Insurance service full-process processing method and device based on multi-agent cooperation

Abstract

The invention discloses an insurance service full-process processing method and device based on multi-agent cooperation, and belongs to the technical field of artificial intelligence. A user request is received through the user interaction layer, a natural language processing technology is utilized to identify a user intention and extract key entity information, and structured user input data is generated; dynamically routing the request to a corresponding AI agent through a task distribution module according to the intention type, wherein the AI agent comprises a customer consultation agent, a risk assessment agent and a claim settlement processing agent; each agent cooperatively processes a business process based on shared session context information, exchanges structured data through a unified communication protocol and executes a cross-link task; and finally, integrating processing results of all agents, generating a service response conforming to insurance business logic, and feeding back the service response to a user through a user interaction layer. According to the invention, efficient automation and intelligent cooperation of the whole insurance service process are realized, and the accuracy and response speed of service processing are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for processing the entire insurance service process based on multi-agent collaboration. Background Technology

[0002] As a crucial infrastructure of the insurance industry, insurance service systems are widely used in core business processes such as customer consultation, risk assessment, and claims processing. With the development of artificial intelligence (AI) technology, a preliminary intelligent insurance service framework has been constructed through the collaborative operation of Natural Language Processing (NLP), image recognition, and machine learning models. Specifically, this technology system covers the entire process from user intent recognition to multimodal data processing, including key steps such as intent recognition, entity extraction, risk scoring calculation, and image feature extraction. The combination of Large Language Models (LLMs) and agent technology provides a new technical path for the automated processing of insurance business; however, existing systems remain at the stage of single-point intelligent application, lacking end-to-end collaborative capabilities.

[0003] However, existing intelligent insurance service methods directly adopt modular AI technology without establishing a dynamic collaboration mechanism among multiple agents. This can lead to fragmented business processes, frequent manual intervention, or delayed service response, thus impacting customer experience and business efficiency. Specifically, traditional systems typically use independently deployed chatbots to handle inquiries, utilize static rule models for risk assessment, and complete the claims process through manual review. However, this approach has three major limitations: First, data sharing between business processes is limited; for example, health information collected during inquiries cannot be automatically used for risk assessment. Second, 80% of claims reviews still require manual intervention, with an average processing cycle of 3-5 business days. Third, personalized services rely on basic tags, and user profile updates take more than 7 days, resulting in a 20% loss of potential customers. Therefore, while existing technologies improve efficiency in some areas, they cannot meet the comprehensive needs of the insurance business for end-to-end intelligent, real-time response, and dynamic personalized services. Summary of the Invention

[0004] This invention provides a method and apparatus for processing the entire insurance service process based on multi-agent collaboration. The method aims to receive user requests through a user interaction layer, identify intents and extract key entity information using natural language processing technology, and generate structured input data. Based on the intent type, a task allocation module dynamically routes the data to the corresponding AI agent, including agents for customer consultation, risk assessment, and claims processing. The agents collaborate with other agents to process business processes based on a shared session context, exchanging structured data and executing cross-stage tasks through a unified communication protocol. Finally, all processing results are integrated to generate a final service response that conforms to insurance business logic and is fed back to the user. This achieves full-process automation, precision, and closed-loop management, improving service efficiency and user experience.

[0005] To achieve the above objectives, this invention provides a method for processing the entire insurance service process based on multi-agent collaboration. The method includes: S1, receiving user requests through a user interaction layer and using natural language processing technology to identify user intent and extract key entity information to generate structured user input data; S2, dynamically routing requests to corresponding AI agents based on the intent type of the structured user input data through a task allocation module; wherein the AI ​​agents include a customer consultation agent, a risk assessment agent, and a claims processing agent; S3, the AI ​​agents collaboratively process business processes with other agents based on shared session context information, exchanging structured data through a unified communication protocol and executing cross-stage tasks to obtain processing results; S4, integrating the processing results of all AI agents to generate a final service response conforming to insurance business logic, and feeding it back to the user through the user interaction layer.

[0006] Furthermore, the step of receiving user requests through the user interaction layer and using natural language processing technology to identify user intent and extract key entity information to generate structured user input data includes: S11, using the BERT-CRF model to perform semantic encoding and entity recognition on the user input, and outputting structured data containing entities such as name, age, and insurance type in the format of BIO annotation; S12, based on the dynamic slot filling mechanism, by maintaining the session_id cache in the session state, to achieve proactive follow-up questions on missing information in cross-turn dialogues and maintain contextual coherence.

[0007] Furthermore, the step of dynamically routing requests to the corresponding AI agents through the task allocation module based on the intent type of the structured user input data includes: S21, matching the intent type recognition result with a preset intent-agent mapping table to determine the target agent type; S22, using a rule-driven routing strategy to further verify the accuracy of agent allocation based on entity information in the user request; wherein, entity information includes policy number and claim type.

[0008] Furthermore, the AI ​​agent collaborates with other agents to process business processes based on shared session context information, exchanges structured data through a unified communication protocol, and executes cross-stage tasks to obtain processing results. The steps include: S31, asynchronous communication is achieved through JSON-formatted message queue content; wherein, the message queue content includes user ID, intent type, entity parameters, and processing status; S32, a priority arbitration mechanism is used to handle conflicts between agents. When multiple agents output contradictory results, the results are verified and corrected according to preset agent priority rules.

[0009] Furthermore, the steps of integrating the processing results of all AI agents to generate a final service response that conforms to the insurance business logic and feeding it back to the user through the user interaction layer include: S41, dynamically filling slot values ​​based on a preset response template to generate a natural language response containing the user's name, processing result type, and key data; S42, evaluating the effectiveness of the generated response through an A / B testing mechanism, and optimizing the template content and filling strategy based on user click-through rate and satisfaction indicators.

[0010] Furthermore, the insurance service end-to-end processing method based on multi-agent collaboration of the present invention further includes: S5, generating a structured user profile containing basic information, behavioral preferences and potential needs through feature extraction and labeling based on the user profile construction module; S6, dynamically adjusting the recommendation algorithm parameters according to the user profile, generating a customized insurance product recommendation list using collaborative filtering and content-based recommendation algorithms, and displaying the recommendation results through the user interaction layer.

[0011] In another aspect, this invention provides a multi-agent collaborative insurance service end-to-end processing device, comprising: a user request receiving and intent recognition module, used to receive user requests through a user interaction layer and use natural language processing technology to recognize user intent and extract key entity information to generate structured user input data; a task dynamic routing module, used to dynamically route requests to corresponding AI agents according to the intent type of the structured user input data through a task allocation module; wherein the AI ​​agents include a customer consultation agent, a risk assessment agent, and a claims processing agent; an agent collaborative processing module, used for AI agents to collaboratively process business processes with other agents based on shared session context information, exchange structured data through a unified communication protocol, and execute cross-stage tasks to obtain processing results; and a service response integration and feedback module, used to integrate the processing results of all AI agents, generate a final service response that conforms to the insurance business logic, and feed it back to the user through the user interaction layer.

[0012] Furthermore, the user request receiving and intent recognition module is also used to: use the BERT-CRF model to perform semantic encoding and entity recognition on user input, and output structured data containing entities such as name, age, and insurance type in the format of BIO annotation; based on the dynamic slot filling mechanism, by maintaining the session_id cache in the session state, to realize proactive follow-up questions on missing information in cross-round dialogue and maintain contextual coherence.

[0013] Furthermore, the task dynamic routing module is also used to: match the intent type identification result with a preset intent-agent mapping table to determine the target agent type; and adopt a rule-driven routing strategy to further verify the accuracy of agent allocation based on entity information in the user request; wherein, entity information includes policy number and claim type.

[0014] Furthermore, the intelligent agent collaborative processing module is also used to: realize asynchronous communication through JSON format message queue content; wherein, the message queue content includes user ID, intent type, entity parameters and processing status; and use a priority arbitration mechanism to handle conflicts between intelligent agents. When multiple intelligent agents output contradictory results, the results are verified and corrected according to the preset intelligent agent priority rules.

[0015] Furthermore, the service response integration and feedback module is also used to: dynamically fill slot values ​​based on preset response templates to generate natural language responses containing user names, processing result types, and key data; evaluate the effectiveness of the generated responses through an A / B testing mechanism, and optimize template content and filling strategies based on user click-through rates and satisfaction indicators.

[0016] Furthermore, the insurance service end-to-end processing device based on multi-agent collaboration of the present invention further includes: a user profile construction module, used to generate a structured user profile containing basic information, behavioral preferences and potential needs based on the user profile construction module through feature extraction and labeling; and a recommendation algorithm parameter adjustment module, used to dynamically adjust the recommendation algorithm parameters according to the user profile, generate a customized insurance product recommendation list using collaborative filtering and content-based recommendation algorithms, and display the recommendation results through a user interaction layer.

[0017] Compared with the prior art, the technical solution provided by the present invention has at least the following beneficial effects: 1. The system receives user requests through the user interaction layer and uses natural language processing technology to identify user intent and extract key entity information to generate structured input data. It then uses the BERT-CRF model for semantic encoding and entity recognition, and combines a dynamic slot filling mechanism to maintain the session state, thereby enabling proactive follow-up questions and contextual coherence across rounds of dialogue.

[0018] 2. Based on the intent type, the task allocation module dynamically routes the task to the corresponding AI agent, including customer consultation, risk assessment and claims processing agents, and verifies the accuracy of the allocation based on entity information through a rule-driven routing strategy.

[0019] 3. The agents collaboratively process business processes based on a shared session context, exchange structured data through a unified communication protocol, implement asynchronous communication using a JSON-formatted message queue, and employ a priority arbitration mechanism to handle conflicts between agents, ensuring consistency and reliability of results.

