Insurance customer portrait generation method and system based on large model and multi-agent cooperation

Abstract

The invention provides an insurance customer portrait generation method and system based on large model and multi-agent cooperation, and relates to the technical field of insurance data processing. According to the method, multi-source heterogeneous data such as customer insurance policies and communication records are collected through an information acquisition module, then customer key information is processed in parallel by special agents in a multi-agent collaborative analysis module, and a special analysis result is generated by combining a long text block processing technology and module retrieval field knowledge fusion correction. And finally, the planning agent integrates results to generate a structured customer portrait report. According to the method, multi-dimensional requirements of customers can be mined, guarantee gaps can be accurately identified, available portrait reports can be efficiently output, a targeted follow-up basis is provided for insurance brokers, and the accuracy of insurance services and the customer adaptation degree are improved.

Description

Technical Field

[0001] This application relates to the field of insurance data processing technology, and in particular to a method and system for generating insurance customer profiles based on large models and multi-agent collaboration. Background Technology

[0002] In the insurance industry, accurately understanding customer needs is crucial for providing suitable insurance solutions. Traditional methods rely on insurance brokers' manual recording and analysis, which is inefficient and prone to missing key information. With the development of artificial intelligence technology, especially the rise of large language models (LLM), new possibilities have emerged for automating the processing of customer service information.

[0003] Currently, traditional user profiling relies on manual analysis and compilation by human insurance planners. While highly professional and accurate, this approach is inefficient and prone to missing crucial information. Some systems utilize natural language summarization, employing models (such as T5 and LLM) for prompts and training, but these methods result in limited profiles that fail to comprehensively reflect customer characteristics. T5, for example, is a pre-trained language model based on the Transformer architecture. Existing technologies have attempted to use natural language processing to analyze customer information, but comprehensive and systematic representation of user information during insurance activities is challenging. Specifically, the lack of specialized processing mechanisms tailored to the insurance industry hinders the effective integration of customer information from multiple channels. Therefore, this technical challenge urgently needs to be addressed. Summary of the Invention

[0004] In view of the above problems, this application is proposed to provide a method and system for generating insurance customer profiles based on large models and multi-agent collaboration to overcome or at least partially solve the above problems. The technical solution is as follows: Firstly, a method for generating insurance customer profiles based on large models and multi-agent collaboration is provided, the method comprising: Acquire unstructured, multi-source, heterogeneous data related to target customers; The system calls a pre-defined large model, which extracts structured target customer information from unstructured multi-source heterogeneous data. The pre-built planning agent is invoked, which decomposes the target customer information into multiple dimensions and generates corresponding profile analysis tasks for each analysis dimension. The profile analysis tasks are then assigned to the corresponding specialized agent, which processes the tasks to obtain the corresponding specialized analysis results. The planning agent generates a target customer profile report based on the specific analysis results returned by each specific agent.

[0005] In one possible implementation, the multi-source heterogeneous data includes one or more of the following: data related to purchased insurance, internet behavior trajectory records, insurance service communication and interaction data, historical profile descriptions, customer data with service personnel notes, insurance product recommendation schemes, and customer source association data.

[0006] In one possible implementation, a large model extracts structured target customer information from unstructured, multi-source heterogeneous data, including: The large model performs logical reasoning on multi-source heterogeneous data based on pre-set thought chain prompts; By combining a preset number of sample examples, the large model can follow the rules in the sample examples to accurately extract customer information; the sample examples include the correspondence between unstructured text and structured information and the extraction rules. Based on the preset output format requirements, the extracted customer information is output as structured target customer information; the output format requirements include field format, dimension classification and presentation specifications.

[0007] In one possible implementation, the specialized intelligent agent includes one or more of the following: family structure analysis intelligent agent, purchased insurance analysis intelligent agent, insurance type intention analysis intelligent agent, health status analysis intelligent agent, and sales reminder analysis intelligent agent.

