[0057] Figure 7 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0059] Big model technology is a new model construction, training and application method that has emerged since the development of natural language processing technology. Through the accumulation of massive text (or multimodal) data in advance and the cleaning and processing of these data, the big model adopts an unsupervised model pre-training method to enable itself to have the ability to understand knowledge in multiple fields. With the help of massive training corpus and sophisticated training operations, this understanding ability has shown far superior results to previous small models on multiple benchmark test sets. Subsequently, the supervised fine-tuning (Supervised Fine-Tune, SFT) and reinforcement learning from human feedback (Reinforcement Learning from Human Feedback, RLHF) process of the big model further enhanced the ability to understand and comply with human instructions, as well as the ability to optimize and improve self-cognition from human feedback. From the final performance, the big model technology has shown excellent results in many aspects of natural language processing, including semantic understanding, intent recognition, dialogue question and answer, etc. Therefore, using big model technology for intelligent question and answer has become a common method used by researchers and developers.

[0060] Agent technology or agent technology based on big models is a system construction mode and framework that has emerged with the development and maturity of big models. According to the consensus of the industry, agents use big models as the core engine and should have four capabilities: task decomposition (planning), long-term and short-term memory, tool calling, and action execution. Therefore, compared with the big models that focus more on semantic understanding, knowledge query and intelligent question-answering capabilities, agent technology can use the capabilities of external tools of the model to cope with more complex tasks and execution requirements. Among them, the task decomposition ability of the agent generally comes from the big model itself. The understanding and reasoning ability of the big model can decompose complex tasks into corresponding action steps; the short-term memory ability relies on the model context window, while the long-term memory ability needs to be realized with the help of persistence means such as knowledge base and database; in terms of tool calling, many big models currently have tool calling capabilities such as function call, in addition to which they can also explicitly tell the big model which tool should be called; in terms of action execution, the big model itself and external tools are combined to complete various question-answering, analytical and operational actions. Intelligent agent technology is widely integrated into products such as digital assistants or intelligent question and answer services because it can provide capabilities beyond the scope of large models for more task-based conversations and scenarios by combining with large models.

[0061] Existing related technologies focus on applying intelligent question-answering technology to assist digital assistant systems in answering professional questions in various fields, including real estate marketing. On the other hand, they use intelligent agent technology to introduce tool capabilities to assist in the decomposition and execution of complex tasks. However, this combination is limited to the direct application of technology and has the following disadvantages:

[0062] (1) Existing solutions are not suitable for the commercial real estate sector: The characteristics of the commercial real estate sector determine that simple intelligent question-and-answer functions are not sufficient to meet its diverse knowledge retrieval, data query and analysis and suggestion needs. Although existing related research and patents have considered using domain data to fine-tune the basic big model to construct personalized, domain-specific big models, including applying them in the real estate marketing field, they cannot actually solve the relatively complex needs of commercial real estate in investment promotion, operation and other scenarios through intelligent question-and-answer alone, such as intelligent analysis and suggestions for operating conditions, indicator query of shopping malls, acquisition and execution of internal computing tools, etc.; in addition to relying on the semantic understanding and task decomposition capabilities of the big model, the above capabilities also require the execution of external tools to complete tasks such as indicator query, information acquisition, and action execution. Therefore, it is urgent to introduce intelligent agent technology to enhance the capabilities of digital assistants, and further, it is necessary to fully expand the tool capabilities of the intelligent agent and organically integrate them into the relevant scenarios and needs in the commercial real estate field. However, according to the current survey results, there is a lack of relevant research and solutions that combine intelligent agent technology and big model technology and apply them in the commercial real estate field.

[0063] (2) The supported task decomposition and execution process is too simple: Although there have been relevant studies in the field of agent-enhanced question answering, which have introduced capabilities or tools outside the large model through agent technology to realize the action execution after the complex task is decomposed, such as introducing search engines, ticket queries, report generation, chart drawing and other tools in some complex tasks to assist in the completion of the overall task flow; however, the existing agent-enhanced question answering technology or digital assistant technology actually only considers the serial task flow, that is, the execution of each subtask is connected in a serial manner, and the next subtask only receives the output of the previous subtask and passes the result to the next subtask after processing. However, in the actual scenario of commercial real estate, there is often a parallel relationship between subtasks, so a subtask may need to receive the output of two previous subtasks at the same time, and then output it to the next subtask node after certain logical processing; taking the analysis of the operation of commercial real estate as an example, it may be necessary to query the operating indicators of the plaza and the indicators of the competing plaza at the same time for comparative analysis, so as to assist in judging whether the current operation of the plaza is improving. In this case, the existing agent framework and technology are not sufficient to support the representation and execution of such complex task flows, and new technologies and solutions are urgently needed.