[0020] 4. Finally, integrate all processing results, dynamically fill slot values ​​based on preset response templates to generate natural language responses that conform to insurance business logic, and optimize response content through A / B testing.

[0021] 5. Introducing a user profile building module generates structured user profiles and dynamically adjusts recommendation algorithm parameters to provide customized insurance product recommendations, thereby achieving full-process automation, precision, and closed-loop management, significantly improving service efficiency, processing accuracy, and user experience.

[0022] This invention discloses a method and apparatus for processing the entire insurance service process based on multi-agent collaboration. It achieves accurate understanding of user intent and maintains contextual coherence through natural language processing and dynamic slot filling, and dynamically allocates tasks to corresponding agents based on rule-driven routing. A unified communication protocol and priority arbitration mechanism ensure data consistency and business reliability in multi-agent collaborative processing. Finally, through templated response generation and user profile-driven personalized recommendations, a complete automated closed loop from request access to service feedback is constructed, significantly improving the efficiency, accuracy, and user experience of insurance business processing, and enhancing the overall service capabilities and operational efficiency of the system in complex insurance service scenarios. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a full-process insurance service method based on multi-agent collaboration, provided as an embodiment of the present invention; Figure 2 A system architecture diagram of a full-process insurance service processing method based on multi-agent collaboration provided in an embodiment of the present invention; Figure 3 A business process diagram of a full-process insurance service processing method based on multi-agent collaboration provided in an embodiment of the present invention; Figure 4 A flowchart of dialogue state management for a multi-agent collaborative insurance service process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an insurance service end-to-end processing device based on multi-agent collaboration, provided in an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] The core idea of ​​this invention is to construct a full-process insurance service system based on multi-agent collaborative decision-making. This system leverages natural language processing technology and a dynamic slot-filling mechanism to accurately interpret user intent and maintain contextual coherence across dialogue rounds. A rule-driven task routing mechanism dynamically allocates user requests to specialized agents such as customer consultation, risk assessment, and claims processing. A unified communication protocol and priority arbitration mechanism ensure data consistency and business logic reliability during collaborative processing. Furthermore, structured user profiles and adaptive recommendation algorithms enable personalized recommendations of insurance products and services. Finally, templated response generation and A / B testing optimization mechanisms form a complete automated closed loop from user request access to service feedback. This transforms the traditional single service process into an intelligent service system that deeply perceives user needs, makes dynamic decisions, and continuously optimizes, significantly improving the efficiency, accuracy, and user experience of insurance business processing, and enhancing the system's overall service capabilities and operational efficiency in complex business scenarios.

[0027] The following description, with reference to the accompanying drawings, describes a method and apparatus for processing the entire insurance service process based on multi-agent collaboration, according to an embodiment of the present invention.

[0028] Example 1 This embodiment provides a method for processing the entire insurance service process based on multi-agent collaboration. For example... Figure 1 As shown, the method includes the following steps: S1 receives user requests through the user interaction layer and uses natural language processing technology to identify user intent and extract key entity information to generate structured user input data.

[0029] Specifically, the technical implementation of this step is based on the collaborative work of deep learning models and context management mechanisms to ensure that user input can be accurately parsed and transformed into actionable business data.

[0030] Specifically, the user interaction layer supports multiple input methods, including text, voice, and graphical interface input, all of which are processed through a standardized interface. The customer consultation agent first uses an intent recognition model (such as BERT, LSTM, etc.) to perform semantic analysis on the user input, identifying the request type, such as "checking policy status," "applying for a claim," or "learning about insurance products." The intent recognition model is typically trained based on a classification task, outputting the intent category of the user's request. Its classification accuracy can reach over 92% on the test set, meeting the high-precision requirements of business scenarios.

[0031] Furthermore, the system incorporates Named Entity Recognition (NER) technology to extract key entity information from user input, such as name, policy number, ID number, and accident time. The NER module employs BERT-CRF or BiLSTM-CRF architectures, using BIO annotation format for entity recognition in the text, outputting structured data in a format such as {Name: “Li Ming”, Policy Number: “PICC2024123456”, Insurance Type: “Auto Insurance”}. This module achieves an accuracy (F1-score) of over 90% on the standard test set and supports contextual memory in multi-turn dialogues, maintaining user state through a session_id mechanism to ensure information continuity and integrity.

[0032] Specifically, the input text length for the intent recognition model is limited to 512 tokens to accommodate the expressive range of most user requests. The entity types supported by the NER module include, but are not limited to, PERSON, POLICY_ID, DATE, INSURANCE_TYPE, etc. The confidence threshold for entity recognition is set to 0.85. Recognition results below this threshold will be marked as "pending confirmation" and trigger the system's proactive follow-up mechanism.

[0033] Specifically, this step is widely used in business processes such as customer inquiries, claims applications, and risk assessment requests. For example, when a user enters "I want to apply for car insurance claims, the policy number is PICC2024123456" through the app, the system will automatically recognize the intent to "apply for claims" and extract key entities such as "car insurance" and "policy number," providing structured input for the subsequent claims processing agent to call relevant services.

[0034] Specifically, by generating structured data, the system's ability to understand and respond to user requests is significantly improved, human intervention is reduced, and a data foundation is laid for subsequent collaborative processing by intelligent agents, thereby realizing the intelligentization and automation of the insurance service process.

[0035] Furthermore, S1 includes: S11 uses the BERT-CRF model to perform semantic encoding and entity recognition on user input, and outputs structured data containing entities such as name, age, and insurance type in the format of BIO annotation.

[0036] Specifically, this step combines deep learning and sequence labeling technology to accurately identify key entity information such as name, age, and insurance type from the user's natural language input, and outputs structured data in BIO labeling format to provide basic support for subsequent business logic processing.

[0037] Specifically, the BERT-CRF model first utilizes BERT (Bidirectional Encoder Representations from Transformers) to perform context-aware semantic encoding on the user input text. BERT extracts the context embedding vector for each word through a multi-layer Transformer structure, with an output dimension of [missing information]. The word vector representation, where Typically, the number of words is 768 or 1024, depending on the BERT version used (such as BERT-base or BERT-large). Subsequently, the CRF (Conditional Random Field) layer performs sequence labeling based on these word vectors, assigning an entity label to each word, such as "B-PER" for the beginning of a name, "I-PER" for a continuation of a name, and "O" for a non-entity word. During the training phase, the BERT-CRF model uses labeled insurance business corpora covering various entity types, such as names, ages, policy numbers, and insurance types, ensuring good generalization ability.

[0038] Furthermore, the input sequence length of the BERT model is typically limited to 512 tokens to accommodate the maximum context length of the model. The transition matrix size of the CRF layer is... ,in The number of entity tags (e.g., in BIO format, if 5 entity types are defined, then...). The model training employs a cross-entropy loss function combined with L2 regularization to optimize the objective function.

[0039] in, For cross-entropy loss, For regularization terms, This is the regularization coefficient, usually set to... to During training, the Adam optimizer was used, and the learning rate was set to [value missing]. The batch size is 32, and the number of training rounds is 3 to 5 to achieve a better convergence effect.

[0040] Specifically, this step is widely applied in business processes such as customer consultation, claims applications, and risk assessments. For example, when a user submits a claims application, the system automatically identifies key information such as name, policy number, and time of incident using the BERT-CRF model, eliminating the need for manual intervention and significantly improving processing efficiency. Furthermore, this model can also be used to extract policyholder information from user voice or text, providing data support for personalized recommendations and risk assessments.

[0041] Specifically, the technical effectiveness of this step is reflected in its high-precision entity recognition capability and excellent contextual understanding. Through the pre-trained language model of BERT, the system can accurately capture semantic information in user input, and combined with the sequence modeling capability of CRF, achieve precise recognition of entity boundaries. In actual testing, the model achieves an F1-score of [insert score here]. The results shown above are significantly superior to traditional rule matching or single-model recognition methods. Therefore, the introduction of the BERT-CRF model not only improves the accuracy of information extraction but also provides high-quality structured input for multi-agent collaboration in the AI ​​Agent layer, serving as a crucial technical support for achieving intelligent processing throughout the entire process.

[0042] S12, based on the dynamic slot filling mechanism, maintains the session_id cache in the session state to proactively ask for missing information in cross-round dialogues and maintain contextual coherence.

[0043] Specifically, this mechanism ensures that users do not need to repeatedly provide information in multiple rounds of interaction through the collaborative work of structured dialogue state management, intent recognition, and entity extraction modules. The system can automatically identify and complete key business parameters, thereby improving interaction efficiency and user experience.

[0044] Furthermore, the system adopts a "state-driven + slot-filling" dialogue management architecture. Each user session is identified by a unique session_id, which is generated at the user interaction layer and passed to the AI ​​Agent layer to associate all contextual information of the current dialogue. In each round of dialogue, the system maintains a structured DialogueState Table (DST) that records the current intent category (such as "apply for claims" or "inquire about products") and the information of the identified entities (such as name, policy number, and accident time). If a key slot is not yet filled or the information is incomplete, the system will proactively initiate follow-up questions based on preset business rules and slot priorities, such as "What is your policy number?" or "Please provide the detailed time of the accident."

[0045] Specifically, the slot filling mechanism involves several key parameters. For example, the slot priority threshold can be set to... This is used to determine which information to prioritize in a multi-turn dialogue; the confidence threshold for intent recognition. Ensure the system triggers the slot filling process only under high confidence levels; entity extraction confidence level. This is used to filter out low-confidence recognition results and avoid incorrect data entry. Furthermore, the system supports cross-round information referencing; for example, when a user mentions "my policy," the system can automatically associate it with the policy_id recorded in previous conversations, ensuring contextual consistency.

