A multi-role group chat task-oriented agent system and implementation method

Abstract

This invention discloses a task-oriented intelligent agent system and its implementation method for multi-role group chats, including a message broker, a message processing module, a session management module, a multimodal semantic understanding module, a large model context management module, a task execution module, and a business system interface module. An intelligent assistant participates in the multi-role group chat in an observer mode. Message processing units are divided using session splitting and message window mechanisms. A complete dialogue context is constructed based on a three-layer dynamic context embedding mechanism, achieving unified semantic fusion processing of multimodal messages. Task execution is completed by calling the business system through a predefined workflow. This invention solves the problems of complex information processing, low task execution efficiency, and difficulty in context understanding in existing technologies for multi-role group chat scenarios. It significantly improves the accuracy and timeliness of task processing without interfering with normal communication, and is applicable to various enterprise business scenarios such as sales ordering, customer support, and after-sales processing.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a task-oriented intelligent system and its implementation method in multi-role group chat. Background Technology

[0002] Currently, enterprises widely use instant messaging tools such as WeChat, DingTalk, and WeChat Work to communicate and collaborate with customers. In typical business scenarios such as sales ordering, customer support, and after-sales processing, multi-role group chats are often used. Enterprise employees need to interact with multiple members of the customer's team and perform various tasks such as checking order status, providing technical support, and handling customer complaints based on the multi-turn dialogue and information in the group chat. Completing such tasks often requires calling multiple business systems within the enterprise, such as CRM systems, order management systems, and knowledge bases, and enterprises have high requirements for the timeliness and accuracy of task processing.

[0003] However, in existing technologies, enterprise employees face numerous technical challenges in multi-role group chat scenarios oriented towards tasks: First, the large and complex amount of information generated by multiple members in a group chat makes it difficult to accurately understand and address the needs and questions of different members; second, traditional manual processing methods require employees to frequently switch between different business systems and tools, resulting in extremely low task execution efficiency and increased processing time and workload; third, the context of multi-turn dialogues includes interactions and historical records between members, which existing technologies struggle to accurately integrate and understand, easily leading to task execution deviations; fourth, the personalized needs of different customer members increase the complexity of task processing, making it difficult for existing solutions to adapt efficiently; and fifth, there are defects in the compatibility and data consistency between various enterprise business systems, making it difficult to integrate them into a unified platform for efficient invocation, resulting in significant system integration challenges.

[0004] To address these issues, some companies have attempted to introduce intelligent customer service to assist employees in handling business. However, existing intelligent customer service solutions are mostly focused on single-user chat scenarios for dialogue, lacking support for multi-role group chat scenarios. They cannot handle complex group chat content with multiple roles and multiple messages, and at the same time, they cannot meet the timeliness and complexity requirements of enterprises for task processing, thus failing to effectively solve the core pain points of enterprises in multi-role group chat business scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a task-oriented intelligent system and its implementation method for multi-role group chats, thereby solving the aforementioned technical problems in the prior art.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A task-oriented intelligent agent system for multi-role group chats, the system includes an intelligent assistant and an intelligent agent server. The intelligent assistant joins a multi-role business group chat containing customers and enterprise employees as an IM client. The intelligent agent server includes a message broker, a message processing module, a session management module, a multimodal semantic understanding module, a large model context management module, a task execution module, and a business system interface module.

[0008] The message broker is used to perform bidirectional message interaction with the IM system, receive group chat messages and forward them to the message processing module, and receive message instructions from the intelligent agent server and send them to the corresponding group chat.

[0009] The message processing module is used to create a unique session and assign a session ID based on the group ID, the member ID of the customer role, and the date. It also sets a resettable fixed time window to split the message flow within the session, and triggers the subsequent processing flow after the window ends.

[0010] The session management module is used for session creation, storage, retrieval and status update, and maintains information such as session members, message records, processing progress, and key system events and stores them in the session database;

[0011] The multimodal semantic understanding module is used to preprocess, multimodally fuse, integrate context, identify intent, and extract entities from messages of text, image, voice, and Excel / PDF file types, and extract unified semantic representations and task-related key entities.

[0012] The large model context management module is used to construct and dynamically embed three layers of context information: business context, session context, and domain event context, to provide the large model with complete task processing context input;

[0013] The task execution module is used to invoke a predefined workflow based on the intent recognition result and execute the task according to standardized steps;

[0014] The business system interface module is used to provide a unified enterprise business system call interface for the task execution module, and to realize interface adaptation, data conversion and error handling.