[0064] (3) Lack of human intervention to guide process execution: Another problem with existing technologies is that they completely rely on the semantic understanding and logical reasoning ability of the model during the execution process. For example, the Chain of Thought (CoT) framework is a method that allows large models to think step by step. By adding prompts such as "think step by step", the large model can break down the task by itself, reason and execute step by step. However, the problem is that the inherent illusion of the large model may cause errors in its step-by-step thinking process. Such errors will be transmitted and accumulated, and eventually lead to huge deviations in the entire execution result or answer. Human intervention or the Human-in-Loop method can effectively alleviate the logical errors caused by the illusion of the large model, and can ensure that human feedback is added when errors occur to tell the large model the missing information or the correct execution path. Therefore, a well-designed human intervention mechanism can improve the accuracy of digital assistants in executing tasks, but unfortunately, the existing technology lacks a human intervention process to guide the execution of the process.

[0065] (4) The thinking framework supported and adapted by the intelligent agent is single: the academic and industrial circles currently generally adopt the method of introducing thinking frameworks to strengthen the observation, reflection and reasoning capabilities of large models. In addition to the CoT framework mentioned above, other frameworks such as ToT (Tree of Thought) and ReAct (Reason and Act) can also make the model's thinking process more rigorous, thereby ensuring that the task decomposition and model output are more accurate. However, the thinking framework that the existing intelligent agent technology can support and adapt is often pre-specified, that is, a single thinking framework is preset for the intelligent agent at the beginning of the design, and users are not supported to select and switch to other newer and more advanced frameworks. Therefore, it is necessary to provide a diverse and convenient way to select thinking frameworks during the design and development of digital assistant systems.

[0066] Based on the above problems, an embodiment of the present invention provides a commercial real estate digital assistant system based on a large model and an intelligent agent. By combining the large model and intelligent agent technology to build a digital assistant system for the commercial real estate field, the effect of efficiently processing complex commercial real estate question and answer tasks can be achieved.

[0067] In this embodiment, a commercial real estate digital assistant system based on a large model and an intelligent agent is provided. Figure 1 is a schematic diagram of the structure of a commercial real estate digital assistant system based on a large model and an intelligent agent according to an embodiment of the present invention. Figure 1 As shown, the system includes: a raw data acquisition module, a data processing and knowledge base construction module, a data and tool application module, a task decomposition and process execution module, a domain large model training module, and a domain large model management module.

[0068] like Figure 1 As shown, the original data acquisition module is used to acquire the original data of commercial real estate, and the original data includes: commercial real estate database, company internal document data, and industry knowledge data.

[0069] Specifically, the raw data acquisition module mainly organizes and processes data in the commercial real estate field to prepare for the subsequent digital assistant system capability building and domain model fine-tuning training. The raw data includes commercial real estate databases, document corpora within commercial real estate companies, and commercial real estate industry knowledge materials. The format of the raw data is not limited and can be files, images, etc.

[0070] like Figure 1 As shown, the data processing and knowledge base construction module is used to perform different processing on the original data to build multiple knowledge bases of different types, including: building an indicator knowledge base based on data governance of the commercial real estate database, building a scenario knowledge base based on the company's internal document data, and building a model training sample library based on industry knowledge data.

[0071] Specifically, the original data is processed differently according to its type to build knowledge bases for different scenarios and functions. Multiple knowledge bases serve as the basis for the Retrieval Augmented Generation (RAG) capability and Natural Language to SQL (NL2SQL) query capability. Specifically, the core indicator database is constructed by data governance of the original commercial real estate database and table, including summarization, merging and reconstruction; for internal documents and corpus, a large number of original document files within the commercial real estate company are processed into vector knowledge bases suitable for large model retrieval, matching and processing through document combing, multimodal recognition and analysis, document content segmentation, document content structuring and slice vector construction, so as to build knowledge bases for various scenarios including rules and regulations knowledge base, investment promotion and marketing knowledge base, etc.; in addition, this module also constructs a model training sample library (training data and test data) suitable for the fine-tuning of large models in the commercial real estate field through processing and effective integration of external industry knowledge and materials. In addition to processing structured data such as database tables and unstructured data such as text, images, and voice, it is also possible to collect and encapsulate tool libraries and interface pools for intelligent calls to solve specific problems and tasks in commercial real estate digital assistant systems. For example, when asking about information such as the geographic location of a commercial real estate project, information such as the geographic location and surrounding traffic can be provided by calling map tools or interfaces.

[0072] For example, Figure 2The figure shows the process of building a document knowledge base. It combines multimodal large models, optical character recognition (OCR), document structuring and other technologies to achieve data processing of complex documents in the commercial real estate field. The specific process includes:

[0073] (1) Upload documents in complex formats into the digital assistant system. Supported formats include txt, pdf, word, ppt, and multiple image formats.