[0046] Specifically, this mechanism is widely applied in multi-round interaction scenarios such as customer inquiries, claims applications, and risk assessments. For example, in the claims application process, a user may only provide their name and accident description in the first round, but the system will automatically ask for key information such as the policy number and the status of uploaded medical certificates in subsequent rounds to ensure the completeness and compliance of the claims process. In customer inquiries, the system can remember information such as the user's occupation and age, avoiding repeated inquiries and improving interaction efficiency.

[0047] Specifically, this mechanism effectively solves problems such as information fragmentation, context loss, and repetitive user input in traditional insurance service systems. Through session_id caching and structured dialogue state management, the system achieves automatic association and completion of information across rounds, significantly improving the coherence and processing efficiency of multi-round dialogues, and providing a solid foundation for intelligent agent collaboration and business process automation in the AI ​​Agent layer.

[0048] S2, based on the intent type of the structured user input data, dynamically routes the request to the corresponding AI agent through the task allocation module; among which, the AI ​​agents include customer consultation agent, risk assessment agent, and claims processing agent.

[0049] Specifically, this step, based on the intent recognition module in Natural Language Processing (NLP) and combined with pre-defined intent-agent mapping rules, achieves accurate classification and routing of user requests. Intent recognition typically employs deep learning-based models, such as BERT and LSTM, to semantically model the user-input text and output an intent label, such as "customer inquiry," "risk assessment," or "claims processing." The output of the intent recognition model must meet a certain confidence threshold (e.g., ...). This ensures the accuracy of routing decisions.

[0050] Furthermore, after receiving the intent recognition result, the task allocation module forwards the request to the corresponding AI agent according to preset routing rules. For example, when the user intent is identified as "claims processing," the system will route the request to the claims processing agent; if it is "risk assessment," it will route it to the risk assessment agent. This routing mechanism is implemented using a rule engine, supporting flexible configuration and expansion, ensuring that the system can adapt to the evolving needs of different business scenarios.

[0051] Specifically, the task allocation module is typically deployed in the coordinator of the AI ​​Agent layer. It is responsible for receiving messages from the user interaction layer and asynchronously forwarding them to the target agent via message queues (such as Kafka or RabbitMQ). The message format uses structured JSON, which includes fields such as user ID, intent category, and key entity information (such as policy number and claim type) to ensure that the agent can quickly parse and initiate the corresponding processing flow.

[0052] Specifically, by combining intent recognition with task allocation, the system achieves end-to-end automation from user input to business processing, reducing manual intervention and lowering service latency. Simultaneously, this mechanism supports multi-agent collaborative processing of complex business processes, providing clear input boundaries and processing paths for subsequent modules such as personalized services, risk assessment, and claims review. It is the core foundation for achieving intelligent, personalized, and efficient collaborative insurance services.

[0053] Furthermore, S2 includes: S21, the target agent type is determined by matching the intent type recognition result with the preset intent-agent mapping table.

[0054] Specifically, this step uses semantic tags output by the intent recognition module in Natural Language Processing (NLP) and predefined business logic mapping relationships to achieve intelligent routing and task allocation of user requests.

[0055] Specifically, intent recognition typically employs deep learning-based models, such as BERT, LSTM, or CRF, to semantically classify user-input text and output one or more intent labels, such as "inquire about insurance products," "apply for claims," ​​or "check claims progress." The recognition results are output in a structured format, including the intent category and its confidence score. The system matches the intent labels against a pre-defined intent-agent mapping table, a static configuration table that defines the agent type and processing priority for each intent. For example, when the intent "apply for claims" is recognized, the system routes the request to the "claims processing agent."

[0056] Furthermore, the output of the intent recognition model needs to meet a certain confidence threshold (e.g., To ensure the accuracy of intent judgment, the mapping table can be configured to correspond to multiple intents with intelligent agents, and supports fuzzy matching and parallel processing of multiple intents. Furthermore, the system supports dynamic updates to the mapping table to adapt to new business scenarios or changes in user expression methods.

[0057] Specifically, this step is widely used in the task distribution process between the user interaction layer and the AI ​​Agent layer. For example, when a user enters "I want to apply for car insurance compensation" through the app, the system recognizes the intent to "apply for compensation," immediately matches the request to the "compensation processing AI agent," and triggers the subsequent document review and compensation process. This mechanism ensures that different business requests can be accurately assigned to AI agents with the corresponding processing capabilities, thereby improving system response efficiency and the accuracy of business processing.

[0058] Specifically, this step effectively achieves automated classification of user requests and accurate matching with intelligent agents, avoiding the inefficiencies of traditional systems that rely on manual judgment or fixed processes. By combining intent recognition with mapping tables, the system maintains logical consistency in multi-turn dialogues and cross-module collaborations, providing a stable and efficient entry mechanism for subsequent modules such as personalized services, risk assessment, and claims processing, significantly improving the overall system's intelligence level and user experience.

[0059] S22 employs a rule-driven routing strategy to further verify the accuracy of agent allocation based on entity information in the user request; the entity information includes the policy number and claim type.

[0060] Specifically, in some implementations, a rule-driven routing strategy is adopted to further verify the accuracy of agent allocation based on entity information (such as policy number and claim type) in the user request. This is one of the key technical steps in the agent collaboration mechanism of this invention. The core objective of this step is to improve the system's response efficiency and accuracy to user requests, ensuring that each request is correctly routed to an agent with the corresponding processing capabilities, thereby achieving efficient collaboration and automated processing of business processes.

[0061] Furthermore, this step is based on a structured message routing mechanism, using predefined business rules and entity recognition results to perform agent matching. Specifically, after the user interaction layer receives a user request, the natural language processing module first performs intent recognition and entity extraction. For example, when a user enters "I want to check the claim progress for policy number 123456789," the system will recognize the intent "claim progress query" and extract the entity "policy number = 123456789." Subsequently, based on the intent and entity information, and combined with a pre-defined routing rule table (such as "claim progress query" → "claim processing agent"), the system forwards the request to the corresponding agent for processing.

[0062] Specifically, the routing rule table is typically stored in key-value pairs, containing a mapping between intent category, entity type, and target agent. For example, a rule could be defined as: if the request contains "policy number" and the intent is "claims inquiry," then the request is routed to the "claims processing agent." Furthermore, the system supports contextual association in multi-turn interactions, maintaining the user's dialogue state through session_id to ensure the continuity and consistency of entity information across multiple rounds of dialogue. If a key entity is missing or its identification confidence level is below a preset threshold (e.g., 85%), the system will proactively request additional information from the user to improve routing accuracy.

[0063] Specifically, this step is widely applied in business processes such as customer consultation, claims application, and risk assessment. For example, in claims processing scenarios, the system accurately routes requests to the claims processing agent by recognizing entities such as "policy number" and "accident time," triggering subsequent processes such as image recognition and document review. In customer consultation scenarios, the system routes requests to the customer consultation agent based on entities such as "insurance product name" and "user identity," enabling personalized Q&A.

[0064] Specifically, this rule-driven routing strategy effectively improves the accuracy and response efficiency of agent allocation, reducing processing delays or errors caused by misrouting. Through structured verification of entity information, the system ensures that subsequent processing modules receive complete and compliant data input, thereby enhancing the stability and automation level of the overall business process. This step, as the foundation of the agent collaboration mechanism, provides crucial support for achieving intelligent processing of the entire insurance service process.

[0065] S3, the AI ​​agent collaborates with other agents to process business processes based on shared session context information, exchanges structured data through a unified communication protocol and executes cross-process tasks to obtain processing results.

[0066] Specifically, the system constructs a session context cache by assigning a unique session_id to each user session. This cache is used to record user intent, entity information, and intermediate processing results during multi-turn interactions. The cache structure typically uses key-value pairs, supporting fast read / write and state updates. This ensures that each agent can obtain consistent context information during collaboration, thereby improving the consistency and accuracy of task processing.

[0067] Furthermore, the communication protocol adopts a structured JSON format, with message types including request messages, response messages, and error messages. Each message type contains necessary metadata (such as session_id, timestamp, task type, parameter fields, etc.). In request messages, user intent and entity information are used as key parameters to trigger the agent's task allocation and processing logic. Response messages include processing results, status codes, and possible suggestions for subsequent operations. Error messages are used to report abnormal information, such as missing data, recognition failure, or model prediction confidence falling below a preset threshold (e.g., ...). )wait.

[0068] Specifically, this mechanism is widely applied to the entire insurance service process, including customer consultation, risk assessment, and claims processing. For example, in the claims process, the customer consultation agent may first identify the user's intent and extract the policy number, then transmit this information to the claims processing agent via a communication protocol. The latter then uses this information to invoke image recognition services to verify the authenticity of the documents and triggers the claims process after confirmation. By sharing context, each agent does not need to repeatedly ask the user for information, significantly improving interaction efficiency and user experience.

[0069] Specifically, this step enables efficient collaboration and information sharing among multiple agents, reduces system redundancy and manual intervention costs, and improves the automation level and response speed of business processing. Simultaneously, through unified communication protocols and structured data exchange, it ensures compatibility and scalability among system modules, providing solid technical support for building an intelligent and personalized insurance service system.

[0070] Furthermore, S3 includes: S31 enables asynchronous communication through JSON-formatted message queue content; the message queue content includes user ID, intent type, entity parameters, and processing status.

[0071] Specifically, the core technical principle of this step lies in combining structured data formats with message queue systems to achieve decoupling and asynchronous processing between various modules of the system, thereby improving overall response efficiency and system stability.

[0072] Furthermore, the message queue uses standard JSON format for encapsulation to ensure data readability and cross-platform compatibility. Each message contains the following key fields: user_id (unique user identifier, usually a string or integer), intent_type (intent type, such as "consultation", "claims", "risk assessment", etc., using enumeration or classification coding), entities (entity parameters, such as name, policy number, accident time, etc., stored in key-value pairs), and status (processing status, such as "pending", "processing", "completed", "failed", etc., usually a string).