[0008] In one possible implementation, the specialized intelligent agent performs task processing to obtain the corresponding specialized analysis results, including: Receive profile analysis tasks assigned by the planning intelligent agent for the corresponding dimension, and obtain data related to the profile analysis tasks from the structured target customer information; To meet the analytical needs of the profiling task, key information is extracted from the relevant data and content that is valuable for completing the profiling task is selected. The extracted key information is analyzed to form specific analysis results that conform to the analysis logic of the corresponding dimension, and the specific analysis results are returned to the planning agent.

[0009] In one possible implementation, a planning agent decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and delegates the profile analysis task to a specialized agent for that dimension. This specialized agent then processes the task to obtain the corresponding specialized analysis results, including: Before the planning agent decomposes the target customer information into multiple dimensions, it determines whether the length of the target customer information exceeds the context window of the large model. If it does, the target customer information is processed as long text and then segmented into long text blocks. The target customer information is obtained through a fact correction step. Before the specialized intelligent agent performs the task processing, it is determined whether the length of the profile analysis task exceeds the context window of the large model. If it does, the profile analysis task is treated as a long text and processed into long text blocks. The corresponding analysis results are obtained through the fact correction step.

[0010] In one possible implementation, long text chunking includes: Obtain the long text to be processed, and split the long text into multiple text blocks according to the context window length of the large model, while maintaining a preset content overlap between adjacent text blocks; Each text block is processed in parallel to obtain the local structured information corresponding to each text block. The processing results of each text block are summarized and integrated to obtain the first target information.

[0011] In one possible implementation, the fact-correction steps include: For the first target information obtained by long text segmentation, retrieve the vector knowledge base that stores relevant domain knowledge, object information and historical data to obtain reference information related to the first target information; Hallucination detection is performed on the first target information, and the first target information is compared with the retrieved reference information to identify and mark the content that exists in the first target information but is not supported by the reference information; Error correction is performed on the first target information. For statements in the first target information that contradict the reference information, replacement or supplementation is made based on the accurate content in the reference information to obtain the corrected final result.

[0012] Secondly, an insurance customer profile generation system based on large models and multi-agent collaboration is provided, the system comprising: The information acquisition module is used to acquire unstructured multi-source heterogeneous data related to the target customer, call the preset large model, and extract structured target customer information from the unstructured multi-source heterogeneous data. The multi-agent collaborative analysis module is used to call a pre-built planning agent, which decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and assigns the profile analysis task to the corresponding special agent for the dimension. The special agent then processes the task to obtain the corresponding analysis results. The output module is used to generate target customer profile reports by the planning agent based on the special analysis results returned by each special agent.

[0013] Thirdly, an electronic device is provided, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the insurance customer profile generation method based on large models and multi-agent collaboration as described above.

[0014] By employing the above technical solutions, the insurance customer profile generation method and system based on large models and multi-agent collaboration provided in this application first collects heterogeneous data from multiple sources, such as customer policies and communication records, through an information acquisition module. Then, each specialized agent in the multi-agent collaborative analysis module processes key customer information in parallel. Combining long text segmentation processing technology and module retrieval domain knowledge fusion correction, specialized analysis results are generated. Finally, the planning agents integrate the results to generate a structured customer profile report. This method can uncover multi-dimensional customer needs, accurately identify protection gaps, and efficiently output usable profile reports, providing insurance brokers with targeted follow-up basis and improving the accuracy and customer fit of insurance services. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0016] Figure 1 The flowchart illustrates the insurance customer profile generation method based on large model and multi-agent collaboration provided in an embodiment of this application; Figure 2 The flowchart of the data structure extraction process of the insurance customer profile generation method based on large model and multi-agent collaboration provided in the embodiments of this application is shown. Figure 3 A flowchart illustrating the special intelligent agent processing provided in an embodiment of this application is shown; Figure 4 A flowchart illustrating the long text judgment and processing provided in an embodiment of this application is shown; Figure 5 This diagram illustrates the structure of the insurance customer profile generation system based on large models and multi-agent collaboration provided in an embodiment of this application. Figure 6 This paper shows a structural diagram of an insurance customer profile generation system based on a large model and multi-agent collaboration, according to another embodiment of this application. Figure 7 A structural diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0017] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0019] In existing technologies, some systems have attempted to use natural language processing (NLP) to analyze customer information, but the following problems still exist: 1. User information in insurance activities is difficult to display comprehensively and systematically. Specifically, it is manifested in the lack of a professional processing mechanism tailored to the characteristics of the insurance field, which makes it impossible to effectively integrate customer information from multiple channels.

