Question and answer information processing method, device, storage medium, and program product

Abstract

This application provides a question-and-answer information processing method, device, storage medium, and program product. In this application embodiment, the main control intelligent module schedules the task intelligent modules and constrains their generation behavior based on the skill constraint descriptions of each task intelligent module, thus resolving the model illusion problem. Simultaneously, a dual-engine asynchronous collaborative mechanism is adopted: synchronously calling the rule engine ensures low-latency response, while asynchronously calling the inference engine generates more reliable solutions, resolving the contradiction between timeliness and reliability. The work order record module, as a global storage unit, enables the persistence of service information. These three components work together to achieve efficient, stable, and controllable intelligent customer service.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a question-and-answer information processing method, device, storage medium, and program product. Background Technology

[0002] With the development of artificial intelligence technology, the shortcomings of traditional human customer service in terms of operating costs, concurrent processing capabilities, and service quality stability have gradually become apparent. In contrast, rule-based customer service robots, with their standardized responses, 24 / 7 service, and significant cost advantages, can efficiently handle a large number of repetitive inquiries and are gradually becoming the mainstay of customer service systems.

[0003] However, rule-based customer service robots mainly rely on predefined rules and keyword matching, which cannot understand vague or long-tail user requests and lack contextual memory capabilities. They are especially unsuitable for after-sales scenarios such as 3C digital products, where problems are complicated, communication is frequent, and response requirements are high, resulting in users repeatedly stating their problems but finding it difficult to resolve them completely.

[0004] Against this backdrop, a new generation of customer service systems based on large models has emerged. These systems are no longer limited by predefined rules, but instead utilize the deep semantic understanding capabilities of large models to accurately capture users' implicit intentions. They also leverage attention mechanisms to maintain long-window contextual memory, enabling smooth multi-turn interactions and ultimately delivering professional and timely responses.

[0005] However, the inherent "illusion" problem of large models may lead them to fabricate incorrect answers that do not conform to actual rules. Therefore, how to effectively improve the reliability of responses without reducing their timeliness, thereby achieving efficient, stable, and controllable intelligent customer service, is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] This application provides a question-and-answer information processing method, device, storage medium, and program product to improve the reliability of responses without reducing their timeliness, thereby achieving efficient, stable, and controllable intelligent customer service.

[0007] This application provides a question-and-answer information processing method applied to a main control intelligent module. The main control intelligent module is associated with multiple task intelligent modules in a target application scenario. The method includes: receiving question information from a user in the current session and determining the work order record module corresponding to the question information; identifying the user's intent based on the question information and constructing a decision context for the current session based on the user's intent and historical service information recorded in the work order record module; planning a first service node based on the decision context; if the first service node is a solution generation node, scheduling a solution decision module corresponding to the solution generation node among the multiple task intelligent modules; the solution decision module is configured to: synchronously schedule a rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule an inference engine to generate a second response solution and write it into the work order record module; if a user's rejection information is received after outputting the first response solution, obtaining the second response solution from the work order record module and processing the question information according to the second response solution.

[0008] This application also provides an electronic device, including: a memory and a processor; wherein the memory is used to store a computer program, and the processor is coupled to the memory and used to execute the computer program to perform the question-and-answer information processing method.

[0009] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the question-and-answer information processing method.

[0010] This application also provides a computer program product, including: a computer program / instructions, which, when executed by a processor, enable the processor to implement the steps in the question-and-answer information processing method.

[0011] In this embodiment, the main control intelligent module schedules the task intelligent modules and constrains their generation behavior based on the skill constraint descriptions of each module. This addresses the issues of illusion and loss of control caused by the model's autonomous behavior, improving the reliability of responses. Simultaneously, a dual-engine asynchronous collaborative mechanism is employed. The rule engine is synchronously invoked to ensure low-latency response, while the inference engine is asynchronously invoked, and the generated solutions are written to the work order record module. This achieves a balance between rapid response and accurate calculation, resolving the conflict between timeliness and reliability. The work order record module, as a global storage unit, ensures the persistence of service information. On one hand, it provides cross-step state memory for the execution of skill constraints, ensuring the controllability of the scheduling path. On the other hand, it provides storage and backtracking capabilities for asynchronous solutions in the dual engines, thus avoiding the problems of missing context memory and inconsistencies in service solutions. These three elements work together to achieve efficient, stable, and controllable intelligent customer service. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram of the structure of a question-and-answer information processing system provided in an exemplary embodiment of this application; Figure 2 A flowchart illustrating a question-and-answer information processing method provided for an exemplary embodiment of this application; Figure 3 A schematic diagram of the architecture of a question-and-answer information processing system provided as an exemplary embodiment of this application in a real-world scenario; Figure 4 Another architectural diagram of the question-and-answer information processing system provided as an exemplary embodiment of this application in a real-world scenario; Figure 5 An interactive flowchart illustrating the improved process of the question-and-answer information processing method provided in an exemplary embodiment of this application in a real-world scenario; Figure 6 Interaction flowchart of the question-and-answer information processing method provided in an exemplary embodiment of this application in a product quality dispute resolution scenario; Figure 7 An interactive flowchart illustrating another improved process of the question-and-answer information processing method provided in an exemplary embodiment of this application in a real-world scenario; Figure 8 A flowchart illustrating the platform-user unilateral negotiation interaction of a question-and-answer information processing system provided as an exemplary embodiment of this application; Figure 9 An interactive flowchart of the bilateral negotiation process of a question-and-answer information processing system provided as an exemplary embodiment of this application; Figure 10 A schematic diagram of the structure of a question-and-answer information processing apparatus provided in an exemplary embodiment of this application; Figure 11 A schematic diagram of the structure of an electronic device provided in an exemplary embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with relevant laws, regulations, and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0015] In customer service systems based on large models, the inherent "illusion" problem of large models can lead to the creation of incorrect answers that do not conform to actual rules. Therefore, how to effectively improve the reliability of responses without reducing their timeliness, thereby achieving efficient, stable, and controllable intelligent customer service, is a pressing technical problem that needs to be solved.