[0073] Specifically, the system employs mature message middleware such as Kafka, RabbitMQ, or RocketMQ, supporting high-throughput and low-latency asynchronous communication. The message publish / subscribe mechanism routes messages based on topics or queues, ensuring that different agents can subscribe to the corresponding message streams according to their intent types. For example, the claims processing agent only subscribes to "claim_submission" type messages, while the customer consultation agent subscribes to "consultation" type messages.

[0074] Furthermore, message queues must meet the following performance requirements: message latency should be controlled within [a certain timeframe]. Throughput not less than In addition, the message format must conform to the JSON Schema specification to ensure the integrity and consistency of fields and avoid processing failures due to formatting errors.

[0075] Specifically, it is the core means to achieve efficient collaboration between intelligent agents. Through asynchronous communication mechanisms, the system can achieve parallel processing of tasks and state tracking, avoiding performance bottlenecks caused by blocking calls. At the same time, the structured message format provides standardized data input for subsequent intent recognition, entity extraction, and business processing, enhancing the system's scalability and maintainability.

[0076] S32 employs a priority arbitration mechanism to handle conflicts between agents. When multiple agents output contradictory results, the results are verified and corrected according to the preset agent priority rules.

[0077] Specifically, this step employs a priority arbitration mechanism to handle conflicts between multiple agents, ensuring the consistency and reliability of the system's output. This mechanism is primarily applied when multiple agents in the AI ​​Agent layer collaboratively process the same user request. For example, during risk assessment and claims processing, different agents may output contradictory judgments based on their respective models or rules. In this case, the system arbitrates using pre-defined agent priority rules, prioritizing the output of higher-priority agents and verifying and correcting the results of lower-priority agents.

[0078] Specifically, the priority arbitration mechanism is technically implemented based on a rule engine and state-driven conflict detection logic. During system initialization, each agent is assigned a different priority weight; for example, risk control agents have higher priority than consultation agents, and claims processing agents have higher priority than product recommendation agents. When multiple agents output different results for the same request, the system first identifies the conflict type (e.g., inconsistent product recommendations, conflicting risk levels), and then selects the output of the highest-priority agent as the final decision-making basis according to the priority rules. For the output of low-priority agents, the system performs consistency verification with the high-priority results. If a significant deviation exists, a correction process is triggered, such as re-calling the model, supplementing data for verification, or introducing a manual review mechanism.

[0079] Furthermore, the priority arbitration mechanism supports dynamic configuration, and priority weights can be adjusted based on business rules, model confidence, or historical accuracy. The conflict detection module uses similarity matching algorithms (such as cosine similarity and Jaccard index) to evaluate the degree of difference in the output results of different agents, setting thresholds (such as similarity...) This serves as the criterion for determining whether arbitration is triggered. The arbitration result is recorded in the system log for subsequent analysis and optimization.

[0080] Specifically, this mechanism is widely used in practical applications in multi-agent collaborative processing scenarios. For example, when users simultaneously submit claims and request risk assessments, the system prioritizes processing the risk assessment results to determine the feasibility of the claim. Through priority arbitration, the system effectively avoids business chaos caused by inconsistent outputs from multiple agents, improves overall processing efficiency and decision-making quality, and is a key technical link in achieving efficient agent collaboration and stable system operation.

[0081] S4 integrates the processing results of all AI agents, generates a final service response that conforms to the insurance business logic, and feeds it back to the user through the user interaction layer.

[0082] Specifically, this step uses a unified response generation engine to structurally integrate the outputs of the customer consultation agent, risk assessment agent, and claims processing agent, ensuring the consistency of the final response in terms of semantics, format, and business logic.

[0083] Specifically, firstly, after completing their tasks, each agent returns the processing results to the response generation module in structured JSON format. For example, the customer inquiry agent returns the semantic parsing results of the user's question and the answer matched by the knowledge base; the risk assessment agent returns the user's risk score (e.g., ...). The system displays the information and its corresponding risk level (e.g., "high risk"); the claims processing agent returns the document review status (e.g., "approved") and the compensation amount (e.g., 50,000 yuan). The response generation module identifies the type and status of the current user request through the session context (session_id) to ensure contextual consistency of information.

[0084] Furthermore, the response generation engine dynamically populates structured data into the response template based on preset business rule templates and Natural Language Generation (NLG) strategies. For example, for a claim application, the system can generate a message such as, "Your claim application has been approved, the compensation amount is 50,000 yuan, and it is expected to be paid out within [time period]." The system uses phrases like "payment within one business day" to adjust the tone and focus of recommended content in personalized service scenarios, combining user profiles with behavioral preferences and risk levels to improve user acceptance and satisfaction.

[0085] Specifically, the response generation module must meet the following requirements: the response generation delay should be controlled within... Within milliseconds, ensure a smooth user experience; response accuracy must reach over 95%, continuously optimized through A / B testing and user feedback mechanisms; support the recall and referencing of multi-turn dialogue states to ensure no information is lost.

[0086] Specifically, this step is widely used in practical applications for scenarios such as user inquiries, claims feedback, and risk assessment report generation. Especially during off-peak hours or under high-concurrency requests, it can significantly reduce the cost of manual intervention and improve service response efficiency and consistency. Through this mechanism, the system achieves closed-loop processing from multi-agent collaboration to final service output, providing users with a precise, compliant, and natural interactive experience, demonstrating the innovation and practicality of this invention in intelligent insurance services.

[0087] Furthermore, S4 includes: S41 dynamically fills in slot values ​​based on a preset response template to generate a natural language response containing the user's name, processing result type, and key data.

[0088] Specifically, the technical implementation of this step is based on a dialogue management mechanism of "state-driven + slot filling". By combining the results of intent recognition and named entity recognition (NER), structured data is mapped to a preset natural language template, thereby generating response content that conforms to business logic and is semantically fluent.

[0089] Specifically, first, the system uses the intent recognition module to determine the type of business requested by the user (such as claims inquiry, risk assessment, product recommendation, etc.), and then uses the NER module to extract key entity information, such as user name, policy number, accident time, risk score, etc. This entity information is stored in the structured state of the current session, organized in the form of key-value pairs, such as {name: “Li Ming”, policy_id: “123456”, result_type: “claim progress”, status: “under review”}. Subsequently, based on the current intent and status information, the system selects a matching template from a preset response template library, such as “Inquiring about the {result_type} of your policy {policy_id}, current status is {status}”.

[0090] Furthermore, the system supports multiple template types, each corresponding to a specific business scenario, and defines the required set of slots and their data types. For example, the claims progress query template needs to be filled with the policy number (string), current status (enumerated value, such as "under review" or "paid"), and estimated processing time (timestamp or relative time expression). The slot filling process must ensure data integrity and consistency. If a key slot is missing (such as the policy number not being identified), the system will trigger a multi-round interaction mechanism to proactively ask the user for the missing information.

[0091] Specifically, this step is widely used in scenarios such as customer inquiries, claims processing, and risk assessment feedback. For example, in the claims processing workflow, after image recognition technology confirms the authenticity of the materials, the system will automatically fill in information such as the user's name, policy number, and compensation amount into a preset template, generating a response such as "Dear {user_name}, your policy {policy_id} has been approved, and the estimated compensation amount is {amount} yuan." This mechanism not only improves the accuracy and efficiency of responses but also significantly reduces the need for manual intervention.

[0092] Specifically, this step automates the generation of natural language responses, ensuring the accuracy and consistency of information delivery. Simultaneously, it supports flexible expansion through template design, adapting to the personalized expression needs of different business scenarios. Combined with context management and slot-filling strategies, the system can generate natural language responses that match the user's context, thereby improving customer experience and service satisfaction.

[0093] S42 uses an A / B testing mechanism to evaluate the effectiveness of the generated response and optimizes the template content and filling strategy based on user click-through rate and satisfaction indicators.

[0094] Specifically, this step aims to improve user interaction experience and business conversion efficiency. At the technical implementation level, the A / B testing mechanism employs a multi-version parallel comparison approach. The response content for the same user request is generated into multiple variations according to a preset strategy (such as different template structures, different filling orders, or different recommended content), and randomly assigned to user groups for display. Each version's response content is dynamically populated using a structured template containing variable slots (such as product name, recommendation reasons, service prompts, etc.). The population strategy is dynamically generated by the dialogue management module based on user profiles, historical behavior, and the current context. During testing, the system records user response behavior for each version using a unique session identifier (session_id), including key metrics such as click-through rate (CTR), dwell time, and satisfaction rating (e.g., a 1-5 scale).

[0095] Furthermore, the system sets up multiple evaluation dimensions, among which click-through rate is the core indicator for measuring users' interest in the response content; satisfaction score... ,in Indicates the first User ratings This represents the total number of users participating in the test. The system also supports setting confidence intervals (e.g., 95%) and minimum sample size (e.g., ...). To ensure the statistical significance of the test results, after the evaluation period ends, the system compares the CTR and average satisfaction of each version and selects the optimal version as the benchmark for subsequent response templates.

[0096] Specifically, this step is widely applied in interactive scenarios such as customer inquiries, product recommendations, and claims notifications. For example, in insurance product recommendations, the system can test the impact of different ways of expressing recommendation reasons (such as "comprehensive coverage" vs. "high cost-effectiveness") on users' willingness to click; in claims notifications, it can test the effect of different template structures (such as segmented vs. list-based) on improving user comprehension efficiency. The test results can be directly fed back to the dialogue management module for updating slot filling strategies or adjusting recommendation algorithm parameters.

[0097] Specifically, by continuously optimizing the logic for generating response content through a data-driven approach, the system enhances user acceptance and trust in its output, thereby increasing user engagement and business conversion rates. Simultaneously, the A / B testing mechanism provides a closed-loop feedback path, enabling the AI ​​Agent to dynamically adjust service strategies based on actual user behavior, achieving truly personalized and intelligent services.