[0020] 2. The lack of a systematic approach to artificial intelligence is specifically reflected in: the lack of a multi-agent collaboration mechanism, resulting in limited processing efficiency and accuracy; and the lack of a state diagram-based workflow management mechanism, which makes it impossible to effectively coordinate multiple analysis tasks.

[0021] 3. Efficiency and accuracy are difficult to meet professional requirements, specifically manifested in: the lack of an external knowledge integration mechanism based on search-enhanced generation (RAG), resulting in limited analytical accuracy; and the lack of a mechanism for processing large-scale user data, leading to low efficiency in processing large-scale customer data.

[0022] The inventors discovered through analysis that traditional user profiling relies on manual analysis and organization by human insurance planners, which is highly professional and accurate but inefficient and prone to missing key information. Some systems use natural language summarization and model-based analysis with prompts and training methods, but the profiles are too simplistic and cannot fully reflect customer characteristics.

[0023] To address the aforementioned technical problems, this application provides a method for generating insurance customer profiles based on large models and multi-agent collaboration. This method integrates multiple customer information sources and utilizes large language models, multi-agent technology, and retrieval-enhanced generation techniques to automatically generate comprehensive and accurate insurance customer profiles. Figure 1 As shown, the insurance customer profile generation method based on large models and multi-agent collaboration may include the following steps S101 to S104: Step S101: Obtain unstructured, multi-source heterogeneous data related to the target customer.

[0024] In this step, multi-source heterogeneous data refers to various unstructured data related to the entire insurance service chain for the target customer. Specifically, it may include, but is not limited to, one or more of the following: data related to purchased insurance, such as historical policy terms, premium payment records, and policy change application documents; internet behavior records, such as browsing logs on insurance product websites, online consultation conversation records, interaction content on insurance-related topics on social media, and data on dwell time and clicks on insurance knowledge pages; insurance service communication and interaction data, such as transcripts of telephone customer service calls, online customer service instant chat records, and email correspondence; historical profile descriptions, such as previously generated customer demand and preference analysis reports; customer data noted by service personnel, such as offline service interview records and handwritten notes on special customer needs; insurance product recommendation plans, such as customized product combination suggestion documents for this customer; and customer source-related data, such as notes on customer referrer information, records of participation in channel promotion activities, and customer registration sources. This embodiment does not limit the specific types, quantities, and combinations of multi-source heterogeneous data and can be flexibly adjusted according to the data source coverage and analysis needs of the actual insurance customer profile generation scenario.

[0025] Step S102: Call the preset large model, which extracts structured target customer information from unstructured multi-source heterogeneous data.

[0026] In this step, the target customer information includes one or more of the following: policy information, customer trajectory, communication records, customer profile, customer information, insurance plan, and business opportunity notes. This embodiment does not limit the specific type, quantity, and combination of target customer information, and can flexibly extract it according to the actual profile analysis dimensions and task requirements.

[0027] Step S103: Invoke the pre-built planning agent, which decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and assigns the profile analysis task to the corresponding specialized agent for the dimension. The specialized agent then processes the task to obtain the corresponding specialized analysis result.

[0028] In this step, specialized intelligent agents can be set up around the core dimensions of insurance customer profile analysis, including but not limited to one or more of the following: a family structure analysis intelligent agent, used to analyze information about the customer's family members, including member roles, ages, occupations, income, etc.; a purchased insurance analysis intelligent agent, used to identify the insurance products the customer has purchased and their detailed information; an insurance type intention analysis intelligent agent, used to analyze the types of insurance the customer may be interested in and their budget; a health status analysis intelligent agent, used to extract health information of the customer and their family members; and a sales reminder analysis intelligent agent, used to generate follow-up suggestions based on the customer's insurance purchase preferences and concerns. This embodiment does not limit the specific type, number, or functional division of specialized intelligent agents and can be flexibly configured according to the analysis scenario of the insurance customer profile and actual business needs.