[0015] Furthermore, the intelligent assistant participates in the group chat using an observer mode, passively observing the conversations between customers and company employees without directly engaging in dialogue with customers, and only sending prompts to the group chat when necessary.

[0016] Furthermore, the time window mechanism of the message processing module is as follows: a fixed-duration waiting window is started each time a new message is received within the session. If a new message is received within the window, the timer is reset from the receiving time. If no new message is received within the window, the window is determined to end. Each message within the window is a semantically complete independent processing unit.

[0017] Furthermore, the multimodal fusion processing of the multimodal semantic understanding module includes: extracting the main content of the image message using the qwen3-vl-plus visual understanding and reasoning model; converting the voice message into text using speech recognition technology; parsing the Excel / PDF file message into text and then extracting the target content using a large model; and directly performing semantic analysis on the text message.

[0018] Furthermore, the three-layer context of the large model context management module is specifically as follows:

[0019] Business context includes business attributes, customer company information, customer addresses, and other business knowledge obtained from enterprise business systems;

[0020] Session context, including message records of the current window and historical sessions retrieved from the session management module, task processing progress, and session member information;

[0021] Domain event context includes key business events that have a significant impact on task processing, such as order creation and product delisting, which are stored in the form of structured data; the large model context management module reserves slots in the large model prompt words and dynamically fills in the corresponding context information according to the currently processed task.

[0022] Furthermore, the business system interface module develops corresponding adapters for different business systems such as enterprise CRM systems, order management systems, and knowledge bases to achieve unified interface calls, and performs format conversion and cleaning on the data returned by the business systems, while handling abnormal situations such as network anomalies, data format errors, and interface call failures.

[0023] A method for implementing a task-oriented intelligent agent system in a multi-role group chat includes the following steps:

[0024] S1. Add the smart assistant as an IM client to the multi-role business group chat that needs service, and grant the smart assistant the permission to receive group chat messages and send notification messages. When customers or enterprise employees send messages in the group chat, the message broker receives the group chat messages in real time and forwards them to the message processing module.

[0025] S2. The message processing module starts the message window mechanism, sets a fixed-duration waiting window for the current session, and resets the window and restarts the timer when a new message is received in the window.

[0026] S3. Determine if the message window has ended. If it has not ended, return to step S2 and continue waiting. If it has ended, execute step S4.

[0027] S4. The session management module creates new sessions or updates existing sessions based on group ID, customer member ID, and date, and stores the message records in the window to the session database.

[0028] S5. The task execution module obtains the current session context from the session management module and starts the multimodal semantic understanding process;

[0029] S6. The multimodal semantic understanding module preprocesses and fuses multimodal messages within the window to extract a unified semantic representation.

[0030] S7, the multimodal semantic understanding module integrates the fused messages of the current window with the content of historical conversations to construct a complete dialogue context;

[0031] S8, the large model context management module, retrieves the business context, session context, and domain event context to build a complete task processing context input for the large model;

[0032] S9. The large model analyzes the integrated context to identify user intent and extract key entity information related to the task.

[0033] S10. The task execution module calls the corresponding predefined workflow to execute the task according to standardized steps based on the identified user intent.

[0034] S11. The task execution module calls the corresponding enterprise business system through the business system interface module. The business system interface module transforms and processes the data returned by the business system to generate the task execution result.

[0035] S12. The session management module records the task execution results and processing progress to the session database and updates the status information of the current session.

[0036] S13. The task execution process corresponding to this message window ends. The agent waits for the next message window to be triggered and repeats the above steps.

[0037] Furthermore, the preprocessing described in step S6 includes filtering invalid file messages and messages with unsupported formats sent by enterprise employees, retaining only valid message types such as text, images, voice, and Excel / PDF files.

[0038] Furthermore, the predefined workflow described in step S10 replaces the autonomous decision-making method of the autonomous intelligent agent, breaking down complex tasks into standardized execution steps to ensure that tasks are executed according to predetermined rules.

[0039] Furthermore, the method can be applied to multi-role group chat business scenarios where enterprises use IM tools to conduct sales ordering, customer support, after-sales processing, and other business activities with customers.

[0040] The beneficial effects of this invention are:

[0041] 1. This invention designs an observer mode participation mechanism for intelligent assistants. The intelligent assistant only passively observes group chat conversations and does not directly interact with customers. It only sends prompts when necessary. This not only allows it to fully obtain the context of the conversation to complete the task, but also avoids interrupting the normal business conversation process between the enterprise and the customer, eliminates the customer's discomfort from being disturbed by the machine, and maintains a natural and smooth communication experience.