[0098] S5, based on the user profile building module, generates structured user profiles containing basic information, behavioral preferences, and potential needs through feature extraction and labeling.

[0099] Specifically, this step integrates multi-dimensional user data from the insurance service system to build a user profile with business value, providing data support for subsequent intelligent recommendations, risk assessments, and claims processing.

[0100] Specifically, the user profile building module first obtains users' historical behavioral data from the user interaction layer, including but not limited to consultation records, click behavior, insurance preferences, and satisfaction feedback, while combining this with static information such as age, gender, occupation, and region. This raw data undergoes structured processing through the feature extraction module, which includes rule mapping, statistical calculation, and model prediction. For example, user occupational information can be converted into risk level labels (such as "high-risk," "medium-risk," and "low-risk") using a predefined mapping table, while user consultation frequency is calculated using statistical methods to characterize their behavioral preferences.

[0101] Furthermore, the tagging process employs a multi-dimensional tagging system, dividing user characteristics into three categories: basic information tags (such as age, gender, and occupation), behavioral preference tags (such as preference for critical illness insurance and frequent consultation times), and potential demand tags (such as family responsibility period and health risk level). Tag generation relies on joint decision-making by machine learning models (such as collaborative filtering and content-based recommendation models) and a business rules engine. For example, based on a user's historical insurance purchase records and similar user behaviors, a collaborative filtering model can predict the types of insurance products a user might be interested in and incorporate them as part of the potential demand tags.

[0102] Furthermore, user profiles are typically updated daily or weekly to ensure timeliness. During feature extraction, missing values ​​are handled using mean imputation or time-series-based interpolation to maintain data integrity. During tagging, tag weights can be configured based on business priorities; for example, health risk levels are weighted higher than occupational categories to more accurately reflect user risk characteristics.

[0103] Specifically, this structured user profile is widely used in personalized insurance product recommendations, risk assessment model input, and contextual understanding in intelligent customer service. For example, in a risk assessment agent, the health risk level in the user profile can be used as one of the model input features to calculate the user's risk score; in a customer consultation agent, behavioral preferences in the profile can be used to optimize response strategies and improve user satisfaction.

[0104] Specifically, by systematically constructing structured user profiles, accurate characterization of user features is achieved, providing a reliable data foundation for subsequent AI Agent intelligent decision-making, thereby significantly improving the personalization level and processing efficiency of insurance services.

[0105] S6. Based on the user profile, dynamically adjust the recommendation algorithm parameters, use collaborative filtering and content-based recommendation algorithms to generate a customized insurance product recommendation list, and display the recommendation results through the user interaction layer.

[0106] Specifically, this step achieves precise matching of insurance products by integrating user profile data with recommendation algorithms, thereby improving user satisfaction and business conversion rates.

[0107] Furthermore, user profiles are composed of multi-dimensional features, including basic user information (such as age, gender, and occupation), behavioral preferences (such as historical consultation records, click behavior, and insurance records), and potential needs (such as family responsibility period and health risk level). The profile data is processed through feature extraction and tagging to form a structured user feature vector. ,in This vector represents the feature dimension. The recommendation algorithm dynamically adjusts the parameters of the collaborative filtering and content-based recommendation models based on this vector, such as adjusting the weights for similarity calculation. Content matching weight To meet the different preferences of different user groups.

[0108] Furthermore, collaborative filtering algorithms recommend insurance products favored by similar users by calculating behavioral similarity (such as cosine similarity or Pearson correlation coefficient) between users and similar users. Content-based recommendations calculate content relevance scores by matching features in user profiles with attributes of insurance products (such as coverage, premium, and sum assured). The final recommendation list is generated by a weighted fusion score, as shown in the following formula:

[0109] in, For the first Recommendation score for each product ,and , The value is dynamically adjusted based on the user profile; for example, a higher risk level could be increased for high-risk users. To enhance the weight of content matching.

[0110] Specifically, this step is widely used in product recommendations before users purchase insurance, personalized advice after user consultation, and risk management services for health insurance customers. Through the user interaction layer, the recommendation results are displayed in the form of images, lists, or cards, allowing users to click further, view details, or directly purchase insurance.

[0111] Specifically, through dynamic parameter adjustment and multi-algorithm fusion, the system significantly improves the personalization and accuracy of recommendations, reduces user selection costs, and enhances the matching efficiency and user satisfaction of insurance products. Simultaneously, by continuously collecting user feedback, the system forms a closed-loop optimization mechanism, further strengthening its adaptability and practicality.

[0112] This invention discloses a multi-agent collaborative insurance service process, which achieves accurate understanding of user intent and contextual consistency maintenance through natural language processing and dynamic slot filling. It dynamically allocates tasks to corresponding agents based on rule-driven routing. A unified communication protocol and priority arbitration mechanism ensure data consistency and business reliability in multi-agent collaborative processing. Finally, through templated response generation and user profile-driven personalized recommendations, a complete automated closed loop from request access to service feedback is constructed, significantly improving the efficiency, accuracy, and user experience of insurance business processing, and enhancing the overall service capabilities and operational efficiency of the system in complex insurance service scenarios.

[0113] Example 2 This invention also provides a system architecture for a multi-agent collaborative insurance service end-to-end processing method, such as... Figure 2 As shown, the system architecture includes: User Interaction Layer: Responsible for interacting with users, receiving user requests, and forwarding them to the AI ​​Agent Layer; AI Agent Layer: Contains multiple intelligent agents (customer consultation agent, risk assessment agent, claims processing agent), responsible for handling user requests and business logic; Customer Consultation Agent: Queries / updates user information and queries insurance product information; Risk Assessment Agent: Reads / writes risk assessment data and invokes risk assessment models; Claims Processing Agent: Reads / updates claims records, invokes external database queries, and invokes API interfaces (such as image recognition services); Tool Invocation Layer: Responsible for invoking external tools and services to support the AI ​​Agent Layer, including database systems, third-party service APIs, machine learning model libraries, etc.; Data Layer: Responsible for storing insurance business-related data and providing data access interfaces, including user information tables, insurance product tables, risk assessment record tables, claims record tables, etc.

[0114] Furthermore, the business processes of user interaction (such as...) Figure 3 (As shown) includes: User Initiates Inquiry: The user initiates an inquiry through the user interaction layer. The customer inquiry agent receives and understands the user's question. Knowledge Base Query / Question-Answer System: The agent queries the knowledge base or uses the question-and-answer system to find relevant answers. Reply Generated and Sent to User: After finding the answer, the agent generates a reply and sends it to the user. User Satisfaction Assessment: The user judges whether they are satisfied with the reply. If not, they provide further explanation; if satisfied, the inquiry ends. User Submits Claim Application: The user submits a claim application through the user interaction layer. The claims processing agent receives and conducts a preliminary review of the claim materials. Image Recognition Verification of Material Authenticity: The agent uses image recognition technology to verify the authenticity and completeness of the materials. Trigger Claims. The claims process is as follows: Payment process and updated claims records: When the materials are authentic and complete, the claims process is triggered, the compensation amount is calculated, and the claims record is updated to the database; Sending compensation result notification to the user: After the claims process is completed, the agent sends a compensation result notification to the user; User requests risk assessment: The user requests a risk assessment through the user interaction layer, and the risk assessment agent collects user information and historical data; Data preprocessing and feature engineering: The agent preprocesses and performs feature engineering on the data to prepare it for input into the risk assessment model; Calculation using the risk assessment model: The agent uses the risk assessment model to perform calculations and generate a risk assessment report; Sending the risk assessment report to the user: The agent sends the risk assessment report to the user.

[0115] Furthermore, the system architecture of an insurance service end-to-end processing method based on multi-agent collaboration according to an embodiment of the present invention also includes: Customer Inquiry Intelligent Agent: Utilizes Natural Language Processing (NLP) technology to understand user inquiries, including intent recognition and entity extraction; generates responses that conform to insurance business logic by calling knowledge bases or question-and-answer systems based on user questions; records user inquiry history to support personalized services. Risk Assessment Intelligent Agent: Receives personal information and historical data provided by users, such as age, gender, occupation, and health status; uses big data analytics and machine learning algorithms to conduct risk assessments, calculating risk scores or risk levels; recommends suitable insurance products or provides risk management advice based on the risk assessment results. Claims Processing Intelligent Agent: Automates the processing of claims applications, receives claim materials submitted by users, such as medical certificates and expense lists; uses image recognition technology to verify the authenticity and completeness of claim materials; upon approval, automatically triggers the claims process, updates the claims record, and notifies the user of the claims result.

[0116] This invention discloses a system architecture for a multi-agent collaborative insurance service process. By constructing a multi-layered service system comprising a user interaction layer, an AI agent layer, a tool invocation layer, and a data layer, it achieves an end-to-end intelligent closed loop for insurance business, from consultation and acceptance to risk assessment and claims processing. This system architecture utilizes natural language processing technology to accurately understand user intent. Through specialized agent collaboration and external service integration, it ensures accurate execution of business logic and data consistency, significantly improving the processing efficiency, automation level, and decision-making accuracy of insurance services. While reducing operating costs, it optimizes the user experience and enhances the system's adaptability and service effectiveness in complex insurance business scenarios.

[0117] Example 3 To achieve the above invention, embodiments of the present invention also provide a detailed implementation of a method for processing the entire insurance service process based on multi-agent collaboration, including: S101, Natural Language Processing. Specifically, it includes: Intent Recognition: Using deep learning models (such as LSTM, BERT, etc.) to semantically understand user input and identify user intent; intent categories include inquiring about insurance products, checking claim status, and applying for claims; Entity Extraction: Extracting key information from user input, such as name, age, and insurance product name; using Named Entity Recognition (NER) technology to achieve entity extraction; Named Entity Recognition (NER): Named Entity Recognition (NER) is a key technology in natural language processing, used to automatically identify and extract specific types of keyword information from user input text, such as name, age, ID number, insurance product name, policy number, and accident time; In intelligent insurance service systems, NER is used in conjunction with intent recognition: while intent recognition determines what the user "wants to do" (such as applying for claims or inquiring about products), NER is responsible for extracting key information such as "who exactly, what product, and when," providing structured data support for subsequent business processing.