[0029] Step S104: According to the preset output format requirements, the extracted customer information is output as structured target customer information; the output format requirements include field format, dimension classification and presentation specifications.

[0030] This application embodiment provides a possible implementation method. In step S102 above, the large model extracts structured target customer information from unstructured multi-source heterogeneous data, such as... Figure 2 As shown, it may specifically include the following steps A1 to A3: Step A1: The large model performs logical reasoning on multi-source heterogeneous data based on the preset thought chain prompts. Step A2 involves combining a preset number of sample examples to enable the large model to accurately extract customer information by following the rules in the sample examples; the sample examples include the correspondence between unstructured text and structured information and the extraction rules. Step A3: Based on the preset output format requirements, output the extracted customer information as structured target customer information; the output format requirements include field format, dimension classification and presentation specifications.

[0031] This embodiment does not impose detailed specifications on the specific content of the thought chain prompt data, the number of sample examples, or the output format requirements. The content of the thought chain prompt data can be flexibly designed according to the specific type of unstructured, multi-source, heterogeneous data. For example, for dialogue-type data, the focus can be on logical reasoning prompts, while for document-type data, the focus can be on key information location prompts. The number of sample examples can be adjusted based on the adaptability of the large model to the insurance customer information extraction task. If the data features are complex, the number of examples can be appropriately increased to improve extraction accuracy; if the data type is simple, the number of examples can be reduced to simplify the processing flow. The detailed specifications of the output format requirements can be determined based on the subsequent use scenarios of the target customer information. For example, to adapt to multi-dimensional decomposition, the dimensional classification specifications can be refined; to facilitate information extraction, field format standards can be clearly defined. Overall, the above elements can be flexibly adjusted according to the characteristics of the unstructured data type and the extraction needs of the target customer information to ensure the extraction efficiency and applicability of structured target customer information.

[0032] This application embodiment provides a possible implementation method. In step S103 above, the specialized intelligent agent performs task processing to obtain the corresponding specialized analysis results, such as... Figure 3 As shown, it may specifically include the following steps B1 to B3: Step B1: Receive the profile analysis task assigned by the planning agent for the corresponding dimension, and obtain the data related to the profile analysis task from the structured target customer information. Step B2: Based on the analysis requirements of the profile analysis task, extract key information from the acquired relevant data and filter out the content that is valuable for completing the profile analysis task. Step B3 involves analyzing the extracted key information to generate specific analysis results that conform to the analysis logic of the corresponding dimensions, and then returning these specific analysis results to the planning agent.

[0033] In this embodiment, the scope of key information extraction can be determined based on the specific dimensions of the profiling analysis task. For example, the family structure analysis dimension can focus on extracting information such as family member composition and income, while the purchased insurance analysis dimension can focus on extracting information such as policy type and coverage period. The presentation format of the special analysis results can be adjusted according to the subsequent integration needs of the planning intelligent agent. For example, the data can be clearly displayed in a structured table format, or the analysis conclusions can be presented in a text description format. This embodiment does not limit the scope of key information extraction or the presentation format of special analysis results.

[0034] This application embodiment provides a possible implementation method. In step S103 above, the planning agent decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and delegates the profile analysis task to the corresponding dimension's specialized agent. The specialized agent then processes the task to obtain the corresponding specialized analysis result, such as... Figure 4 As shown, it may also include the following steps C1 to C2: Step C1: Before the planning agent decomposes the target customer information into multiple dimensions, it determines whether the length of the target customer information exceeds the context window of the large model. If it does, the target customer information is processed as long text and then segmented into long text blocks. The target customer information is obtained through the fact correction step. In step C2, before the specialized intelligent agent performs task processing, it is determined whether the length of the profile analysis task exceeds the large model context window. If it does, the profile analysis task is treated as a long text and processed into long text blocks. The corresponding analysis results are obtained through the fact correction step.

[0035] Another possible implementation provided in this application embodiment, the long text block processing in steps C1 to C2 above may specifically include the following steps: Obtain the long text to be processed, and split the long text into multiple text blocks according to the context window length of the large model, while maintaining a preset content overlap between adjacent text blocks; Each text block is processed in parallel to obtain the local structured information corresponding to each text block. The processing results of each text block are summarized and integrated to obtain the first target information.