[0074] (2) Use the image understanding capabilities of the multimodal large model and combine it with the OCR model to identify, parse, and process document materials in formats such as images;

[0075] (3) Call the document structuring algorithm to perform structural recognition on the entire text and distinguish the directory structure, etc.

[0076] (4) Use a custom algorithm to slice the document content to ensure more accurate recall and recognition in the future.

[0077] (5) The sliced ​​results are processed using a vectorization algorithm to build a domain document knowledge base.

[0078] like Figure 1 As shown, the data and tool application module is used to query the corresponding knowledge base or call the tool according to the target question, generate the answer to the target question and display it.

[0079] Specifically, the data and tool application module provides the digital assistant system with scenario-based intelligent question and answer, data query, analysis suggestions and tool usage capabilities, which are the key basic capabilities that the digital assistant system should have as an intelligent entity. Based on the actual needs of the commercial real estate field, it provides basic query, reply and visual display functions.

[0080] like Figure 1 As shown, the task decomposition and process execution module is used to decompose the target problem into multiple task processes, generate execution links, and execute tasks in combination with the task processes and execution links.

[0081] Specifically, the user's target problem division instructions are decomposed into complex task processes, execution links are generated, and the workflow execution module is combined to execute each complex task according to the link.

[0082] like Figure 1 As shown, the domain large model training module is used to obtain multiple domain large models based on the original data.

[0083] Specifically, the main function of the domain big model is to realize intention understanding, task decomposition and task execution according to the target problem, and finally interact with users through intelligent question and answer, tool call or data display. The domain big model is obtained by fine-tuning and training with relevant knowledge in the commercial real estate field.

[0084] The domain large model training module uses different types of raw data to train multiple domain large models. The training process mainly includes domain fine-tuning data construction, model fine-tuning training and model effect testing.

[0085] like Figure 1 As shown, the domain big model management module is used to store multiple domain big models and call different domain big models according to the target problem.

[0086] Specifically, the models trained by the domain big model training module will be placed in the domain big model resource pool for dedicated storage and call. The domain big model management module includes the domain big model resource pool, and different models can be called according to the scenarios and actual needs in the commercial real estate field.

[0087] The commercial real estate digital assistant system based on big models and intelligent agents provided in this embodiment fine-tunes the training domain big model based on relevant application scenarios and industry knowledge in the commercial real estate field, and combines big models and intelligent agent technology to build a digital assistant system for the commercial real estate field. The system can communicate with users through question-and-answer interactions and better cope with complex tasks and needs in the commercial real estate field.

[0088] In some optional implementations, querying a corresponding knowledge base or calling a tool according to the target question, generating an answer to the target question and displaying it includes:

[0089] ,By converting natural language queries into structured query language, dynamic ,query and visual display of the indicator knowledge base are ,realized.

[0090] Specifically, based on NL2SQL, natural language queries (Natural Language) are converted into structured query language (Structured Query Language, SQL), and data query and visualization are performed based on SQL to achieve dynamic query and visualization of core indicator data.

[0091] Based on the internal knowledge retrieval and query generated by retrieval enhancement, the query of the scene knowledge base is realized.

[0092] Specifically, based on the internal knowledge retrieval and query of RAG, the query and reply of internal rules and regulations, investment promotion and marketing knowledge, etc. The specific query method is a mature existing technology and will not be described here.

[0093] By calling tools or interfaces, external resources are combined with the knowledge base to achieve intelligent question and answer.

[0094] Specifically, based on the generation and execution of intelligent agent-based analysis and suggestion processes, users are provided with reference suggestions on investment attraction and operation. At the same time, tool calling and interface calling functions are combined with large models through methods such as Function Call, and external resources and knowledge bases are used as retrieval and query objects to achieve more accurate intelligent question and answer.

[0095] The commercial real estate digital assistant system based on a large model and an intelligent agent provided in this embodiment uses different retrieval methods to retrieve the corresponding knowledge base according to the target question, and visualizes the query results to meet the needs of basic commercial real estate queries, supports calling tools or interfaces, and improves the flexibility of the digital assistant system.

[0096] In some optional implementations, the target problem is decomposed into multiple task flows, an execution link is generated, and the task is executed in combination with the task flow and the execution link, including:

[0097] Step a1: Combine historical conversation data to rewrite and filter the target question to obtain complete semantic content.

[0098] Specifically, Figure 3 As shown in the figure, it is a schematic diagram of the process of disassembling and generating execution links of complex tasks. Question rewriting and filtering: Since the questions and instructions input by users may be vague in meaning and may contain referential words in multiple rounds of dialogue, it is necessary to combine the historical dialogue database to substitute and rewrite the information of multiple rounds before task disassembly, and rewrite and expand referential words and descriptions such as "it", "the square", "then" into complete semantic expressions. The process of introducing historical dialogue data is: according to the preset rounds or the context window of the large model, the historical dialogue text content of appropriate length is selected and brought into the large model.