[0016] To address the aforementioned technical issues, this application embodiment employs a modular architecture that links a main control intelligent module with multiple task intelligent modules. The work order record module serves as the system's global storage unit, enabling unified management and traceability of service data such as user session information and platform response schemes. The main control intelligent module identifies the user's intent and, combined with historical service information from the work order record module, constructs a decision context to achieve precise planning of service nodes. Based on the skill constraint descriptions of each task intelligent module, the behavior of the task intelligent modules is constrained to prevent them from acting autonomously and causing illusions or loss of control.

[0017] For solution generation service nodes, a dual-track parallel scheduling mechanism of rule engine and inference engine is adopted: the rule engine is scheduled synchronously to generate the first response solution and output it in real time to ensure the timeliness of the response; at the same time, the inference engine is scheduled asynchronously to generate the second response solution and pre-write it into the work order record module to adapt to complex requests; when the user's rejection information for the first response solution is received, the pre-generated second response solution is directly retrieved from the work order record module to complete the response processing.

[0018] By employing a dual-track parallel scheduling of the rule engine and the inference engine, the problem of insufficient adaptability of the rule engine to complex scenarios is solved, while effectively avoiding the illusionary defects of large models without compromising response timeliness. The work order recording module, as a context carrier, provides cross-step state memory and backtracking capabilities for the asynchronous processing of the two engines and the execution of skill constraints. The three components work together to achieve efficient, stable, and controllable intelligent customer service. The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 This is a schematic diagram of the structure of a question-and-answer information processing system provided as an exemplary embodiment of this application. Figure 1 As shown, the question-and-answer information processing system can adopt a distributed architecture with multiple intelligent modules working together, including: a main control intelligent module 110, multiple task intelligent modules 120, and a work order recording module 130.

[0020] The information question-and-answer system in this embodiment differs from the single general large model of end-to-end reasoning. The question-and-answer information processing system decomposes the full-process customer service capabilities into intelligent modules with clear responsibility boundaries, skill constraints, and independent execution capabilities. Through the orderly collaboration of multiple intelligent modules, it achieves fully controllable, illusion-free, and highly adaptable intelligent customer service.

[0021] In this embodiment, the intelligent module refers to an intelligent system unit with autonomous perception, memory, decision-making, interaction, and execution capabilities. Based on the collaborative work of the underlying model, rule engine, and work order recording module, it achieves automated question answering and task processing in specific application scenarios. In this embodiment, the specific implementation of the underlying model associated with each intelligent module is not limited. For example, a deep learning model with a relatively large model parameter scale can be used. Here, "large model" is merely an example, and this embodiment does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning model involved in this embodiment can be an AI-based LLM (Large Language Mode) or MLLM (Multimodal Large Language Model).

[0022] Among them, the main control intelligent module 110 is the central scheduling and decision-making brain of the entire question-and-answer information processing system, responsible for path planning and compliance verification of the entire service process, as well as the unified scheduling and collaborative management of the various task intelligent modules 120. In terms of implementation, the main control intelligent module 110 can be implemented as one or more service planning intelligent agents.

[0023] Unlike general-purpose large-scale models that generate reasoning in a black box without boundaries, this service planning agent is pre-configured with core responsibilities such as session management, intent recognition, context construction, and service node planning and scheduling. It has built-in standardized service flow topology and node admission rules, and does not generate content without limits. Based on the user's question information and the information from the current session, it dynamically plans service nodes. Each node transition has clear rules, ensuring that the entire service process is traceable and controllable. This fundamentally avoids the shortcomings of general-purpose large-scale models, such as uncontrollable processes and unjustified transitions in multi-round interactions, thus guaranteeing the compliance and stability of the customer service process.