[0036] In this embodiment, the length judgment threshold can be flexibly set according to the actual context window parameters of the large model used, ensuring that the segmented text can adapt to the model's processing capabilities; the preset overlap ratio can be adjusted according to the continuity requirements of the text information to avoid the breakage of key information due to segmentation; the parallel processing method for text blocks can be selected according to the hardware computing power and processing efficiency requirements, such as using multi-threaded processing or distributed processing. Overall, it can be determined according to the performance of the large model, the characteristics of the text data, and the task processing requirements.

[0037] Another possible implementation provided in this application embodiment, the fact correction steps in steps C1 to C2 above, specifically include the following steps: For the first target information obtained by long text segmentation, retrieve the vector knowledge base that stores relevant domain knowledge, object information and historical data to obtain reference information related to the first target information; Hallucination detection is performed on the first target information, and the first target information is compared with the retrieved reference information to identify and mark the content that exists in the first target information but is not supported by the reference information; Error correction is performed on the first target information. For statements in the first target information that contradict the reference information, replacement or supplementation is made based on the accurate content in the reference information to obtain the corrected final result.

[0038] In this embodiment, the information storage type of the vector knowledge base can be flexibly expanded according to the actual content scenario of the first target information. For example, for target customer information scenarios in the insurance field, it can store information such as insurance product terms and conditions, historical customer service examples, and industry compliance standards. The comparison dimensions of illusion detection can be adapted and adjusted according to the information authenticity verification requirements. For example, comparative analysis can be carried out from the dimensions of information source reliability, data time validity, and logical self-consistency to achieve accurate identification of unfounded content. The specific implementation method of error correction can be determined according to the specific type of contradictory expression. If the data value is incorrect, it can be directly replaced with the accurate data in the reference information. If the information content is missing, the complete content in the reference information can be supplemented.

[0039] The above introduces Figure 1 The embodiments shown have various implementation methods for each step. The following will further explain the insurance customer profile generation method based on large model and multi-agent collaboration of this application through specific embodiments.

[0040] This specific embodiment mainly addresses the issues of accuracy in information extraction, targeted dimensional decomposition, and structured result integration during the generation of insurance customer profiles. Relying on the collaborative cooperation of planning intelligent agents and specialized intelligent agents, it efficiently extracts key information from multi-source customer data and generates usable profile reports, avoiding information omissions or dimensional biases caused by single-module processing, and providing accurate data support for subsequent insurance services.

[0041] (1) Data acquisition and structured extraction stage 1. Acquisition of multi-source heterogeneous data: Unstructured multi-source heterogeneous data of the target customer (customer Xiaoming) is acquired through the information acquisition module. Specifically, this includes data related to purchased insurance ("customer Xiaoming's policy includes an accident insurance policy"), insurance service communication and interaction data ("the communication record mentions that he wants to purchase medical insurance for his parents"), and historical profile description ("historical customer profile shows that his family has parents and one child"). 2. Large-scale model structuring: The large-scale model is called to process the unstructured data mentioned above. First, the large-scale model performs logical reasoning based on the pre-set thought chain prompts (such as "sorting out customer policy, family, and needs-related information from the text"). Then, combined with two sets of sample examples (Example 1: unstructured text "Customer A has a critical illness insurance policy" corresponds to structured information "Purchased Insurance: Critical Illness Insurance"; Example 2: text "Customer B's family has a spouse and 1 child" corresponds to "Family Members: Spouse, 1 Child"), the model completes accurate information extraction by following the extraction rules in the examples. Finally, according to the pre-set output format requirements of "field format (key-value pairs), dimension classification (purchased insurance, family structure, insurance type intention), and presentation standard (concise text)," the structured target customer information is obtained. 3. Long text judgment and processing: After the planning agent receives the structured target customer information, it first judges whether its length exceeds the context window of the large model. If it does, it splits the text into multiple text blocks according to the length of the context window, maintains 20% content overlap, processes them in parallel and summarizes them to obtain the first target information, and directly enters the task decomposition stage.