[0099] Step a2: The target question is divided and identified according to its complete semantic content, the user's intent is determined, and the task type of the target question is judged according to the user's intent.

[0100] Specifically, the purpose of diverting and identifying intents is that in the application scenarios of commercial real estate digital assistant systems, user intentions are diverse, which may be single-node knowledge retrieval, data query, tool usage and other tasks or simple chatting, or they may be operational analysis suggestions and research report generation involving multiple nodes and complex logic. Therefore, fine-grained identification and division of the intent of the target question based on the complete semantic content will help provide users with clearer answer outputs.

[0101] Step a3: If the task type of the target problem is a multi-node task, the target problem is decomposed into tasks and links are generated according to a preset algorithm. The formula of the preset algorithm is:

[0102] D r =F t (q,rule|m,history)

[0103] Among them, D r represents the generated analysis link graph, F t It represents the selected big model thinking framework, q represents the target question input by the user, rule represents the business rule matching the target question, m represents the selected domain big model, and history represents the historical conversation records.

[0104] Specifically, if the result of the intention diversion and identification in step a2 is a single-node task, it is directly output in the form of a "single process node". If it involves a complex task, it enters the "multi-node task disassembly link", such as Figure 3 As shown in the figure, depending on the complexity of the task, it may include serial tasks or parallel tasks. By introducing additional parallel task prompts (Few-Shot) and disassembly templates in the prompt words of the large model, examples of similar analysis links are given in the business rule library to implement the disassembly and process generation algorithm of parallel tasks. In addition, since the existing intelligent agent support and adaptation framework is relatively simple, the multi-node task disassembly part of this module supports the dynamic adaptation of the large model thinking framework. The formulaic expression of the preset algorithm for task disassembly and link generation is:

[0105] D r =F t (q,rule | m,history) (1)

[0106] Among them, D r Represents the analysis link graph generated by this module (usually a directed acyclic graph); F t It indicates the selected big model thinking framework (for example, the chain of thought (CoT), reasoning execution (ReAct) or reflection (Reflexion) method, which can support self-selection when understanding and decomposing tasks); q indicates the target question (query) input by the user; rule indicates the business rule matching the target question; m indicates the selected text big model (generally a domain big model that has been fine-tuned and trained, with better business understanding capabilities); history indicates historical conversation records.

[0107] Step a4: After task decomposition and link generation, the operations corresponding to each task are executed in the order of tasks in the link.

[0108] Specifically, after generating the results and links of task decomposition, multiple subtasks of the complex task decomposition and the execution order of each subtask are determined, and the operations corresponding to each task are executed according to the task order in the link to complete the answer to the target question.

[0109] The commercial real estate digital assistant system based on a large model and an intelligent agent provided in this embodiment realizes the understanding and generation of complex parallel tasks through task decomposition and link generation, thereby improving the digital assistant system's decomposition accuracy and task completion rate for various complex tasks, making the digital assistant system suitable for real application scenarios in commercial real estate.

[0110] In some optional implementations, after the tasks are decomposed and the links are generated, the operations corresponding to the tasks are performed in the order of the tasks in the links, including:

[0111] Step a41, the task decomposition and link generation are in the form of a task description configuration file, and the task description configuration file is parsed and formatted according to a preset method of the test script language to obtain a parsed task file.

[0112] Specifically, the decomposition results of single-node tasks and multi-node tasks are in the form of task description configuration files, which facilitate the unified understanding and format conversion of the subsequent digital assistant system. The task description configuration files are described in the form of a unique task-specific language (TSL), which explicitly represents the type and input parameters of each task node, thereby providing clear guidance for the execution of subsequent tasks.

[0113] Parse and convert the format according to the TSL conventions to cooperate with the process engine of the digital assistant system. At the same time, extract the parameters contained in the task description configuration file and push them together with other additional parameters to the process engine for execution.

[0114] Step a42, extracting and parsing the parameters of the task file, and executing the operations corresponding to each task using the task flow execution engine according to the task sequence in the link.

[0115] Specifically, after receiving the task execution instructions and related parameters, the task process execution engine executes related operations in the order of the task link. If some task nodes need to call tools or interfaces during the execution process, appropriate tools or interfaces are selected from the preset tool or interface pool for calling. The tool call is triggered under appropriate instructions to implement the tool call (Function Call) function based on the large model, and the parameters generated in the TSL are passed to the tool or interface to complete the task execution.

[0116] The commercial real estate digital assistant system based on large models and intelligent agents provided in this embodiment unifies the formats of task decomposition and link generation to provide clear guidance for the execution of subsequent tasks, cooperates with the process execution engine, and improves the accuracy of task decomposition and link generation as well as the efficiency of task execution.