[0024] Among them, multiple task intelligence modules 120 are specialized intelligent agents corresponding to specific service nodes in the entire question-and-answer information processing system. Each task intelligence module has independent skill constraint descriptions and execution capabilities. In terms of implementation, the task intelligence module 120 can be implemented as a skill-based intelligent agent corresponding to the target application scenario, and different skill-based intelligent agents can be configured for different application scenarios. Unlike the generalization capability of general large models across all scenarios, each task intelligence module 120 is customized for a specific service link in the target application scenario, with clear responsibility boundaries, skill constraint descriptions, and execution capabilities, and does not perform redundant generative reasoning.

[0025] The system allows for pre-configuration, customized construction, and dynamic generation of the task intelligence module 120 based on the characteristics, service processes, and compliance requirements of different application scenarios, achieving precise matching between scenarios and skills. For example, for 3C digital after-sales scenarios, modules such as solution decision-making, voucher verification, intelligent outbound calling, solution negotiation, and comprehensive adjudication can be configured; for e-commerce retail after-sales scenarios, modules such as refund verification, logistics tracking, and return / exchange processing can be configured; and for financial consulting scenarios, modules such as compliance verification, qualification verification, and product explanation can be configured. Each task intelligence module corresponding to any application scenario can focus on a single service link within that application scenario, thereby achieving scenario-based, lightweight, and highly reliable execution capabilities.

[0026] Each task intelligence module 120 can have a built-in skill constraint description. This skill constraint description defines the responsibility boundaries, execution rules, compliance requirements, output format, and accessible tool / engine permissions of the corresponding task intelligence module 120, thus defining the execution scope and output boundaries of the task intelligence module 120. Unlike general large models that can generate text without limit and are prone to producing illusionary content that deviates from service rules, the task intelligence module 120 in this embodiment can only perform operations and generate outputs within the scope limited by the skill constraint description. The output content must comply with predefined service rules and compliance requirements, and will not generate content that exceeds the skill boundaries, thus eliminating the illusion problem of large models from the root of the architecture.

[0027] Taking the 3C digital after-sales scenario as an example, the skill constraint descriptions and execution configurations of each task's intelligent module 120 are as follows: The solution decision module is configured as a skill-based intelligent agent for solution generation. The skill constraint description limits it to calling the rule engine and inference engine to execute response solution generation tasks, following a dual-engine scheduling rule that combines synchronous and asynchronous operations, and the output content must conform to the after-sales response specifications.

[0028] The voucher verification module is configured as a skill-based intelligent agent for verifying after-sales vouchers. The skill constraint description limits it to verifying the compliance, authenticity, and validity of after-sales vouchers uploaded by users. Its output is only the verification results and compliance judgment basis, and it has no additional content generation permissions.

[0029] The intelligent outbound calling module is configured as a skill-based intelligent agent for reaching users. The skill constraint description limits it to performing predefined outbound calling processes, user information verification, and notification delivery tasks.

[0030] The solution negotiation module is configured as a skill-based intelligent agent for after-sales solution negotiation. The skill constraint description limits it to negotiating solutions with users within the predefined after-sales policy scope and it cannot generate any commitments beyond the policy boundaries. The integrated adjudication module is configured as a skilled intelligent agent for adjudicating complex disputes. Its skill constraint description limits it to generating standardized dispute adjudication solutions based on work order context information, service rules, and after-sales policies.

[0031] In summary, the question-and-answer information processing system adopts a multi-agent collaborative architecture, which has three core advantages compared with a single general-purpose model: First, full controllability, with the entire service process dynamically planned and node-controlled by the master intelligent module. Each execution path has clear rules, making the entire process traceable and controllable. Second, low illusion risk, with each task intelligent module having its responsibilities and output boundaries defined by skill constraints, thus solving the inherent illusion defects of general-purpose models at the root of the architecture. Third, high scenario adaptability, allowing for flexible pre-configuration and construction of task intelligent modules with corresponding skills based on different application scenarios, quickly adapting to customer service needs in different fields and service stages.

[0032] In this embodiment, the main control intelligent module 110 can receive the user's problem information in the current session and determine the corresponding work order record module 130. Specifically, it can determine the resource object (e.g., an order) to which the problem information belongs and obtain the work order record module 130 corresponding to that resource object. Each time a user accesses customer service through any entry point (including but not limited to online customer service, offline customer service, telephone customer service, etc.), it is considered as starting a session; when the user actively exits the current entry point, the session times out and disconnects, or the user accesses again after switching devices / channels, it is considered as a new session. Different sessions are independent of each other and are associated and bound through resource object identifiers (e.g., order IDs). For each resource object, a unique work order record module 130 is pre-built to form a "one resource object, one work order" management model, such as "one order, one work order," to ensure the integrity and independence of the service information related to the resource object.