[0042] 2. This invention uses a session splitting and message window mechanism in the message processing module to divide the disordered group chat message flow into independent sessions based on group ID, customer member ID, and date. It also uses a resettable time window to split the sessions into semantically complete processing units, transforming complex group chat scenarios involving multiple people, multiple topics, and multiple time periods into multiple independent processing units. This significantly reduces the complexity of information processing and allows subsequent semantic understanding and task execution to be carried out within clear contextual boundaries.

[0043] 3. This invention achieves standardized integration of multiple business systems within an enterprise through a business system interface module, providing a unified interface call layer to solve compatibility issues in system integration; at the same time, the task execution module adopts a predefined workflow approach to automate task execution, eliminating the time overhead of frequent system switching and manual operation, transforming traditional manual processing into automated processing by intelligent agents, significantly shortening task processing time and improving enterprise business operation efficiency.

[0044] 4. This invention records the message history and processing progress of the session completely through the session management module. At the same time, the large model context management module constructs a dynamic embedding system of three layers of context: business, session, and domain event. This provides the large model with all the relevant information required to solve the task, enabling the large model to accurately understand multi-turn dialogues spanning multiple time windows, correctly grasp the coherence and logical relationship of the dialogue, effectively solve the problem of difficult context understanding in the prior art, and improve the accuracy of task execution.

[0045] 5. This invention, through a multimodal semantic understanding module, employs targeted processing techniques for various message types such as text, images, voice, and Excel / PDF files to achieve unified semantic fusion of multimodal messages, extracting a unified semantic representation. This breaks through the limitation of existing technologies that only support text messages, enabling a comprehensive understanding of various information in group chats, more comprehensive and accurate acquisition of customer needs, and improved completeness and accuracy of information understanding.

[0046] 6. The business system interface module of this invention develops dedicated adapters for different business systems of an enterprise, realizes unified interface calls, and establishes a sound error handling mechanism to handle various abnormal situations during the interface call process, solves compatibility and data consistency issues between various business systems of an enterprise, improves the stability and reliability of system integration, and adapts to the diverse business system architecture of an enterprise. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0048] Figure 1 This is a logical architecture diagram of an embodiment of the present invention;

[0049] Figure 2 This is a system flowchart of the implementation method of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, this embodiment combines three typical multi-role group chat business scenarios—enterprise sales ordering, customer support, and after-sales processing—to provide a detailed and complete practical explanation of the intelligent agent system and implementation method of this invention. In this embodiment, the intelligent assistant of the intelligent agent system joins the corresponding business group chat as an IM client of DingTalk, WeChat Work, and WeChat, respectively, and participates in the group chat interaction in the observer mode throughout the process; the seven core modules of the intelligent agent server work together to complete various business tasks, and a 30-second message waiting window is uniformly set in all scenarios. Each session is uniquely identified according to the rule of group ID + customer role member ID + date, ensuring the standardization and consistency of the technical implementation.

[0052] Before implementation in each scenario, the intelligent agent system of this invention has completed the entire process of pre-deployment:

[0053] 1. Completed in-depth integration with the enterprise's internal CRM system, order management system, after-sales maintenance system, enterprise knowledge base and other business systems. Dedicated adapters have been developed for each business system interface module to achieve standardized and unified interface calls.

[0054] 2. The large model context management module has completed the preset of enterprise business knowledge and domain event types, and reserved three types of context slots in the prompt words: business, conversation, and domain event, which can be dynamically filled according to task requirements.

[0055] 3. The multimodal semantic understanding module has completed the deployment and debugging of the qwen3-vl-plus visual understanding model and speech recognition engine, supporting the processing of all types of multimodal messages, including text, images, voice, and Excel / PDF files.

[0056] 4. The task execution module has preset standardized workflows for the three major business scenarios of sales, support, and after-sales, including sub-processes such as order inquiry, product consultation, repair progress inquiry, and complaint handling, which can be accurately invoked according to user intent.

[0057] like Figure 2 As shown, the intelligent agent system of the present invention follows a fully automated processing logic of "message reception - window processing - semantic understanding - task execution - result feedback", and the specific steps are as follows:

[0058] S1. Smart Assistant Joining the Group and Message Generation and Forwarding: Add the smart assistant to the multi-role business group chat that needs service as an IM client, grant the smart assistant permission to receive group chat messages and send notification messages, and the smart assistant enters observer mode to prepare to receive group chat messages.