[0118] The NER implementation process includes: text input: receiving natural language input from users (e.g., "My name is Li Ming, I want to check the progress of my car insurance claim"); semantic encoding: using a pre-trained model (e.g., BERT) to convert the text into context-sensitive word vectors; entity recognition: using a deep learning model (e.g., BERT-CRF or BiLSTM-CRF) to label each word and identify entities of predefined categories (e.g., name, insurance type, etc.), using the BIO annotation format; and result output: extracting structured information (e.g., {name: "Li Ming", insurance type: "car insurance"}) for subsequent business logic.

[0119] Among them, dialogue management (such as Figure 4As shown, it includes: Design goals: Understanding user intent and key entities (from the intent recognition and NER modules); Supporting multi-turn interactions: Able to ask for missing information and clarify ambiguous requests; Maintaining contextual coherence: Remembering key information from historical dialogues (such as provided names, policy numbers, etc.); Generating natural and accurate responses: Dynamically generating responses by combining business rules and dialogue status.

[0120] Furthermore, the implementation process of dialogue management includes: Dialogue flow design: adopting a "state-driven + slot filling" model. The system determines the required information (such as name, policy number, accident time, etc.) based on the user's current intent (such as "apply for a claim"), and gradually guides the user to complete the missing content until the business execution conditions are met; Context management: each round of dialogue is associated through a unique session ID (session_id). The system maintains a structured dialogue state, recording the identified intents and extracted entities (i.e., "slots"). Historical information remains available throughout the session's validity period and supports cross-round referencing; Multi-round interaction mechanism: if key information is missing or ambiguous, the system proactively asks follow-up questions (such as "What is your policy number?"); if the information is complete, it calls backend services (such as the claims query interface) and generates a natural language response to return to the user; Response generation method: dynamically filling slot values ​​based on preset templates (such as "Checking the claims progress of policy {policy_id} for you..."), balancing accuracy and response efficiency. In the future, it can be expanded to generate more flexible statements using a lightweight NLG model.

[0121] S102, Risk Assessment Algorithm. Specifically, it includes: Data preprocessing: Cleaning, deduplication, and normalization of user-provided personal information and historical data; handling missing values ​​by using interpolation, mean imputation, and other methods to complete the data; Feature engineering: Extracting risk assessment-related features, which are typically constructed from raw data through rule mapping, statistical calculations, or domain knowledge. Examples include: Age groups: converting birth dates to integer ages or dividing them into intervals (e.g., "18–30", "31–50"); Occupation categories: mapping occupational codes or text descriptions to predefined risk levels (e.g., "high risk", "medium risk", "low risk"); Health indicators: such as BMI (Body Mass Index) calculated using formulas. ; Historical claims frequency: Statistics on the number of claims made over the past N years.

[0122] Furthermore, the implementation process of feature engineering includes: extracting raw fields from business databases or user forms; applying rules or functions to generate new features; selecting features that are highly correlated with the target variable (such as "high risk") (which can be evaluated through Pearson correlation coefficient, information gain, etc.); feature encoding: since machine learning models can only process numerical inputs, categorical variables need to be encoded. Common encoding methods include one-hot encoding and label encoding: One-hot encoding: converts each category value into a binary vector. For example, "Occupation = Teacher" → [1, 0, 0], "Doctor" → [0, 1, 0]. Suitable for cases where the categories are unordered and few in number; Label encoding: maps categories to integers (such as "low = 0, medium = 1, high = 2"). Suitable for ordered categories or tree models (such as random forests), but may introduce spurious size relationships.

[0123] Specifically, model training includes: training a risk assessment model using machine learning algorithms (such as logistic regression, random forest, XGBoost, etc.); and optimizing the model by selecting the optimal parameter combination to improve model accuracy.

[0124] Furthermore, the implementation steps for model training include: data preparation: dividing the dataset into training, validation, and test sets (in a 7:2:1 ratio); feature input: inputting the encoded feature matrix. (Each row represents a user, and each column represents a feature) and label vector (0 = low risk, 1 = high risk); Model selection and training: Select algorithms suitable for structured data (such as logistic regression, random forest, XGBoost), and train by optimizing the following general objective function: Where L is the cross-entropy loss, The regularization term is used to balance prediction accuracy and model complexity; hyperparameter tuning: grid search or Bayesian optimization is used on the validation set to adjust key parameters (such as learning rate, tree depth, regularization strength, etc.) and select the optimal combination using metrics such as AUC and F1-score; model evaluation and deployment: the final model performance is evaluated on the test set, and after confirming that it meets the business accuracy requirements, it is encapsulated as a service interface for online risk scoring to call.

[0125] Specifically, risk assessment includes: inputting user data into a trained model to calculate a risk score or risk level; providing users with personalized insurance product recommendations or risk management advice based on the risk score or risk level; Risk score: a predicted probability value between 0 and 1 output by the model, reflecting the likelihood that the user belongs to a "high-risk customer" (e.g., 0.82 indicates that the user's characteristics are highly similar to historical high-risk groups); Risk level: dividing the score into several levels, for example: low risk: score < 0.3; medium risk: 0.3 ≤ score < 0.7; high risk: score ≥ 0.7; The system automatically provides differentiated services to users based on risk level: recommending more comprehensive insurance products (such as high-value medical insurance) and matching them with health management advice (such as "suggesting to increase regular physical examinations") for high-risk users; matching standard products with higher cost-effectiveness for low-risk users to improve experience and conversion efficiency.

[0126] S103, Image Recognition Technology. Specifically, this includes: Image Preprocessing: Preprocessing user-submitted claim materials, such as cropping, rotating, and denoising; converting the image into a digital signal for subsequent processing; User-uploaded claim images (such as invoices, ID cards, etc.) are essentially continuous optical signals. To facilitate computer processing, they need to be converted into digital signals, i.e., a numerical matrix composed of pixels.

[0127] Furthermore, image recognition technology mainly includes two steps: sampling: dividing the image into pixels at fixed intervals (e.g., 1024×768 pixels); quantization: mapping the color or grayscale value of each pixel to an integer. For example, in a grayscale image, each pixel is represented by an integer from 0 to 255 (0 for black, 255 for white). Ultimately, an image is represented as a two-dimensional (grayscale) or three-dimensional (color, including R / G / B channels) digital matrix. Where H and W are the image height and width, and C = 1 (grayscale) or 3 (color). This matrix can be directly input into subsequent image recognition models for processing.

[0128] Specifically, image recognition technology also includes: feature extraction: using deep learning models such as convolutional neural networks (CNN) to extract image features; features include text content, image structure, color information, etc.; image recognition: inputting the extracted features into a classifier to perform image recognition; The classifier is the core module in image recognition, used to determine the category of an image based on extracted image features. Commonly used classifiers include: Support Vector Machine (SVM): which separates feature vectors of different categories by finding the optimal separating hyperplane, suitable for small to medium-sized data with well-defined features; Neural Networks (such as CNN): which automatically learn the hierarchical features of images through multi-layer nonlinear transformations, especially good at processing complex images (such as tickets, scene photos), and performs excellently in tasks such as Optical Character Recognition (OCR) and object detection. Specifically, in this system, the classifier receives preprocessed image features and outputs recognition results, such as "medical invoice", "front of ID card", "vehicle accident scene" or the text content therein, to provide structured information support for claims review.

[0129] Furthermore, the results are verified: the recognition results are verified to ensure accuracy; if the recognition results are inaccurate, they can be manually adjusted or re-recognized.

[0130] Furthermore, the system automatically verifies the image recognition results, mainly based on the following standards: Confidence threshold: The classification or OCR result must reach the preset confidence level (e.g., ≥90%), otherwise it is considered low confidence; Business rule verification: For example, whether the invoice number conforms to the format, whether the ID number is legal, whether key fields are missing, etc.; Consistency check: Whether the information between multiple materials is logically consistent (e.g., whether the name and policy number match).

[0131] Specifically, depending on the circumstances, the following results may occur: High confidence and compliance: Automatic approval, proceeding to the next step of the claims process; Low confidence or abnormal format: Marked as "Pending manual review", prompting customer service intervention; Obvious error (such as identifying images not related to claims): Triggering a re-upload or re-identification process.

[0132] This mechanism ensures efficiency while effectively controlling business risks caused by identification errors.

[0133] S104, Agent Cooperation Mechanism. Specifically, it includes: Communication Protocol: Design a unified communication protocol to define the format and process of information exchange between intelligent agents. Messages adopt a structured JSON format and include three categories: Request messages: initiating a task (including user ID, intent, and parameters); Response messages: returning processing results; Error messages: providing feedback on exceptions and error codes.

[0134] Furthermore, all communication is delivered asynchronously via message queues, ensuring that the process is orderly and traceable.