[0033] The work order record module 130 is the global state storage unit of the question-and-answer information processing system. As a trusted data carrier throughout the entire service lifecycle, its lifecycle is bound to the resource object (e.g., an order) processed by the system, spanning the entire execution process of the main control intelligent module 110 and various task intelligent modules 120 for the same resource object. A resource object can be associated with one or more user sessions. The user inquiries, interactive actions, session status, historical service information, execution results of each intelligent agent, and generated response schemes generated by these sessions are all uniformly recorded in the work order record module 130 corresponding to the resource object.

[0034] The work order recording module 130 is used to uniformly store service information associated with a resource object at the resource object level. The service information includes session information and execution status data. The session information includes user input, system output, and action logs during a single interaction, which is used to support the contextual coherence of multi-turn dialogues. The execution status data includes user intent, various response schemes, confirmation results, and the final executed action, which is used to ensure service continuity, scheme consistency, and full-process traceability across sessions.

[0035] Unlike the temporary context window storage mechanism of general large models, the work order recording module 130 possesses persistent, end-to-end, and traceable information storage and read / write capabilities. For different sessions associated with the same resource object, the session initiation entry point is not limited, and can cover multiple channels such as online customer service and offline customer service. The work order recording module 130 provides a unified context data read / write interface for each agent, ensuring that all agent decisions and executions are based on the same set of globally trusted data. This avoids context loss and information inconsistency issues in multi-turn interactions or multi-session scenarios, while also providing complete data support for compliance auditing and path tracing throughout the service process.

[0036] Taking a 3C digital product after-sales refund scenario as an example, a user might initiate a refund inquiry on the first day (first round of conversation), confirm the refund request and preliminary solution, and then exit. On the third day, they might revisit to inquire about the refund progress (second round of conversation), and on the fifth day, they might submit supplementary refund vouchers (third round of conversation). All the key actions generated by these different conversations spanning multiple days (request confirmation, solution notification, progress inquiry, voucher submission, etc.) will be structured and recorded in the work order record module 130 corresponding to the refund order. When the user submits vouchers in the third round of conversation, the context information of the first two rounds of conversation can be quickly retrieved through the work order record module 130, allowing for precise advancement of the subsequent review process without requiring the user to repeat their statements.

[0037] It should be noted that the after-sales service process may span several hours or even days, involving multiple interactions, document submissions, and negotiations. Therefore, this embodiment adopts an "event-driven + state persistence" architecture: user sessions, external events, and intelligent module return results can all serve as trigger events to drive the system's dynamic response; the work order record module, as the core carrier of working memory information, stores dynamic information such as current requests, executed actions, service nodes, and incomplete tasks. Each time any intelligent module is triggered (whether through online interaction or offline processing), the working memory in the corresponding work order is automatically reloaded to ensure that decisions are based on the latest service status and avoid information lag.

[0038] In this embodiment, the main control intelligent module 110 can identify the user's intent based on the problem information, and construct the decision context for the current session based on the user's intent and the historical service information recorded in the work order record module 130. After constructing the decision context for the current session, a first service node can be planned based on the decision context. The planning of the first service node can be completed by the main control intelligent module 110. For example, the constructed decision context can be used as input, and based on the built-in standardized service flow topology diagram, the service node that best matches the current session can be determined through a combination of rule matching and probabilistic reasoning. In this embodiment, the types of service nodes include, but are not limited to: scheme generation node, voucher review node, intelligent outbound call node, scheme negotiation node, comprehensive adjudication node, and manual transfer node, etc. Different service nodes correspond to different task intelligent modules 120.

[0039] For example, when the decision context shows that the user is initiating an inquiry for the first time and explicitly states the problem of "the phone cannot be charged", the main control intelligent module 110 will plan the first service node as the solution generation node; when the decision context shows that the user has submitted after-sales vouchers and is waiting for review, the main control intelligent module 110 will plan the first service node as the voucher review node; when the decision context shows that the user has repeatedly rejected the standardized solution, the main control intelligent module 110 will directly plan the first service node as the human transfer node to ensure service experience.

[0040] In this embodiment, the specific implementation method of the main control intelligent module 110 associated with the underlying model is not limited, and can be flexibly configured as a multi-model collaborative architecture or a single-model end-to-end architecture according to the application scenario.

[0041] Optionally, the main control intelligent module 110 can be associated with an AI (Artificial Intelligence) model. This AI model adopts an end-to-end unified architecture, integrating functions such as intent recognition, context building, and service decision-making within a single model. It completes the entire reasoning process from user questions to service node planning through a unified input / output interface. The single-model architecture is simpler in terms of system deployment, model management, and inference chain, and is suitable for scenarios with relatively fixed application scenarios and controllable service process complexity.