[0117] In some optional implementations, combining the task flow and the execution link to execute the task includes:

[0118] Step a5, manually determine whether there are errors in the task flow and execution link. If there are errors in the task flow and execution link, manually supplement the information and regenerate the task flow and execution link until it is manually determined that there are no errors in the task flow and execution link.

[0119] Specifically, Figure 4 As shown in the figure, the human-in-loop mechanism mainly plays a role in the two links of task decomposition and task execution. After realizing task decomposition and link generation based on the capabilities of the large model and intelligent agent, a "user confirmation / regeneration" mechanism is added to the digital assistant system. The user can manually check whether the generated task link is correct. If the generated task link is correct, the user can confirm that the task link is correct and continue to execute the task; otherwise, the execution link of the task can be regenerated after supplementing the relevant information in the human-computer interaction page. The human-in-loop mechanism helps to further clarify the user's intentions and operation instructions, and also avoids the inaccurate task execution link caused by the large model illusion, further improving the accuracy of task decomposition and link generation.

[0120] Step a6, for each task, determine whether manual intervention is required. If manual intervention is required, wait for manual information supplement and confirmation until manual confirmation is made, and then start executing the corresponding task. If manual intervention is not required, directly execute the corresponding task.

[0121] Specifically, after the task description configuration file is generated and parsed by the digital assistant system, the entire task process (serial process or parallel process) will be executed in sequence, with a time interval between each subtask execution ( Figure 3 Taking the first subtask and the second subtask as an example), a judgment mechanism for user intervention is designed, so that the workflow generated by the task decomposition module is expressed in the following formula at each specific step:

[0122]

[0123] Among them, A n Indicates the execution action (Action) or task description of the current nth subtask. The execution action or task description configuration text will be converted into TSL format and passed to the workflow engine for execution; A n-1and R n-1 They represent the execution action and execution result of the n-1th subtask node respectively. The execution of the current nth subtask needs to rely on the result of the previous action. For example, the analysis of a problem in commercial real estate needs to rely on the result of the previous data query. n It indicates the additional information added to the current subtask after the user intervenes. If the user finds errors or deviations in the execution of the current subtask, he can add additional information or clarify the intention to make the task execution more accurate. init Indicates the user's initial task instruction, which will also serve as a reference for the current subtask execution; It represents the execution function of each subtask node under the specified large model m and thinking framework t. For example, if it is a knowledge retrieval node, it represents the function from the query statement or the final result. If it is a data query node, it represents the node from the query statement to the query code generation to the final code execution to obtain the data result. This is only an example, but not limited to this.

[0124] In addition to introducing the information supplement link, the user intervention mechanism also designs a process waiting mechanism to ensure the balance between the two needs of automatic execution of the task process and timely user intervention in the interaction method. That is, on the one hand, it ensures that users can intervene in time at the link where intervention is required, and on the other hand, it ensures that users will not be disturbed too much, and the generated task process can be automatically executed without the user's operation.

[0125] The commercial real estate digital assistant system based on large models and intelligent agents provided in this embodiment introduces a manual intervention mechanism in the task disassembly and link generation process to achieve retry of task disassembly and manual assistance in the execution process, thereby alleviating the logical errors or execution deviations caused by the large model illusion, avoiding the gradual accumulation of errors in the previous step of execution, and improving the accuracy of digital assistant disassembly and execution of tasks.

[0126] In some optional implementations, multiple domain models are trained based on the original data, including:

[0127] Various forms of training data are constructed based on the original data, and the training data are preprocessed to obtain training samples.

[0128] Specifically, various forms of training data such as question-answer pairs and extraction tasks are constructed based on the existing data, documents and various materials in the commercial real estate field, and are manually screened and verified to eventually become high-quality training samples that can be used for model fine-tuning.

[0129] Use the open source fine-tuning framework to fine-tune the general large model to obtain the domain large model.

[0130] Specifically, model fine-tuning training uses an open source fine-tuning framework (such as Lora, etc.) to fine-tune the general large model to obtain a proprietary model that can understand the meaning of domain knowledge, which serves as the basis of the domain large model.

[0131] The training samples are used to train the domain big model to obtain a trained domain big model.

[0132] Specifically, the domain model is trained using training samples. After the trained domain model is obtained, it is necessary to evaluate the domain model in terms of domain capabilities and general capabilities. The fine-tuned model needs to enhance the ability to understand domain knowledge while maintaining the ability to understand general knowledge as much as possible, so as to avoid the phenomenon of "catastrophic forgetting".