[0059] When customers or company employees send messages in a group chat, the message broker captures the group chat messages in real time and forwards the messages to the message processing module on the intelligent agent server in accordance with the IM system protocol.

[0060] S2. Message Window Management: The message processing module performs session matching on the received messages. If a corresponding session exists, it updates the session directly. If not, it creates a new session and assigns a unique session ID. At the same time, it starts the message window mechanism, sets a fixed-duration waiting window for the current session, and resets the window and restarts the timer when a new message is received in the window.

[0061] S3. Window End Determination: Determine whether the message window has ended. If there are still new messages in the window, return to step S2 and continue waiting. If no new messages are received in the window, determine that the window has ended and execute the subsequent processing flow.

[0062] S4. Session Creation / Update: The session management module creates new sessions or updates existing sessions based on group ID, customer member ID, and date. It also stores all message records in the current window into the session database and initializes or updates session status information.

[0063] S5. Semantic understanding process start-up: The task execution module quickly retrieves and obtains the complete context information of the current session from the session management module, and starts the multimodal semantic understanding process;

[0064] S6. Multimodal Fusion Processing: The multimodal semantic understanding module preprocesses the multimodal messages in the current window, filters out invalid and unsupported messages, and retains only valid message types. Then, it performs multimodal fusion processing on the valid messages and uses dedicated technology to extract a unified semantic representation for different types of messages.

[0065] S7. Dialogue Context Integration: The multimodal semantic understanding module deeply integrates the fused messages of the current window with the content of historical conversations to build a complete dialogue context covering multiple rounds of dialogue.

[0066] S8. Multi-layer context construction: The large model context management module retrieves the business context from the enterprise business system and retrieves the session context and domain event context from the session management module to construct a complete task processing context input for the large model, which includes three layers of context, and dynamically fills the corresponding information in the prompt word slots.

[0067] S9. Intent Recognition and Entity Extraction: The large model performs in-depth analysis of the integrated multi-layered context to accurately identify the user's business intent and extract key entity information related to task execution, such as order number, product name, complaint type, etc.

[0068] S10. Workflow Execution and Business System Invocation: Based on the user intent identified by the large model, the task execution module invokes the corresponding predefined business workflow of the enterprise, breaks down complex tasks into standardized execution steps, and starts the task execution process sequentially according to predetermined rules.

[0069] S11. The corresponding enterprise business system is called through the business system interface module. The business system interface module realizes unified interface calls through a dedicated adapter, and transforms, cleans and processes the data returned by the business system to generate execution results that meet the task requirements.

[0070] S12, Session Status Update: The session management module records key information such as the execution result and processing progress of this task to the session database in real time, updates the current session status, and ensures the traceability of session content;

[0071] S13, Task Flow Ends: The task execution flow corresponding to this message window officially ends. The agent returns to observer mode and waits for the next message window to be triggered. The above steps are repeated to achieve continuous automated processing of tasks in multi-role group chat.

[0072] Scenario 1: Sales and Ordering Scenario - Customers inquire about the order shipment status in a group chat and send an order list Excel file.

[0073] The company's sales staff, along with the customer's purchasing manager and finance manager, form a DingTalk business group chat. The customer's purchasing manager inquires about the shipping status of order number "SD2026xxxx001" in the group and sends an Excel order list containing product details and order quantity. The company is then asked to confirm whether the shipped product category and quantity match the list. The intelligent agent system of this invention assists the sales staff in completing the entire process of order information verification and shipping status inquiry.

[0074] Group chat participation and message reception: The intelligent assistant joins the sales group chat in advance as a DingTalk client, enters observer mode, and only passively receives group chat messages without participating in active dialogue; after the customer's purchasing manager sends a text inquiry message + Excel order list, the message broker captures the multimodal messages in the group chat in real time, strictly encapsulates the messages according to DingTalk's message protocol, and then forwards them to the message processing module of the intelligent agent server.

[0075] Message processing and session creation: After receiving a message, the message processing module creates a unique session according to the rule of DingTalk group ID + customer procurement manager ID + 2026xxxx, assigns a session identifier XSS2026xxxx001, and starts a 30-second message waiting window. If no new group chat messages are generated in this window, the message processing module determines that the current message window has ended after the timer expires, and triggers the subsequent task processing process.