[0042] Optionally, the main control intelligent module 110 can be associated with multiple functionally decoupled AI models, including an intent recognition model, a context construction model, and a service decision model. Unlike traditional designs that integrate all functions such as intent recognition, context understanding, and service decision-making into a single complex large model, this embodiment adopts a multi-model collaborative approach. It breaks down the complex service decision-making task into three independent and clearly defined sub-tasks, each handled by a dedicated AI model with a single function, thus achieving functional decoupling. The computational resources, time cost, and number of labeled samples required to train a small model responsible for a single function are far lower than training an all-encompassing large model covering the entire process. Furthermore, when service rules change, only the individual model at the corresponding stage needs fine-tuning, without retraining the entire system. Each model only needs to construct labeled samples for its specific task; for example, the intent recognition model only needs to label user intent tags, not the complete service process, resulting in a lower labeling threshold and higher efficiency. Moreover, individual models can be independently optimized, upgraded, or replaced without affecting the normal operation of other models, enabling rapid response to changes in service needs across different industries and scenarios.

[0043] To improve the synergy and overall decision-making accuracy among multiple AI models and avoid inconsistencies in output caused by independent training of multiple models, this embodiment employs a multi-stage joint training approach to uniformly optimize the aforementioned AI models. Specifically, during training, a complete customer service session sample sequentially passes through the intent recognition model, the context construction model, and the service decision model. For this sample, three stages of annotation results are constructed: the first stage annotates the user's true intent label, the second stage annotates standard decision context features, and the third stage annotates the correct service node label.

[0044] In the loss calculation stage, a dual calculation mechanism combining single-stage loss functions and joint loss functions can be adopted: on the one hand, the single loss function of each model in its corresponding stage is calculated separately to ensure the accuracy of each model in its own task; on the other hand, the joint loss function among the outputs of the three models is calculated to measure the logical consistency and synergy between the intent recognition result, the context construction result, and the final service decision. By weighted summing the single-stage loss function and the joint loss function to obtain the total loss function, and backpropagating based on the total loss function to update the parameters of all models, multi-model collaborative optimization is achieved, effectively improving the linkage effect between models and the overall decision accuracy.

[0045] Among them, the intent recognition model can be used to perform deep semantic understanding and intent classification of user question information, and its built-in intent classification system can be flexibly configured and expanded according to different application scenarios: In e-commerce after-sales scenarios, the intent classification system includes at least the following core categories: return intent, exchange intent, repair intent, reissue of voucher intent, negotiation intent, progress inquiry intent, complaint intent, and cancellation intent. Specifically, return intent identifies user requests for product returns; exchange intent identifies user requests for product replacements; repair intent identifies user requests for product repairs; reissue of voucher intent identifies user requests for reissue of invoices, warranty cards, order vouchers, etc.; and negotiation intent identifies user requests to negotiate after-sales solutions, compensation amounts, and shipping costs.

[0046] In e-commerce pre-sales scenarios, the intent classification system includes at least the following core categories: product inquiry intent, price inquiry intent, inventory inquiry intent, promotional activity inquiry intent, purchase guidance intent, logistics inquiry intent, and after-sales policy inquiry intent. Specifically, product inquiry intent identifies user needs to inquire about product parameters, functions, materials, usage methods, and compatibility; price inquiry intent identifies user needs to inquire about product prices, discounts, and price protection policies; and inventory inquiry intent identifies user needs to inquire about product inventory status, shipping time, and shipping warehouse.

[0047] In non-e-commerce scenarios, corresponding intent classification systems can be customized according to specific industry characteristics. For example, in financial consultation scenarios, intents such as account opening consultation, loan application, wealth management product consultation, bill inquiry, and loss reporting and replacement can be configured; in government service scenarios, intents such as document processing, policy consultation, application progress inquiry, and complaint reporting can be configured; and in medical consultation scenarios, intents such as appointment booking, symptom consultation, examination report inquiry, and medication guidance can be configured.

[0048] The context construction model can be used to deeply integrate the user intent and confidence level output by the intent recognition model with the historical service information corresponding to the order recorded in the work order record module 130 to construct a complete decision context for the current session. The decision context may include: semantic features of the user's current question, the identified user intent and confidence level, basic information of the order (such as product information, order amount, order time, payment status, etc.), historical session records, the chain of executed key service actions, the current service node status, the list of incomplete tasks, and the user's historical preference information.

[0049] Based on the constructed decision context, the service decision model can accurately plan subsequent service nodes and schedule corresponding intelligent task modules 120. Through the collaborative work of multiple AI models, it can accurately understand the user's true intentions and fully grasp the historical service status of orders, providing a reliable decision-making basis for subsequent accurate planning of service nodes and scheduling of corresponding intelligent task modules, ensuring a smooth and efficient service process.