[0042] (2) Multi-agent collaborative analysis stage This mainly includes the following specialized intelligent agents: Family structure analysis agent: used to extract and analyze information about customer family members, including member roles, ages, occupations, incomes, etc. Purchased Insurance Analysis Agent: Used to identify the insurance products that customers have purchased and their detailed information; Insurance Intention Analysis Agent: Used to analyze the types of insurance and budgets that customers may be interested in; Health status analysis agent: used to extract health information of customers and their family members; Sales reminder analysis agent: used to generate follow-up suggestions based on customer insurance purchase preferences and concerns.

[0043] 4. Planning the intelligent agent task decomposition: The planning intelligent agent decomposes the structured target customer information into multiple dimensions and determines 3 core analysis dimensions: family structure, purchased insurance, and insurance type intention. Correspondingly, 3 profile analysis tasks are generated and assigned to the corresponding specialized intelligent agents (family structure analysis intelligent agent, purchased insurance analysis intelligent agent, and insurance type intention analysis intelligent agent). 5. Specialized intelligent agent task processing: The family structure analysis intelligent agent receives the task of "extracting family members," obtains relevant data ("historical customer profile shows that their family has parents and 1 child"), extracts key information in a targeted manner, and then analyzes it to obtain the specialized analysis result of "family members: parents, 1 child"; The purchased insurance analysis intelligent agent receives the task of "identifying existing insurance policies," filters information from relevant data ("customer Xiaoming's insurance policy includes an accident insurance policy"), and obtains the result of "purchased insurance: accident insurance"; The insurance type intention analysis intelligent agent receives the task of "sorting out insurance needs," and based on communication record data, obtains the result of "insurance type intention: wants to purchase medical insurance for parents"; 6. Task length judgment and correction: Before processing, each special agent judges the length of the profile analysis task; if it exceeds the window, the task is divided into long text blocks, and the analysis results are obtained through fact correction. The processing is completed directly and the results are returned to the planning agent.

[0044] (3) Fact correction and profile report generation stage 7. Fact Correction: If the first target information obtained by segmenting long text during the processing of the special intelligent agent (such as information after segmenting complex insurance policy terms) is involved, then the fact correction step is performed: retrieve the vector knowledge base (which stores reference information such as "Accident Insurance Coverage" in insurance product terms and customer historical service records), compare the first target information with the reference information, identify unsubstantiated content, and obtain the corrected result; 8. Planning Agent Summary Output: After receiving the analysis results from all specialized agents, the planning agent summarizes the results, eliminates information redundancy, and generates a structured target customer profile report. The report includes: family structure (parents, 1 child), purchased insurance (accident insurance), and insurance preferences (plan to purchase medical insurance for parents). Finally, the report is output through the output module for subsequent insurance services (such as recommending medical insurance products for parents).

[0045] It should be noted that the sequence numbers of the steps in the above embodiments do not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. In practical applications, all the above possible implementation methods can be arbitrarily combined in a combined manner to form possible embodiments of this application, which will not be described in detail here.

[0046] Based on the insurance customer profile generation method based on large model and multi-agent collaboration provided in the above embodiments, and based on the same inventive concept, this application also provides an insurance customer profile generation system based on large model and multi-agent collaboration.

[0047] Figure 5 This is a structural diagram of the insurance customer profile generation system based on large models and multi-agent collaboration provided in the embodiments of this application. Figure 5 As shown, the device for identifying abnormal behavior of consultants may specifically include an information acquisition module 210, a multi-agent collaborative analysis module 220, and an output module 230.

[0048] The information acquisition module 210 is used to acquire unstructured multi-source heterogeneous data related to the target customer, call the preset large model, and extract structured target customer information from the unstructured multi-source heterogeneous data. The multi-agent collaborative analysis module 220 is used to call the pre-built planning agent, which decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and assigns the profile analysis task to the special agent of the corresponding dimension, which then processes the task to obtain the corresponding analysis result. Output module 230 is used to generate a target customer profile report by the planning agent based on the special analysis results returned by each special agent; This application embodiment provides a possible implementation, wherein the information acquisition module 210 is further configured to: The large model performs logical reasoning on multi-source heterogeneous data based on pre-set thought chain prompts; By combining a preset number of sample examples, the large model can follow the rules in the sample examples to accurately extract customer information; the sample examples include the correspondence between unstructured text and structured information and the extraction rules. Based on the preset output format requirements, the extracted customer information is output as structured target customer information; the output format requirements include field format, dimension classification and presentation specifications.