[0076] Session storage and context retrieval: The session management module formally creates session XSS2026xxxx001 according to the instructions of the message processing module, stores the text messages and Excel file messages in the window completely in the session database, and initializes the session status to "pending processing". It records that the session members include the enterprise sales staff, the customer's purchasing manager, and the customer's financial manager. The task execution module accurately retrieves the complete context information of the session from the session database and starts the multimodal semantic understanding process.

[0077] Multimodal semantic understanding processing: The multimodal semantic understanding module completes all stages of processing according to a preset process. First, message preprocessing is performed to filter out invalid and interfering information in the group chat, retaining only the two valid message types: text and Excel files. Then, multimodal fusion is performed, directly performing semantic analysis on text messages. After automatically parsing the Excel files into text format, the large model extracts the core information "Product A 50 pieces, Product B 30 pieces". Next, context integration is performed, combining the fused message in the current window with the historical conversation content of the group chat to supplement the background information that the customer placed the order in 2026xxxy, constructing a complete dialogue context. Finally, the large model completes intent recognition and entity extraction, accurately identifying the user's core intent as "to query the shipping status of order SD2026xxxx001 and verify whether the shipped product category / quantity is consistent with the order list", while extracting the key entities "order number SD2026xxxx001, product A, product B, and order quantity 50 / 30".

[0078] Multi-layered dynamic context construction: Based on extracted key entities, the large model context management module dynamically fills in three layers of context in preset prompt word slots, providing a complete basis for task processing for the large model. The business context is retrieved from the enterprise CRM system and includes "This customer is an A-level cooperative customer of the enterprise, the shipping address is XX province XX city, and enjoys priority shipping rights"; the conversation context is retrieved from the conversation database and includes "Customer's order record for 2026xxxy, consultation messages and Excel list information in the current window, and the composition of conversation members"; the domain event context is retrieved from the enterprise order management system and includes "This order was created in 2026xxxy, financial review and payment confirmation were completed on the same day, and it has entered the inventory preparation stage".

[0079] Task Execution and Business System Invocation: Based on the identified core user intent, the task execution module automatically invokes the sales order query workflow and executes the task according to standardized steps. First, through the dedicated adapter of the business system interface module, it calls the enterprise order management system to verify the validity of order number SD2026xxxx001 and query the real-time shipping status. Then, the system automatically compares the product details in the order management system with the categories and quantities in the customer's Excel list. Finally, the business system interface module performs format conversion and cleaning processing on the raw data returned by the order management system to generate a clear task execution result: "Order SD2026xxxx001 has been prepared and is expected to be shipped in 2026xxxz. The category / quantity is completely consistent with the order list."

[0080] Session status update and notification message sending: The session management module will record the execution result and processing progress of this task, including "completed order shipment status query + order list category / quantity verification", in the session database. The status of session XSS2026xxxx001 will be updated from "pending" to "processed". After the smart assistant completes the task processing, it will return to observer mode. After the enterprise sales staff confirms the result in the group, the smart assistant will send a notification message to the DingTalk group chat through the message broker, synchronizing the order-related information without interfering with the normal communication between the enterprise and the customer.

[0081] Scenario 2: Customer Support Scenario – Customers inquire about product technical parameters and send product images to confirm the model in a group chat.

[0082] Enterprise technical support staff, customer technical engineers, and purchasing personnel form a business group chat on WeChat. When a customer technical engineer inquires about the detailed technical parameters of an industrial product in the group and cannot accurately confirm the specific product model, they simultaneously send a picture of the product to request model confirmation. The intelligent agent system of this invention assists technical support staff in completing the task of intelligently identifying the product model and accurately querying the technical parameters.

[0083] Group chat participation and message reception: The intelligent assistant joins the customer support group chat as an Enterprise WeChat client and enters observer mode; after the customer's technical engineer sends a text inquiry message and a picture of the product, the message agent receives the text and image messages in the group chat in real time and forwards them to the message processing module of the intelligent agent server in accordance with the Enterprise WeChat message protocol.

[0084] Message processing and session creation: The message processing module creates a unique session by using the WeChat group ID + customer technical engineer ID + 2026xxxx, assigns a session identifier KFC2026xxxx001, starts a 30-second message waiting window, and determines the end of the window after the timer expires if no new messages are generated in the window, triggering the subsequent processing flow.

[0085] Session storage and context retrieval: The session management module creates session KFC2026xxxx001, stores text and image messages in the session database, initializes the session status to "pending processing", and records the session members as enterprise technical support staff, customer technical engineers, and customer purchasing personnel; the task execution module retrieves the session context from the session database and starts the multimodal semantic understanding process.