[0050] After constructing the decision context for this round of conversation, the first service node can be planned based on the decision context. The planning of the first service node can be completed by the service decision model within the main control intelligent module 110. The service decision model takes the constructed decision context as input and, based on a built-in standardized service flow topology diagram, determines the service node most suitable for this round of conversation through a combination of rule matching and probabilistic reasoning. In this embodiment, the types of service nodes include, but are not limited to: scheme generation nodes, credential verification nodes, intelligent outbound call nodes, scheme negotiation nodes, comprehensive adjudication nodes, and manual transfer nodes, etc. Different service nodes correspond to different task intelligent modules 120.

[0051] For example, when the decision context shows that the user is initiating an inquiry for the first time and explicitly states the problem of "the phone cannot be charged", the service decision model will plan the first service node as the solution generation node; when the decision context shows that the user has submitted after-sales vouchers and is waiting for review, the service decision model will plan the first service node as the voucher review node; when the decision context shows that the user has repeatedly rejected the standardized solution, the service decision model will directly plan the first service node as the human transfer node to ensure service experience.

[0052] Regardless of which method is used to plan the first service node, if the first service node is a solution generation node, then the solution decision module 1201 corresponding to the solution generation node among the multiple task intelligence modules 120 is scheduled. The main control intelligence module 110 can send scheduling instructions to the corresponding solution decision module 1201. Correspondingly, the solution decision module 1201 can be configured to: synchronously schedule a rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule a reasoning engine to generate a second response solution and write it to the work order record module.

[0053] The following sections will provide a detailed description of the two processes: "the synchronous scheduling rule engine generates the first response scheme" and "the asynchronous scheduling inference engine generates the second response scheme".

[0054] The synchronous scheduling rule engine generates the first response scheme: The rule engine can have a built-in standardized service rule base and knowledge base for the target application scenario. For example, in the 3C digital after-sales scenario, this service rule base can include common fault troubleshooting procedures, return and exchange policies, warranty terms, shipping rules, etc. The rule engine can generate a first response solution by combining keyword matching and rule reasoning. The content of the first response solution can come from the predefined service rule base and comply with service specifications and compliance requirements, thus the risk of illusion is low. Taking the 3C digital after-sales scenario as an example, when a user asks "What should I do if my phone won't turn on?", the rule engine will generate a standardized first response solution: "Hello, you can try the following operations first: 1. Press and hold the power button for 10 seconds to force restart; 2. Connect to the original charger and charge for 30 minutes, then try to turn it on again; 3. Check for water damage or drops. If the above operations are ineffective, you can apply for after-sales inspection." The asynchronous scheduling inference engine generates a second response scheme: While the synchronous scheduling rule engine generates the first response scheme, the scheme decision module 1201 asynchronously schedules the inference engine, which has deep semantic understanding capabilities, to generate the second response scheme. The inference engine is built on large model technology and can accurately capture implicit information, personalized needs, and ambiguous requests in user questions, generating more targeted and flexible response content.

[0055] The generation process of the inference engine is controlled by the skill constraints described in the solution decision module 1201: its input is limited to the decision context and rule base content of the current dialogue, and the output can conform to predefined service rules and response formats. Optionally, the generated second response solution can also be reviewed by the built-in compliance verification module. After confirming that there is no violation, it is written to the work order record module 130 for pre-storage. The entire asynchronous generation process is completed in the background and will not affect the output speed of the first response solution.

[0056] Continuing with the example above, if a user adds in their question that "my phone won't turn on after yesterday's system upgrade," the inference engine will generate a more targeted second response: "Hello, we understand that your phone is experiencing a boot problem after a system upgrade. In addition to basic troubleshooting steps, you can also try entering recovery mode to repair the system: 1. Press and hold the power button and volume up button simultaneously until you enter recovery mode; 2. Select the 'Clear cache partition' option; 3. Restart your phone. If the repair fails, it may be due to an anomaly during the system upgrade process. We recommend that you apply for official after-sales flashing service, which is free during the warranty period." In this embodiment, after receiving the first response scheme, the main control intelligent module 110 can output it to the user without waiting for the second response scheme to be generated. If the user's approval information (such as "Okay, I'll try" or "Thank you") is received after outputting the first response scheme, the service process of this scheme generation node is completed. Afterwards, the next service node can be planned based on user feedback.

[0057] If a rejection message is received from the user after outputting the first response solution (such as "I've tried all these methods and they didn't work" or "I don't want to do this"), the main control intelligent module 110 does not need to reschedule the inference engine for real-time generation. Instead, it can directly obtain the pre-generated second response solution from the work order record module and process the user's question information according to the second response solution. This approach avoids the delay of users waiting for the large model to be generated and also solves the deficiency of rule engines in handling complex and personalized problems, thus improving the quality of responses.

[0058] In some optional embodiments, when the main control intelligent module 110 determines the work order record module corresponding to the problem information, it can achieve unified binding and management of work orders across channels, sessions, and all entry points based on the following steps: Step 111: Identify the resource object to which the problem information belongs. The resource object is the smallest resource unit in the target application scenario that can independently create a work order.