[0135] Specifically, the intelligent agent collaboration mechanism also includes: Task allocation: Based on user requests and preset business rules, the coordinator dynamically allocates tasks to the most suitable intelligent agent. For example, when a user inquires about insurance products, the system automatically routes the request to the "customer consultation intelligent agent"; when applying for a claim, the task is handled by the "claims processing intelligent agent." The allocation logic is based on the intent recognition result and the task-intelligent agent mapping table, requiring no complex formulas and primarily driven by rules; Information sharing: Intelligent agents exchange necessary data, including basic user information, risk assessment results, and claims records, by sharing session context (such as a cache based on the user's session_id). The criteria for "necessary" are: whether the information is used by other intelligent agents for decision-making, whether it affects business consistency, and whether it has passed privacy compliance verification. Information sharing improves collaboration efficiency and processing accuracy; Conflict resolution: A conflict resolution mechanism is designed to handle conflicts and contradictions between intelligent agents. When multiple intelligent agents simultaneously process the same request or output contradictory results (such as recommending different products), the system automatically arbitrates through a priority mechanism (such as risk control intelligent agents having higher priority than consultation intelligent agents) or the principle of latest timestamp winning. If the conflict cannot be resolved automatically, it will be transferred to manual review to ensure that the final result is consistent and reliable.

[0136] S105, Personalized Service Mechanism. Specifically, it includes: User profile construction: Based on users' historical interaction data (such as consultation records, click behavior, and insurance preferences) and static information (such as age, occupation, and region), a structured user profile is constructed through feature extraction and tagging. The profile includes three core dimensions: basic information (such as age and gender); behavioral preferences (such as frequently checking critical illness insurance and preferring online services); and potential needs (such as family responsibility period and health risk level).

[0137] At the same time, all features are updated regularly to ensure the timeliness of the profile.

[0138] Specifically, the personalized service mechanism also includes: personalized model training: training a personalized recommendation model using machine learning algorithms such as collaborative filtering and content-based recommendation. Model training specifically includes: Collaborative filtering: Recommendations are made based on the choices of similar users; Content-based recommendation: Insurance product attributes (such as education savings and accident insurance) are matched according to user profile characteristics (such as "having children" or "high-risk occupation"). Furthermore, the model learns preferences through historical user-product interaction data (such as clicks and insurance purchases) and generates a customized product recommendation list for each user.

[0139] Specifically, the personalized service mechanism also includes: service optimization: the system continuously collects user feedback on the recommendation results (such as whether to click, whether to purchase insurance, and satisfaction rating), and optimizes the service accordingly.

[0140] Specific optimization steps include: dynamically adjusting recommendation algorithm parameters; updating user profiles to reflect the latest behavior; and verifying the effectiveness of different strategies through A / B testing, selecting the solution with higher conversion rate and satisfaction.

[0141] Ultimately, this closed-loop mechanism continuously improves recommendation accuracy and user experience.

[0142] This invention discloses a detailed implementation of a multi-agent collaborative insurance service process. By integrating natural language processing, risk assessment algorithms, and image recognition technologies, it achieves accurate parsing of user intent and maintenance of contextual coherence. Furthermore, based on a unified communication protocol and task allocation mechanism, it ensures data consistency and business reliability in multi-agent collaborative processing. By combining personalized service mechanisms and user profile construction, it dynamically generates customized insurance product recommendations and service strategies, constructing a complete intelligent closed loop from request reception to service feedback. This significantly improves the automation level, decision accuracy, and user experience of insurance business processing, and enhances the system's adaptability and operational efficiency in complex insurance service scenarios.

[0143] Example 4 This invention also provides an application scenario for a multi-agent collaborative insurance service end-to-end processing method, including: Scenario 1: Intelligent handling of customer inquiries. Specifically, this includes: Scenario Description: Mr. Zhang is a potential insurance buyer who is very interested in a certain life insurance product but has many questions that need answering. He initiated an inquiry through the insurance company's official website or mobile application.

[0144] System processing flow: The user interaction layer receives Mr. Zhang's consultation request and forwards it to the customer consultation agent in the AI ​​Agent layer; the customer consultation agent uses natural language processing (NLP) technology to understand Mr. Zhang's question, including intent recognition and entity extraction; the agent queries the knowledge base or uses a question-answering system to find answers related to Mr. Zhang's question; a reply is generated and sent to Mr. Zhang, and the reply is accurate, clear, and in line with insurance business logic; Mr. Zhang is satisfied with the reply, and the consultation ends.

[0145] Specifically, the entire process is fully automated, requiring no human customer service intervention, which greatly improves consultation efficiency and customer satisfaction.

[0146] Scenario 2: Personalized recommendations for risk assessment. This includes: Scenario Description: Ms. Li plans to purchase a life insurance policy, but she hopes to receive a personalized risk assessment and insurance product recommendation based on factors such as her age, health condition, and family situation.

[0147] System processing flow: The user interaction layer receives Ms. Li's risk assessment request and forwards it to the risk assessment agent in the AI ​​Agent layer; the risk assessment agent collects Ms. Li's personal information and historical data, such as age, gender, occupation, and health status; it uses big data analysis and machine learning algorithms to conduct a risk assessment and calculate a risk score or risk level; based on the risk assessment results, the agent recommends suitable insurance products or provides risk management advice to Ms. Li; and sends the risk assessment report and insurance product recommendations to Ms. Li.

[0148] Specifically, the system can provide customized risk assessments and insurance product recommendations based on users' individual needs, enhancing the personalization and targeting of the service.

[0149] Scenario 3: Automated processing of claims applications. This includes: Scenario Description: Mr. Wang unfortunately suffered an accident that resulted in his injury and hospitalization. He hopes to submit a claim through the insurance company's mobile application and receive compensation as soon as possible.

[0150] System processing flow: The user interaction layer receives Mr. Wang's claim application and forwards it to the claim processing intelligent agent in the AI ​​Agent layer; the claim processing intelligent agent receives and preliminarily reviews the claim materials submitted by Mr. Wang, such as medical certificates and expense lists; it uses image recognition technology to verify the authenticity and completeness of the materials; after the materials are approved, the claim process is automatically triggered, the claim amount is calculated, and the claim record is updated to the database; after the claim process is completed, the intelligent agent sends a claim result notification to Mr. Wang.

[0151] Specifically, the entire claims application process is fully automated, which greatly shortens the claims cycle and improves claims efficiency and customer satisfaction.

[0152] Scenario 4: Intelligent Recommendation of Insurance Products. This includes: Scenario Description: Ms. Zhao is looking for suitable insurance products, but she has limited knowledge of the insurance market and doesn't know how to choose. She hopes to obtain personalized insurance product recommendations through the insurance company's intelligent system.

[0153] System processing flow: The user interaction layer receives Ms. Zhao's insurance product recommendation request; the system learns Ms. Zhao's basic information, insurance needs and preferences through the AI ​​Agent layer's customer consultation intelligent agent; based on Ms. Zhao's information, the risk assessment intelligent agent assesses her risk tolerance; according to the assessment results and Ms. Zhao's needs, the system recommends a suitable combination of insurance products for her through an algorithm model; the recommendation results are displayed to Ms. Zhao, along with detailed insurance product information and explanations.

[0154] Specifically, through the intelligent recommendation system, Ms. Zhao can easily obtain insurance products that meet her needs without spending a lot of time on market research and comparison.

[0155] Scenario 5: Insurance Service Satisfaction Survey and Analysis. This includes: Scenario Description: Insurance companies want to use intelligent systems to survey and analyze customer satisfaction with insurance services in order to understand service quality and improve service processes.

[0156] System Processing Flow: The user interaction layer periodically sends satisfaction surveys to a subset of customers; customers provide feedback and suggestions on insurance services through the surveys; the AI ​​Agent layer collects and analyzes the survey data, identifying areas of customer satisfaction and dissatisfaction; the system generates service improvement suggestion reports based on the analysis results and provides them to the insurance company's management; the insurance company adjusts its service processes and policies based on the reports to improve customer satisfaction.

[0157] Specifically, through intelligent systems, insurance companies can efficiently collect and analyze customer feedback, promptly identify problems and make improvements, thereby enhancing overall service quality.

[0158] Scenario Six: Intelligent Customer Service 24 / 7 Online Support. Specifically includes: Scenario Description: Mr. Chen encountered an insurance-related problem outside of working hours and needed immediate assistance. He sought support through the insurance company's intelligent customer service system.

[0159] System processing flow: The user interaction layer receives Mr. Chen's request and forwards it to the customer consultation agent in the AI ​​Agent layer; the customer consultation agent uses natural language processing technology to understand Mr. Chen's question and queries the knowledge base or calls relevant services to provide an answer; if the question is complex or requires human intervention, the agent will escalate the question to human customer service (if available), or have human customer service follow up after working hours; the agent provides Mr. Chen with an immediate answer or informs him that the question has been escalated; Mr. Chen receives a satisfactory answer or is informed that the question will be further processed.

[0160] Specifically, intelligent customer service systems can provide customers with real-time support around the clock, solving the problem that traditional customer service systems cannot provide services outside of working hours.

[0161] Scenario 7: Insurance Knowledge Promotion and Education. This includes: Scenario Description: Xiaoming is a young man who doesn't know much about insurance. He hopes to learn about basic insurance concepts, product types, and coverage through a smart platform.

[0162] System processing flow: The user interaction layer receives Xiaoming's insurance knowledge learning request; the AI ​​Agent layer's customer consultation intelligent agent extracts relevant insurance knowledge content from the knowledge base according to Xiaoming's learning needs; the system presents insurance knowledge to Xiaoming in various forms such as text, images, and videos through the user interaction layer; Xiaoming can ask questions at any time during the learning process, and the intelligent agent will answer his questions instantly; the system can also recommend further insurance knowledge learning resources to him based on Xiaoming's learning progress and interests.

[0163] Specifically, through an intelligent insurance knowledge dissemination and education platform, Xiaoming is able to conveniently access and learn insurance knowledge, thereby improving his insurance awareness and knowledge level.

[0164] Scenario 8: Insurance Claim Progress Tracking and Reminders. This includes: Scenario description: Xiaohong submitted an insurance claim and hopes to be able to track the progress of the claim in real time and receive reminders at key stages.