[0086] Multimodal semantic understanding processing: The multimodal semantic understanding module first completes message preprocessing, retaining text and image as the two valid message types; then, it performs multimodal fusion processing, analyzing the text message to obtain the core requirement "inquire about product technical parameters, and need to confirm the product model first," while calling the qwen3-vl-plus visual understanding and reasoning model to intelligently identify the product image and accurately extract the product model as "GY-2026 industrial controller"; next, it integrates the historical conversation content of the group chat, adding the background information "the customer purchased this series of industrial controllers for production line transformation in 202601," constructing a complete dialogue context; finally, it completes intent recognition and entity extraction through a large model, identifying the user's core intent as "confirm the product model is GY-2026 and query the detailed technical parameters of this model," and extracting the key entities "product model GY-2026, industrial controller, technical parameters."

[0087] Multi-layer context dynamic construction: The large model context management module completes the dynamic filling of three layers of context. The business context is retrieved from the CRM system and includes "This customer is an industrial equipment manufacturing enterprise that purchases this product for intelligent transformation of the workshop production line and has specific requirements for the communication protocol and interface type of the equipment"; the session context includes "Customer 202601's purchase record, the consultation message and product image in the current window, and the composition of the session members"; the domain event context is retrieved from the enterprise knowledge base and includes "The GY-2026 industrial controller completed a version upgrade in 202601, adding remote control and fault self-diagnosis functions, and the technical parameters were updated synchronously".

[0088] Task execution and business system invocation: The task execution module invokes the product technical parameter query workflow and executes the task step by step; firstly, it calls the enterprise knowledge base through the business system interface module to accurately retrieve the complete technical parameters of the GY-2026 industrial controller; then, based on the customer's business scenario (intelligent transformation of the production line), the system automatically filters out the core technical parameters required, including operating voltage, rated power, communication protocol, interface type, etc.; finally, the business system interface module converts the structured data returned by the knowledge base to generate a clear and organized list of technical parameters.

[0089] Session status update and notification message sending: The session management module records the task execution result "The product model has been intelligently identified as GY-2026 industrial controller, and the core technical parameter query has been completed" in the session database and updates the session status to "processed". After the enterprise technical support staff confirms the parameter information, the intelligent assistant sends a notification message to the enterprise WeChat group chat through the message broker, synchronizes the product model and core technical parameters, and automatically sends the complete technical parameter document to the group chat to provide customers with complete reference.

[0090] Scenario 3: After-sales service scenario – Customers report product malfunctions in group chats and send fault videos / voices, inquiring about repair progress.

[0091] The company's after-sales staff, customer equipment operators, and purchasing managers form a WeChat business group chat. The customer's equipment operator reports a malfunction in the purchased GY-2026 industrial controller in the group, and simultaneously sends a video and voice description of the malfunction (frequent equipment crashes, screen displaying error code E03). At the same time, they inquire about the progress of the equipment's return to the factory for repair. The intelligent agent system of this invention assists the after-sales staff in completing the entire process of malfunction analysis and repair progress inquiry.

[0092] Group chat participation and message reception: The intelligent assistant joins the after-sales service group chat as a WeChat client and enters observer mode; after the customer's equipment operator sends a fault feedback text message + fault site video + voice message, the message agent receives the multimodal messages in the group chat in real time, encapsulates them according to WeChat's message protocol, and forwards them to the message processing module on the intelligent agent server. The customer's purchasing manager sends a text message "This equipment has been returned to the factory for 3 days" within 30 seconds, and the message agent forwards it to the message processing module simultaneously.

[0093] Message processing and session creation: The message processing module creates a unique session using WeChat group ID + customer equipment operator ID + 2026xxxx, assigns a session identifier SH2026xxxx001, and starts a 30-second message waiting window; if a supplementary message from the customer's purchasing manager is received in the window, the message processing module restarts the timer from the time the supplementary message was received. After the timer restarts, no new messages are generated in the window, and the window is finally determined to end.

[0094] Session storage and context retrieval: The session management module creates session SH2026xxxx001, stores the video, voice, and multiple text messages in the session database, initializes the session status to "pending processing", and records the session members as enterprise after-sales staff, customer equipment operators, and customer procurement managers; the task execution module retrieves the complete context of the session and starts the multimodal semantic understanding process.