[0059] Resource objects can be flexibly configured according to different application scenarios. For example, in e-commerce after-sales scenarios, resource objects can be user orders; in financial consulting scenarios, resource objects can be user accounts or single transactions; and in 3C digital after-sales scenarios, resource objects can be after-sales orders for a single product.

[0060] If the user's question information does not explicitly include the identifier of the resource object, all valid resource objects under the currently logged-in user account can be queried, and the resource object can be accurately identified by intelligently guiding the user to select.

[0061] Step 112: Using the resource object's identifier as the query key, query the work order mapping table to see if the resource object is already bound to the work order record module. The work order mapping table refers to a pre-built global mapping relationship table of the question-and-answer information processing system, which uses the resource object's identifier as the primary key to store the one-to-one correspondence between resource objects and work order record modules.

[0062] Step 113: If the resource object is already bound, the work order record module that is already bound to it will be determined as the work order record module corresponding to the problem information; if it is not bound, a work order record module will be created and bound to the resource object in the work order mapping table, and the created work order record module will be determined as the work order record module corresponding to the problem information.

[0063] In this context, the same resource object corresponds to the same work order record module under each service entry point and in each round of conversation in the target application scenario.

[0064] In this embodiment, omnichannel customer service access is supported, including but not limited to: online access channels (mini-program customer service channel, in-app native customer service channel, web-based online customer service channel), offline access channels (offline message channel, email feedback channel), and voice access channels (intelligent telephone customer service channel, human telephone transfer channel). Regardless of which channel a user initiates an inquiry from, regardless of the round of conversation initiated by the user, or even if the same user initiates multiple inquiries about the same resource object on different channels and at different times, the question-and-answer information processing system can locate the same work order record module in the work order mapping table through the identifier of the resource object.

[0065] For example, a user first initiates a refund inquiry for 3C digital products through a mini-program (first round of conversation), then checks the refund progress through in-app customer service (second round of conversation), and then submits supplementary after-sales documents through telephone customer service. These three interactions from different channels and at different times will all be associated with the work order record module corresponding to the same refund order.

[0066] Interaction information, action records, and processing results generated by each channel are all written into the same work order record module. This ensures that the main control intelligent module and each task intelligent module can obtain the complete full-link service context of the resource object regardless of which channel it is triggered from. This effectively solves the pain points of information isolation and the need for users to repeatedly state their problems in traditional multi-channel customer service systems. It also provides underlying data support for the event-driven and state persistence architecture adopted in this embodiment.

[0067] Optionally, the work order mapping table can have a work order access interface, and all service entry points in the target application scenario are uniformly mapped to this work order access interface, forming an architecture that combines multi-entry access with single-interface access. This work order access interface encapsulates all core operations of the work order record module, such as creation, querying, writing, and updating, serving as the standard entry point for providing services externally and decoupling the customer service access layer from the work order data layer. Based on this, when creating and binding a work order record module for a resource object in the work order mapping table, a binding relationship can be established between the resource object's identifier and the work order access interface. This binding relationship can be stored in the work order mapping table as a global routing rule. When a user initiates a request for the resource object from any service entry point, the request first reaches the unified work order access interface. The work order access interface can automatically locate the corresponding work order record module based on the resource object identifier carried in the request, through the binding relationship in the work order mapping table, and complete the information reading or writing operation. Afterwards, the resource object's identifier can be used as a unique primary key to create a blank work order entry in the work order mapping table, serving as the resource object's work order record module. Blank work order entries can have predefined standardized field structures, including but not limited to resource object basic information fields, session history fields, action record chain fields, service status fields, and response scheme storage fields.

[0068] In this way, the unified binding effect of cross-channel and cross-session work orders can be strengthened. It not only breaks down the data barriers between different customer service channels, but also significantly improves the scalability of the system. When a new service entry is needed (such as a new customer service channel for short video platforms or smart speakers), the new entry only needs to be mapped to the globally unified work order access interface. There is no need to modify the core logic of the work order system or redesign the data storage structure, which greatly reduces the cost and risk of system iteration.

[0069] In some optional embodiments, when the main control intelligent module 110 constructs the decision context of the current session based on the user intent and the historical service information recorded in the work order record module, it can be implemented based on the following steps, thereby quantitatively evaluating the relevance of historical service information through multi-dimensional features, effectively filtering irrelevant historical information, solving the decision bias problem caused by information redundancy and too many interference items in traditional context construction, and significantly improving the accuracy and effectiveness of the decision context.

[0070] Step 121: Obtain the target service entry point and initiation time corresponding to this round of conversation. The target service entry point is any one of the multiple service entry points in the target application scenario, including but not limited to the mini-program customer service channel, the in-app native customer service channel, the web-based online customer service channel, the offline message channel, the email feedback channel, the intelligent telephone customer service channel, etc. mentioned above.