[0165] System processing flow: The user interaction layer receives Xiaohong's claim progress tracking request; the AI ​​Agent layer's claim processing intelligent agent tracks Xiaohong's claim application progress in real time and updates it to the database; the system displays real-time claim progress information to Xiaohong through the user interaction layer, such as "reviewed" or "pending payment"; at key claim milestones, such as document approval or payment amount confirmation, the system sends reminders to Xiaohong via SMS, email, etc.; Xiaohong can follow up on claim matters promptly based on the reminder information.

[0166] Specifically, through the intelligent claims progress tracking and reminder system, Xiaohong can understand the claims progress in real time and receive timely reminders at key stages, improving the transparency and efficiency of the claims process.

[0167] Scenario Nine: Health Management and Prevention Services for Insurance Clients. This includes: Scenario description: Mr. Li is a customer who has purchased health insurance. He hopes that the insurance company can provide health management and prevention services to help him better manage his health.

[0168] System Processing Flow: The user interaction layer receives Mr. Li's request for health management and prevention services; the risk assessment agent in the AI ​​Agent layer develops a personalized health management plan for him based on his personal information and historical health data; the system displays the health management plan to Mr. Li through the user interaction layer and provides relevant health information, dietary advice, exercise suggestions, etc.; the risk assessment agent also regularly monitors Mr. Li's health status and adjusts and optimizes the health management plan as needed; if Mr. Li's health status becomes abnormal, the system will promptly send him a warning message and advise him to take appropriate preventive measures or seek medical attention.

[0169] Specifically, through an intelligent health management and prevention service system, Mr. Li can obtain a personalized health management plan and understand his health status in a timely manner, thereby taking more effective preventive measures to maintain his health.

[0170] This invention, through an application scenario of a multi-agent collaborative insurance service process, effectively solves the core problems of slow response, fragmented processes, and insufficient personalization in traditional service models. The system achieves an end-to-end service loop from customer consultation and risk assessment to claims processing. Through dynamic routing and collaboration mechanisms between agents, combined with natural language processing, image recognition, and risk assessment algorithms, it significantly improves the automation level and decision-making accuracy of business processing. In practical applications, the system can provide customers with 24 / 7 intelligent services, greatly shortening business response time and optimizing service experience. Simultaneously, through continuous learning and optimization mechanisms, it continuously enhances the system's adaptability and service efficiency in complex insurance scenarios.

[0171] Example 5 This invention also provides a multi-agent collaborative insurance service end-to-end processing device 10, such as... Figure 5 As shown, the device includes: The user request receiving and intent recognition module 100 is used to receive user requests through the user interaction layer and use natural language processing technology to recognize user intent and extract key entity information to generate structured user input data.

[0172] Specifically, the BERT-CRF model is used to perform semantic encoding and entity recognition on user input, and structured data containing entities such as name, age and insurance type is output in the format of BIO annotation; based on the dynamic slot filling mechanism, by maintaining the session_id cache in the session state, the system can actively ask for missing information in cross-turn dialogue and maintain contextual coherence.

[0173] The task dynamic routing module 200 is used to dynamically route requests to the corresponding AI agents through the task allocation module based on the intent type of the structured user input data; among which, the AI ​​agents include customer consultation agents, risk assessment agents, and claims processing agents.

[0174] Specifically, the target agent type is determined by matching the intent type identification result with a preset intent-agent mapping table; a rule-driven routing strategy is adopted to further verify the accuracy of agent allocation based on entity information in the user request; the entity information includes policy number and claim type.

[0175] The intelligent agent collaborative processing module 300 is used by AI intelligent agents to collaboratively process business processes with other intelligent agents based on shared session context information. It exchanges structured data through a unified communication protocol and executes cross-stage tasks to obtain processing results.

[0176] Specifically, asynchronous communication is achieved through message queue content in JSON format; the message queue content includes user ID, intent type, entity parameters and processing status; a priority arbitration mechanism is used to handle conflicts between agents, and when multiple agents output contradictory results, the results are verified and corrected according to the preset agent priority rules.

[0177] The service response integration and feedback module 400 is used to integrate the processing results of all AI agents, generate the final service response that conforms to the insurance business logic, and feed it back to the user through the user interaction layer.

[0178] Specifically, based on a preset response template, slot values ​​are dynamically filled to generate a natural language response containing the user's name, processing result type, and key data; the generated response is evaluated through an A / B testing mechanism, and the template content and filling strategy are optimized based on user click-through rate and satisfaction indicators.

[0179] Furthermore, an insurance service end-to-end processing device based on multi-agent collaboration also includes: a user profile construction module, used to generate a structured user profile containing basic information, behavioral preferences, and potential needs through feature extraction and labeling based on the user profile construction module; and a recommendation algorithm parameter adjustment module, used to dynamically adjust the recommendation algorithm parameters according to the user profile, generate a customized insurance product recommendation list using collaborative filtering and content-based recommendation algorithms, and display the recommendation results through a user interaction layer.

[0180] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0181] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0182] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for processing an insurance service whole process based on multi-agent cooperation, characterized in that, Comprise: S1, receiving user request through user interaction layer and identifying user intention and extracting key entity information by natural language processing technology, generating structured user input data; S2, according to the intention type of structured user input data, the request is dynamically routed to the corresponding AI agent through task allocation module; wherein, the AI agent includes customer consultation agent, risk assessment agent and claim processing agent; S3, AI agent cooperates with other agents to process business process based on shared conversation context information, exchanges structured data through unified communication protocol and executes cross-link task, and obtains processing result; S4, integrating the processing results of all AI agents, generating the final service response conforming to the insurance business logic, and feeding back to the user through the user interaction layer.

2. The method of claim 1, wherein, Said S1 comprises: S11, adopting BERT-CRF model to perform semantic coding and entity recognition on user input, and outputting structured data containing name, age and insurance type entity through BIO labeling format; S12, based on dynamic slot filling mechanism, through maintaining session_id cache in conversation state, active questioning and context coherence maintenance of missing information in cross-turn dialogue are realized.

3. The method of claim 1, wherein, Said S2 comprises: S21, matching the identification result of intention type with the preset intention-agent mapping table to determine the target agent type; S22, adopting rule-driven routing strategy, further verifying the accuracy of agent allocation according to entity information in user request; wherein, the entity information includes policy number and claim type.

4. The method of claim 1, wherein, Said S3 comprises: S31, realizing asynchronous communication through JSON format message queue content; wherein, the message queue content contains user ID, intention type, entity parameter and processing state; S32, adopting priority arbitration mechanism to process conflicts between agents, when multiple agents output contradictory results, according to the preset agent priority rule, the results are checked and corrected.

5. The method of claim 1, wherein, Said S4 comprises: S41, based on the preset response template, the slot value is dynamically filled to generate natural language reply containing user name, processing result type and key data; S42, through A / B test mechanism, the effect of generated response is evaluated, and the template content and filling strategy are optimized according to user click rate and satisfaction index.

6. The method of claim 1, wherein, Further comprising: S5, based on user portrait construction module, generating structured user portrait containing basic information, behavior preference and potential demand through feature extraction and labeling; S6, according to the user portrait, dynamically adjusting the recommendation algorithm parameters, generating customized insurance product recommendation list by using collaborative filtering and content-based recommendation algorithm, and displaying the recommendation result through user interaction layer.

7. An insurance service whole-process processing device based on multi-agent cooperation, characterized in that, Comprise: User request receiving and intention identifying module, for receiving user request through user interaction layer and identifying user intention and extracting key entity information by natural language processing technology, generating structured user input data; The task dynamic routing module is configured to route the request to a corresponding AI agent through the task allocation module according to the intention type of the structured user input data, wherein the AI agent includes a customer consultation agent, a risk assessment agent and a claim processing agent. The agent collaborative processing module is configured to process the business process by the AI agent based on the shared conversation context information and other agents, exchange structured data through a unified communication protocol and perform cross-link tasks to obtain a processing result. The service response integration and feedback module is configured to integrate the processing result of all AI agents, generate a final service response in line with the insurance business logic and feed back to the user through the user interaction layer.

8. The apparatus of claim 7, wherein, The user request receiving and intention recognition module is further configured to: perform semantic coding and entity recognition on the user input by using a BERT-CRF model and output structured data containing name, age and insurance type entity in a BIO tagging format; based on a dynamic slot filling mechanism, maintain a session_id cache in the conversation state to actively ask for missing information and maintain context coherence in cross-round dialogues.

9. The apparatus of claim 7, wherein, The task dynamic routing module is further configured to: determine the target agent type by matching the recognition result of the intention type with a preset intention-agent mapping table; adopt a rule-driven routing strategy to further verify the accuracy of agent allocation according to the entity information in the user request, wherein the entity information includes a policy number and a claim type.

10. The apparatus of claim 7, wherein, The agent collaborative processing module is further configured to: achieve asynchronous communication through JSON format message queue content, wherein the message queue content includes a user ID, an intention type, entity parameters and a processing state; adopt a priority arbitration mechanism to process conflicts between agents, and when multiple agents output contradictory results, perform result verification and correction according to a preset agent priority rule.

11. The apparatus of claim 7, wherein, The service response integration and feedback module is further configured to: based on a preset response template, dynamically fill in slot values to generate a natural language reply containing the user's name, the processing result type and key data; perform effect evaluation on the generated response through an A / B test mechanism, and optimize the template content and filling strategy according to the user click rate and satisfaction index.

12. The apparatus of claim 7, wherein, Further comprising: a user portrait construction module configured to generate a structured user portrait containing basic information, behavior preferences and potential needs by feature extraction and labeling based on the user portrait construction module; a recommendation algorithm parameter adjustment module configured to dynamically adjust the recommendation algorithm parameters according to the user portrait, generate a customized insurance product recommendation list by using collaborative filtering and content-based recommendation algorithm and display the recommendation result through the user interaction layer.

Images