[0049] This application embodiment provides a possible implementation, wherein the multi-agent cooperative analysis module 220 is further configured to: Receive profile analysis tasks assigned by the planning intelligent agent for the corresponding dimension, and obtain data related to the profile analysis tasks from the structured target customer information; To meet the analytical needs of the profiling task, key information is extracted from the relevant data and content that is valuable for completing the profiling task is selected. The extracted key information is analyzed to form specific analysis results that conform to the analysis logic of the corresponding dimension, and the specific analysis results are returned to the planning agent.

[0050] This application embodiment provides a possible implementation, wherein the multi-agent cooperative analysis module 220 is further configured to: Before the planning agent decomposes the target customer information into multiple dimensions, it determines whether the length of the target customer information exceeds the context window of the large model. If it does, the target customer information is processed as long text and then segmented into long text blocks. The target customer information is obtained through a fact correction step. Before the specialized intelligent agent performs the task processing, it is determined whether the length of the profile analysis task exceeds the context window of the large model. If it does, the profile analysis task is treated as a long text and processed into long text blocks. The corresponding analysis results are obtained through the fact correction step.

[0051] This application provides one possible implementation method, such as... Figure 6 As shown above, Figure 5 The insurance customer profile generation system demonstrated, based on large models and multi-agent collaboration, may also include a long text segmentation processing module 240 and a retrieval enhancement generation module 250.

[0052] The long text segmentation processing module 240 is used to acquire the long text to be processed, split the long text into multiple text blocks according to the context window length of the large model, and maintain a preset content overlap between adjacent text blocks; process each text block in parallel to obtain the local structured information corresponding to each text block, and summarize and integrate the processing results of each text block to obtain the first target information.

[0053] The retrieval enhancement generation module 250 is used to retrieve relevant domain knowledge, object information, and historical data from a vector knowledge base for the first target information obtained by the long text segmentation processing module, and obtain reference information related to the first target information; perform illusion detection on the first target information, compare the first target information with the retrieved reference information, identify and mark content in the first target information that exists but is not supported by the reference information; and perform error correction on the first target information, replacing or supplementing statements in the first target information that contradict the reference information based on the accurate content in the reference information, to obtain the corrected final result.

[0054] Based on the same inventive concept, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the insurance customer profile generation method based on large model and multi-agent collaboration of any of the above embodiments.

[0055] In an exemplary embodiment, an electronic device is provided, such as Figure 7 The illustrated electronic device 300 includes a processor 301 and a memory 303. The processor 301 and the memory 303 are connected, for example, via a bus 302. Optionally, the electronic device 300 may also include a transceiver 304. It should be noted that in practical applications, the transceiver 304 is not limited to one type, and the structure of this electronic device 300 does not constitute a limitation on the embodiments of this application.

[0056] Processor 301 may be a CPU (Central Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 301 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0057] Bus 302 may include a pathway for transmitting information between the aforementioned components. Bus 302 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 302 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0058] The memory 303 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0059] The memory 303 stores computer program code that executes the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the computer program code stored in the memory 303 to implement the content shown in the foregoing method embodiments.

[0060] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0061] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of this application, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to leave the protection scope of this application.

Claims

1. A method for generating insurance customer profiles based on large models and multi-agent collaboration, characterized in that, The method includes: Acquire unstructured, multi-source, heterogeneous data related to target customers; The system calls a pre-defined large model, which extracts structured target customer information from unstructured multi-source heterogeneous data. The pre-built planning agent is invoked, which decomposes the target customer information into multiple dimensions and generates corresponding profile analysis tasks for each analysis dimension. The profile analysis tasks are then assigned to the corresponding specialized agent, which processes the tasks to obtain the corresponding specialized analysis results. The planning agent generates a target customer profile report based on the specific analysis results returned by each specific agent.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data includes one or more of the following: data related to purchased insurance, internet behavior trajectory records, insurance service communication and interaction data, historical profile descriptions, customer data with service personnel notes, insurance product recommendation plans, and customer source association data.