[0133] The commercial real estate digital assistant system based on a large model and an intelligent agent provided in this embodiment, through fine-tuning training of the general large model, enables the trained domain large model to better handle problems in the commercial real estate field and enhance the ability to understand knowledge in the commercial real estate field, while maintaining the ability to understand general knowledge and avoiding the phenomenon of "catastrophic forgetting".

[0134] In some optional implementations, different domain models are called according to the target problem, including:

[0135] If the target question type is knowledge retrieval, intelligent question and answer, or text generation, the large text model is called to perform knowledge retrieval, answer generation, and task decomposition.

[0136] If the type of the target problem is data query or indicator display, the code-based large model is called to convert the target problem into code and execute it to achieve data or indicator display.

[0137] If the type of the target question is image material in a document, the multimodal large model is called to perform image parsing and format conversion.

[0138] Specifically, a variety of different models are stored in the domain big model resource pool. Different domain big models are called from the resource pool according to the type of target question, and the most appropriate domain big model is called according to the scenario and actual needs in the commercial real estate field. Capabilities such as knowledge retrieval, intelligent question and answer, and copywriting generation mainly rely on the capabilities of text-based big models to achieve accurate intent understanding, knowledge retrieval, answer generation, and task decomposition. In terms of data query and indicator display, the code-based big model is called to convert the user's query intent and instructions into code (such as SQL, Python), and then the code execution module is called to execute the code to obtain the corresponding data, or to display data charts through code. For materials such as pictures in documents, it is necessary to call the capabilities of multimodal big models to perform image parsing and format conversion, so as to assist subsequent intelligent question and answer functions. Models in the big model resource pool, such as text-based big models and code-based big models, need to be fine-tuned and trained to enhance their ability to understand domain knowledge.

[0139] The commercial real estate digital assistant system based on big models and intelligent agents provided in this embodiment selects different domain big models according to the type of target questions, answers questions in a targeted manner, and improves answering efficiency and accuracy.

[0140] In this embodiment, a commercial real estate question-answering method based on a commercial real estate digital assistant system in any of the previous embodiments is provided, which can be used in the above-mentioned commercial real estate digital assistant system or computer system. Figure 5 is a flow chart of a commercial real estate question-and-answer method according to an embodiment of the present invention. Figure 5 As shown, the process includes the following steps:

[0141] Step S101, obtain the target question, and use the preset domain model to perform intent diversion and identification on the target question to obtain the intent identification result.

[0142] Specifically, the function realization of the commercial real estate digital assistant system relies on the interactive front-end, service back-end, and algorithm capability end. The interactive front-end realizes the front-end page operation for dialogue and visual interaction with users; the service back-end undertakes the middle-layer service encapsulation of algorithm capabilities and interactive front-end, and uses the workflow engine to handle the execution of complex tasks in the commercial real estate field; the algorithm capability end relies on the text big model and code big model fine-tuned by the domain corpus to provide different algorithm capabilities. The target question input by the user is obtained through the interactive front-end, and the intention recognition capability of the algorithm capability end is used to realize the intention diversion and recognition of the target question, and obtain the intention recognition result.

[0143] Step S102, calling the domain big model according to the intention recognition result, realizing task decomposition and initial link generation, optimizing the initial link in combination with the manual intervention mechanism, and obtaining the execution link of multiple tasks.

[0144] Specifically, the service backend calls the appropriate domain model from the algorithm capability end based on the intent recognition results, performs task decomposition and process generation, obtains the initial link of multiple tasks, and combines the manual intervention mechanism to add user confirmation or regenerate manual operations. If the generated initial link is correct, the user can confirm the execution; otherwise, the execution link of the corresponding task can be regenerated after supplementing the relevant information in the interactive page.

[0145] Step S103, executing tasks according to the execution links of multiple tasks, combining with the manual intervention mechanism, obtaining the execution results of each task and displaying them.

[0146] Specifically, the task execution is driven by the workflow engine, and the knowledge retrieval capability and data query capability of the algorithm capability end are called according to the needs during the execution process. This is only an example, but not limited to this. In the process of process generation, the judgment of whether manual intervention is needed is added to improve the flexibility and accuracy of complex task disassembly and process generation.

[0147] The commercial real estate digital assistant system and question-answering method based on big models and intelligent agents provided in this embodiment draw on the method of fine-tuning and training big models with domain data in the prior art to construct a text big model and a code big model for commercial real estate as the basis of the entire digital assistant system; on this basis, after data cleaning, governance, parsing and screening, core capability components such as databases, knowledge bases and capability tool pools suitable for knowledge retrieval, data query, analysis reports, chart presentation and tool calling are constructed; the core of the digital assistant system is the task decomposition and execution engine implemented based on big models and intelligent agent technology, which can decompose user questions and instructions into task links executed by sub-tasks, and complete task execution sequentially with the help of the workflow engine, and finally give answers that meet user needs, or complete specific tasks in accordance with user instructions.