[0095] Multimodal semantic understanding processing: The multimodal semantic understanding module first performs message preprocessing, retaining three valid message types: video, voice, and text. Then, it performs multimodal fusion processing, calling the speech recognition engine to convert the customer's voice messages into text, resulting in "Frequent equipment crashes, screen displays error code E03." Through visual understanding model analysis of the fault scene video, the core fault symptom is extracted: "The equipment suddenly goes black during operation, displaying error code E03, and the fault persists after restarting." The core information from all text messages is integrated: "GY-2026 industrial controller malfunction, returned to factory 3 days ago, seeking repair consultation." The process begins by integrating the historical conversation content of the group chat, supplementing it with background information such as "the device was returned to the factory in 2026xxxx, with order number SD2026xxxx001," thus constructing a complete dialogue context. Finally, the large model is used to complete intent recognition and entity extraction, identifying the user's core intent as "analyzing the cause of the GY-2026 industrial controller's failure and querying the device's return-to-factory repair progress," and extracting key entities such as "product model GY-2026, industrial controller, failure symptoms (frequent crashes, error E03), return-to-factory duration 3 days, and order number SD2026xxxx001."

[0096] Multi-layer context dynamic construction: The large model context management module completes the dynamic filling of three layers of context. The business context is retrieved from the CRM system and includes "This customer is an enterprise A-level cooperative customer and enjoys priority after-sales repair and one-on-one technical support services"; the conversation context includes "Equipment 2026xxxx return record, fault feedback multimodal message in the current window, and supplementary information for 3 days after return"; the domain event context is retrieved from the enterprise after-sales repair system and includes "GY-2026 industrial controller error E03 is a core board contact problem, which is a common minor fault after the upgrade of this model, and the repair time is about 3-5 days".

[0097] Task Execution and Business System Invocation: Based on the identified core user intent, the task execution module invokes the after-sales maintenance progress query workflow and executes the task according to standardized steps. First, it calls the enterprise's after-sales maintenance system through the business system interface module, inputs the equipment model and order number, and queries the real-time maintenance progress of the equipment. Then, it combines the fault analysis records in the after-sales maintenance system to provide a professional analysis of the cause of the E03 error. Finally, the business system interface module performs format conversion and processing on the maintenance progress and fault analysis data returned by the maintenance system to generate a clear execution result: "The fault cause of the GY-2026 industrial controller is poor contact of the core board. Repair and preliminary testing have been completed. It can be sent back after final quality inspection. All after-sales processes are expected to be completed in 2026xxxs."

[0098] Session status update and notification message sending: The session management module records the task execution result and processing progress "Completed GY-2026 equipment fault cause analysis + return to factory repair progress query" to the session database and updates the session status to "Processed"; After the enterprise's after-sales staff confirms the repair information, the intelligent assistant sends a notification message to the WeChat group chat through the message agent, synchronizes the fault cause and repair progress, and informs the customer that the logistics information will be synchronized as soon as the quality inspection is completed. After completing this task processing, it returns to the observer mode and waits for the next message window to be triggered.

[0099] The above three scenarios are typical implementations of task-oriented intelligent agent systems in multi-role group chats. The technical solution of this invention is not limited to the above scenarios, but can also be widely applied to various multi-role group chat business scenarios such as enterprise product consultation, complaint handling, and contract docking. By simply adjusting the predefined workflow of the task execution module and the adapter of the business system interface module according to different business needs of the enterprise, the automated and standardized processing of different business tasks can be achieved, which has strong versatility and scenario adaptability.

[0100] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A multi-role group chat task-oriented agent system, characterized in that, The system includes an intelligent assistant and an intelligent agent server. The intelligent assistant joins a multi-role business group chat containing customers and enterprise employees as an IM client. The intelligent agent server includes a message broker, a message processing module, a session management module, a multimodal semantic understanding module, a large model context management module, a task execution module, and a business system interface module. The message broker is used to perform bidirectional message interaction with the IM system, receive group chat messages and forward them to the message processing module, and receive message instructions from the intelligent agent server and send them to the corresponding group chat. The message processing module is used to create a unique session and assign a session ID based on the group ID, the member ID of the customer role, and the date. It also sets a resettable fixed time window to split the message flow within the session, and triggers the subsequent processing flow after the window ends. The session management module is used for creating, storing, retrieving and updating the status of sessions, maintaining session members, message records, processing progress and key system event information and storing them in the session database; The multimodal semantic understanding module is used to preprocess, multimodally fuse, integrate context, identify intent, and extract entities from messages of text, image, voice, and Excel / PDF file types, and extract unified semantic representations and task-related key entities. The large model context management module is used to construct and dynamically embed three layers of context information: business context, session context, and domain event context, to provide task processing context input for the large model; The task execution module is used to invoke a predefined workflow based on the intent recognition result and execute the task according to standardized steps; The business system interface module is used to provide a unified enterprise business system call interface for the task execution module, and to realize interface adaptation, data conversion and error handling.