3. The method according to claim 1, characterized in that, The large model extracts structured target customer information from unstructured, multi-source, heterogeneous data, including: The large model performs logical reasoning on multi-source heterogeneous data based on pre-set thought chain prompts; By combining a preset number of sample examples, the large model can follow the rules in the sample examples to accurately extract customer information; the sample examples include the correspondence between unstructured text and structured information and the extraction rules. Based on the preset output format requirements, the extracted customer information is output as structured target customer information; the output format requirements include field format, dimension classification and presentation specifications.

4. The method according to any one of claims 1 to 3, characterized in that, Specialized intelligent agents include one or more of the following: family structure analysis intelligent agents, purchased insurance analysis intelligent agents, insurance type intention analysis intelligent agents, health status analysis intelligent agents, and sales reminder analysis intelligent agents.

5. The method according to any one of claims 1 to 3, characterized in that, The specific intelligent agent performs task processing to obtain corresponding specific analysis results, including: Receive profile analysis tasks assigned by the planning intelligent agent for the corresponding dimension, and obtain data related to the profile analysis tasks from the structured target customer information; To meet the analytical needs of the profiling task, key information is extracted from the relevant data and content that is valuable for completing the profiling task is selected. The extracted key information is analyzed to form specific analysis results that conform to the analysis logic of the corresponding dimension, and the specific analysis results are returned to the planning agent.

6. The method according to claim 1, characterized in that, The planning agent breaks down target customer information into multiple dimensions, generates corresponding profile analysis tasks for each analysis dimension, and delegates these tasks to specialized agents for that dimension. These specialized agents then process the tasks to obtain the corresponding specialized analysis results, including: Before the planning agent decomposes the target customer information into multiple dimensions, it determines whether the length of the target customer information exceeds the context window of the large model. If it does, the target customer information is processed as long text and then segmented into long text blocks. The target customer information is obtained through a fact correction step. Before the specialized intelligent agent performs the task processing, it is determined whether the length of the profile analysis task exceeds the context window of the large model. If it does, the profile analysis task is treated as a long text and processed into long text blocks. The corresponding analysis results are obtained through the fact correction step.

7. The method according to claim 6, characterized in that, Long text segmentation processing includes: Obtain the long text to be processed, and split the long text into multiple text blocks according to the context window length of the large model, while maintaining a preset content overlap between adjacent text blocks; Each text block is processed in parallel to obtain the local structured information corresponding to each text block. The processing results of each text block are summarized and integrated to obtain the first target information.

8. The method according to claim 7, characterized in that, The fact-correction steps include: For the first target information obtained by long text segmentation, retrieve the vector knowledge base that stores relevant domain knowledge, object information and historical data to obtain reference information related to the first target information; Hallucination detection is performed on the first target information, and the first target information is compared with the retrieved reference information to identify and mark the content that exists in the first target information but is not supported by the reference information; Error correction is performed on the first target information. For statements in the first target information that contradict the reference information, replacement or supplementation is made based on the accurate content in the reference information to obtain the corrected final result.

9. An insurance customer profile generation system based on large models and multi-agent collaboration, characterized in that, The system includes: The information acquisition module is used to acquire unstructured multi-source heterogeneous data related to the target customer, call the preset large model, and extract structured target customer information from the unstructured multi-source heterogeneous data. The multi-agent collaborative analysis module is used to call a pre-built planning agent, which decomposes the target customer information into multiple dimensions, generates a corresponding profile analysis task for each analysis dimension, and assigns the profile analysis task to the corresponding special agent for the dimension. The special agent then processes the task to obtain the corresponding analysis results. The output module is used to generate target customer profile reports by the planning agent based on the special analysis results returned by each special agent.

10. An electronic device, characterized in that, The invention includes a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the insurance customer profile generation method based on large model and multi-agent collaboration as described in any one of claims 1 to 8.

Images