[0148] In particular, in the commercial real estate digital assistant system, the task decomposition engine can accurately understand user intentions and instructions, and plan and decompose complex tasks. By separating the intention recognition module, introducing the business rule library, and decoupling the thinking framework module, it supports parallel process tasks and can accurately distinguish between single-node tasks and multi-node tasks, making task decomposition more accurate. It will not generate overly lengthy and repetitive task links due to the illusion of large models and the defect of easy over-exertion. It can also accurately refer to business rules to provide reasonable, strict and commercial real estate business cognition-compliant operating procedures, which significantly improves the accuracy of process decomposition and task generation.

[0149] At the same time, a manual intervention mechanism is introduced during the task link generation and task execution stages to ensure that each step of the action is under manual supervision. When necessary, more information can be added or retried manually to allow the task to be executed more smoothly and in line with expectations. On the one hand, it solves the problem of error accumulation and amplification caused by an error in a certain step due to fully automatic execution of existing technologies. On the other hand, it avoids causing too much disturbance to users through reasonable design, and only intervenes to take over the process or provide guidance when necessary. The digital assistant system with the added manual intervention mechanism has significantly improved the completion rate of complex tasks.

[0150] The digital assistant system for commercial real estate provided in this embodiment is more field-specific, has more accurate problem understanding, supports a wider range of tasks, has more precise task planning, and has a higher degree of task execution completion.

[0151] In a specific embodiment, the service backend is mainly used as a carrier for function implementation and data flow, and the service backend is used as the execution subject, such as Figure 6 As shown, the working process of the commercial real estate digital assistant system includes:

[0152] (1) The interactive front end receives the target question input by the user, and the service back end rewrites and encapsulates the question. The specific implementation functions include: filtering the conversation records according to the conversation rounds and adding them to the user's target question query, etc.

[0153] (2) The service backend calls the intent recognition capability of the large model to perform intent diversion and recognition. In terms of specific implementation, the present invention adopts large model technology for intent recognition and classification. Other implementation methods can also use traditional natural language processing (Natural Language Process, NLP) small-size models, such as text classification models based on bidirectional encoder representation models (Bidirectional Encoder Representations from Transformers, BERT) or convolutional neural networks to realize intent recognition.

[0154] (3) Based on the intention recognition results, the task understanding, planning and decomposition capabilities of the large model are called upon, and a specific thinking framework is combined to realize the decomposition and process generation of complex tasks in the commercial real estate field. The decomposed process is displayed on the interactive front end, and the user confirms the generated task link or regenerates it if there is a problem. The current task understanding and planning capabilities rely on the logical reasoning and task planning capabilities of the large text model. For specific model selection, relatively large-size models such as Llama and Qwen can be selected, or closed-source models such as GPT can be selected when permitted.

[0155] (4) After the algorithm outputs the task analysis and link generation results, the service backend needs to read and parse the task flow file according to the agreed format, such as the TSL language mentioned above, and convert it into content that can be recognized by the workflow execution engine.

[0156] (5) The service backend is based on the workflow execution engine and executes tasks according to the specified task chain. Each step of the execution will visualize the execution process and results to the user on the interactive front end, and the user can view the execution results and decide whether human intervention (Human-in-Loop) is required to assist the execution of the process. At the same time, considering the needs of the commercial real estate field, the tasks that may be included in the task execution stage, such as asking questions about investment promotion, operation or internal corporate rules and regulations, need to rely on knowledge retrieval, data query and other capabilities to answer; asking questions about core indicators and other issues requires the use of data query capabilities to query data tables; asking for analysis of operating conditions and improvement suggestions requires the use of analysis summary and logical reasoning capabilities; asking for internal information that cannot be obtained by the big model may require the use of tools or interfaces to answer; the realization of the above capabilities relies on the text big model or code big model after fine-tuning the domain, among which the tool calling capability is realized by the big model's Function Call capability, and the choice of workflow engine can be connected and modified by the open source workflow engine.

[0157] (6) The results of each execution step will eventually be formatted, such as in the form of a chart, and then displayed on the interactive front end. Users can interact with the displayed results and provide feedback on the results output by the digital assistant system, such as how to correct the deviations in the execution.

[0158] (7) After the previous round of conversation is completed, the user can continue to enter questions on the interactive front end to start the next round of conversation. The service back end needs to process the historical conversation information and then continue to start the next round of question rewriting and packaging.

[0159] The technical implementation in the specific embodiment is only an example, but not limited to this.

[0160] The embodiment of the present invention also provides a computer device having the above Figure 1 The commercial real estate digital assistant system based on large models and intelligent agents is shown.

[0161] See also Figure 7 , Figure 7 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 7As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 7 A processor 10 is taken as an example.