2. The multi-role group chat task-oriented agent system according to claim 1, wherein, The intelligent assistant participates in group chats using an observer mode, passively observing the conversations between customers and company employees, and sending prompts to the group chat when necessary. 3.The multi-role group chat task-oriented agent system according to claim 1, wherein, The time window mechanism of the message processing module is as follows: a fixed-duration waiting window is started every time a new message is received in the session. If a new message is received in the window, the timer is restarted from the receiving time. If no new message is received in the window, the window is determined to end. Each message in the window is a semantically complete independent processing unit. 4.The multi-role group chat task-oriented agent system according to claim 1, wherein, The multimodal fusion processing of the multimodal semantic understanding module includes: extracting the main content of image messages using visual understanding and reasoning models; converting speech messages into text using speech recognition technology; parsing Excel / PDF file messages into text and then extracting the target content using a large model; and directly performing semantic analysis on text messages.

5. The multi-role group chat task-oriented agent system according to claim 1, wherein, The three-layer context of the large model context management module are as follows: Business context includes business attributes, customer company information, and customer address business knowledge obtained from the enterprise business system; Session context, including message records of the current window and historical sessions retrieved from the session management module, task processing progress, and session member information; Domain event context, including key business events such as order creation and product delisting stored in structured data form that have a significant impact on task processing; the large model context management module reserves slots in the large model prompt words and dynamically fills in the corresponding context information according to the currently processed task.

6. The multi-role group chat task-oriented agent system according to claim 1, wherein, The business system interface module develops corresponding adapters for different business systems such as enterprise CRM system, order management system, and knowledge base, realizes unified interface calls, and performs format conversion and cleaning on the data returned by the business systems.

7. An implementation method of a task-oriented agent system in a multi-role group chat, based on the agent system of any one of claims 1-6, characterized in that, Includes the following steps: S1. Add the smart assistant as an IM client to the multi-role business group chat that needs service, and grant the smart assistant the permission to receive group chat messages and send notification messages. When customers or enterprise employees send messages in the group chat, the message broker receives the group chat messages in real time and forwards them to the message processing module. S2. The message processing module starts the message window mechanism, sets a fixed-duration waiting window for the current session, and resets the window and restarts the timer when a new message is received in the window. S3. Determine if the message window has ended. If it has not ended, return to step S2 and continue waiting. If it has ended, execute step S4. S4. The session management module creates new sessions or updates existing sessions based on group ID, customer member ID, and date, and stores the message records in the window to the session database. S5. The task execution module obtains the current session context from the session management module and starts the multimodal semantic understanding process; S6. The multimodal semantic understanding module preprocesses and fuses multimodal messages within the window to extract a unified semantic representation. S7, the multimodal semantic understanding module integrates the fused messages of the current window with the content of historical conversations to construct a complete dialogue context; S8, the large model context management module, retrieves the business context, session context, and domain event context to build a complete task processing context input for the large model; S9. The large model analyzes the integrated context to identify user intent and extract key entity information related to the task. S10. The task execution module calls the corresponding predefined workflow to execute the task according to standardized steps based on the identified user intent. S11. The task execution module calls the corresponding enterprise business system through the business system interface module. The business system interface module transforms and processes the data returned by the business system to generate the task execution result. S12. The session management module records the task execution results and processing progress to the session database and updates the status information of the current session. S13. The task execution process corresponding to this message window ends. The agent waits for the next message window to be triggered and repeats the above steps.

8. The implementation method of multi-role group chat task-oriented agent system according to claim 7, characterized in that, The preprocessing described in step S6 includes filtering invalid file messages and messages with unsupported formats sent by enterprise employees, while retaining valid message types such as text, images, voice, and Excel / PDF files.

9. The implementation method of multi-role group chat task-oriented agent system according to claim 7, characterized in that, The predefined workflow described in step S10 replaces the autonomous decision-making method of the autonomous agent, breaking down complex tasks into standardized execution steps to ensure that tasks are executed according to predetermined rules.

10. The implementation method of multi-role group chat task-oriented agent system according to claim 7, characterized in that, The method can be applied to a multi-role group chat business scenario of sales ordering, customer support and after-sales processing of a customer developed by an enterprise through an IM tool.

Images