[0162] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0163] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0164] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0165] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0166] The computer device further comprises a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0167] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0168] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A commercial real estate digital assistant system based on a large model and an intelligent agent, characterized in that: The system includes: a raw data acquisition module, a data processing and knowledge base construction module, a data and tool application module, a task decomposition and process execution module, a domain large model training module, and a domain large model management module, wherein: The original data acquisition module is used to acquire the original data of commercial real estate, and the original data includes: commercial real estate database, company internal document data, and industry knowledge data; The data processing and knowledge base construction module is used to perform different processing on the raw data to construct different types of multiple knowledge bases, including: constructing an indicator knowledge base based on data governance of the commercial real estate database, constructing a scenario knowledge base based on the company's internal document data, and constructing a model training sample library based on industry knowledge data; The data and tool application module is used to query the corresponding knowledge base or call the tool according to the target question, generate the answer to the target question and display it; The task decomposition and process execution module is used to decompose the target problem into multiple task processes, generate execution links, and execute the task in combination with the task processes and execution links; The domain large model training module is used to train multiple domain large models based on the original data; The domain big model management module is used to store multiple domain big models and call different domain big models according to the target problem.

2. The system according to claim 1, characterized in that According to the target question, query the corresponding knowledge base or call the tool to generate the answer to the target question and display it including: By converting natural language queries into structured query language, dynamic query and visual display of indicator knowledge base can be achieved; Based on the internal knowledge retrieval and query generated by retrieval enhancement, the query of the scene knowledge base is realized; By calling tools or interfaces, external resources are combined with the knowledge base to achieve intelligent question and answer.

3. The system according to claim 1, characterized in that Decompose the target problem into multiple task flows, generate execution links, and execute tasks in combination with the task flows and execution links, including: Combined with historical conversation data, the target question is rewritten and filtered to obtain complete semantic content; Divide and identify the intention of the target question according to the complete semantic content, determine the user intention, and judge the task type of the target question according to the user intention; If the task type of the target problem is a multi-node task, the target problem is decomposed and links are generated according to the preset algorithm. The formula of the preset algorithm is: D r =F t (q,rule|m,history) Among them, D r represents the generated analysis link graph, F t Indicates the selected big model thinking framework, q indicates the target question input by the user, rule indicates the business rule that matches the target question, m indicates the selected domain big model, and history indicates the historical conversation record; After the tasks are decomposed and the links are generated, the operations corresponding to each task are executed in the order of the tasks in the link.

4. The system according to claim 3, characterized in that After the tasks are broken down and the links are generated, the operations corresponding to each task are performed in the order of the tasks in the link, including: The task decomposition and link generation are in the form of a task description configuration file, which is parsed and format-converted according to a preset method of the test script language to obtain a parsed task file; Extract the parameters of the parsed task file, and use the task flow execution engine to execute operations corresponding to each task according to the task sequence in the link.

5. The system according to claim 1 or 3, characterized in that: Executing tasks in combination with the task flow and execution link includes: Manually determine whether there are errors in the task flow and execution link. If there are errors in the task flow and execution link, manually supplement the information and regenerate the task flow and execution link until it is manually determined that there are no errors in the task flow and execution link. For each task, determine whether human intervention is required. If human intervention is required, wait for human intervention to supplement and confirm the information, and then start executing the corresponding task after human confirmation; if human intervention is not required, execute the corresponding task directly.

6. The system according to claim 1, characterized in that Based on the original data, we train large models in multiple fields, including: Constructing various forms of training data based on the original data, and preprocessing the training data to obtain training samples; Use the open source fine-tuning framework to fine-tune the general large model to obtain the domain large model; The domain big model is trained using the training samples to obtain a trained domain big model.

7. The system according to claim 1, characterized in that Different domain models are called according to the target problem, including: If the target question type is knowledge retrieval, intelligent question answering, or text generation, the large text model is called to perform knowledge retrieval, answer generation, and task decomposition; If the type of the target problem is data query or indicator display, the code-based large model is called to convert the target problem into code and execute it to achieve data or indicator display; If the type of the target question is image material in a document, the multimodal large model is called to perform image parsing and format conversion.

8. A commercial real estate question-answering method based on the commercial real estate digital assistant system according to any one of claims 1 to 7, characterized in that: The method comprises: Obtain the target question, and use the preset domain model to perform intent diversion and recognition on the target question to obtain the intent recognition result; According to the intention recognition result, the domain big model is called to realize task decomposition and initial link generation, and the initial link is optimized by combining the manual intervention mechanism to obtain the execution link of multiple tasks; Tasks are executed according to the execution links of the multiple tasks, and the execution results of each task are obtained and displayed in combination with a manual intervention mechanism.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method of claim 8 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method described in claim 8.

Images