[0071] It should be noted that different service entry points have different interactive characteristics and information densities. For example, telephone customer service channels are mainly based on voice interaction, and the information is relatively conversational and fragmented; text-based customer service channels are mainly based on text interaction, and the information is more structured and complete. At the same time, the session initiation time directly reflects the timeliness of historical service information. Therefore, incorporating these two dimensions of information into the context construction process can significantly improve the accuracy of context matching.

[0072] Step 122: Construct a model by inputting the target service entry point, initiation time, and user intent context corresponding to this round of conversation.

[0073] The context building model is one of the multi-stage joint training AI models mentioned above. It can learn the intrinsic relationship between service entry characteristics, time decay rules and intent transfer logic through pre-training, and can accurately quantify the relevance of different historical service information to the current session.

[0074] Step 123: Based on the target service entry point, initiation time, and user intent corresponding to this round of conversation, calculate the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information recorded in the work order record module.

[0075] The channel consistency feature can be used to measure whether the service entry point corresponding to historical service status information is the same as the service entry point corresponding to the current session. For example, considering that continuous inquiries initiated by users from the same service entry point usually have a stronger contextual relevance, if historical service information and the current session come from the same service entry point, the feature value is 1; if they come from different service entry points, a feature value between 0 and 1 is assigned based on the degree of information interoperability between different channels. For example, the information interoperability between the mini-program customer service channel and the in-app customer service channel is relatively high, and the feature value can be set to 0.8; while the information interoperability between the telephone customer service channel and the email feedback channel is relatively low, and the feature value can be set to 0.5.

[0076] The temporal proximity feature measures the time interval or step decay between the session initiation time and the current session for historical service status information. Shorter time intervals result in higher feature values; longer time intervals lead to exponential decay. For example, historical sessions within 24 hours have a feature value of 0.9, those within 7 days have 0.5, and those over 30 days have 0.1. For multiple historical records within the same session, step decay calculations can be applied, meaning that historical records closer to the current session have higher feature values. This feature effectively filters outdated historical information, ensuring the timeliness of the decision-making context.

[0077] Among them, the intent coherence feature can be used to measure whether the historical intent corresponding to historical service status information is the same as the user's intent in the current session, or whether there is an intent connection or jump relationship. In other words, the intent coherence feature can quantify the degree of correlation between the historical intent corresponding to historical service information and the user's intent in the current session, including three relationships: identical intent, intent connection, and intent jump. For example, if the historical intent and the current intent are exactly the same, the feature value is 1; if there is a clear logical connection relationship (such as asking "return process" first and then asking "shipping cost sharing"), a feature value between 0.6 and 0.9 is assigned based on the built-in intent transition probability matrix; if there is a reasonable intent jump relationship (such as asking "repair policy" first and then asking "exchange conditions"), a feature value between 0.3 and 0.6 is assigned; if the intents are completely unrelated, the feature value is 0. The intent coherence feature can accurately capture the evolution logic of user demands, ensuring the coherence of the context.

[0078] Step 124: Based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information, select the target historical service information that has a contextual relationship with the current session.

[0079] This allows us to obtain the dynamic feature priority corresponding to the current session. The dynamic feature priority is used to indicate the service entry point, initiation time, and user intent. Figure 3 The relative importance of each dimension in this round of conversation. It's important to note that the importance of the three dimensions varies significantly across different conversations. The dynamic feature priority is not fixed in advance but is calculated and generated in real-time based on the service entry point attributes and user intent type for this round of conversation. Specifically, a basic priority can be pre-set for different service entry points. The priority is determined by the user's urgency, interaction characteristics, and service importance of that service entry point. For example, the intelligent telephone customer service channel has the highest basic priority (users typically choose telephone consultation when the problem is urgent), followed by the in-app native customer service channel, then the mini-program customer service channel and the web-based customer service channel, and finally the offline message channel and email feedback channel have the lowest priority. A conversation initiated through a higher-priority service entry point indicates a higher level of user attention and urgency regarding the issue, which will automatically increase the priority weight of the service entry point dimension and prioritize matching historical service information from the same entry point.

[0080] This system allows for pre-setting basic priorities for different types of user intents. The priority level is determined by the urgency of the service, the complexity of the problem, and the intensity of the user's demand. For example, in e-commerce after-sales scenarios, complaint intents, refund intents, and exchange intents have the highest basic priority; repair intents and reissue of vouchers intents have the next highest priority; and progress inquiry intents and consultation intents have the lowest priority. When a high-priority intent is identified in the current session, its priority weight can be automatically increased, prioritizing the matching of historical service information logically related to the current intent.

[0081] Based on the aforementioned dynamic feature priorities, corresponding dynamic weight coefficients are assigned to the channel consistency feature, temporal proximity feature, and intent coherence feature of each historical service information entry. The weight coefficients are positively correlated with the feature priority; the higher the priority of a dimension, the larger the assigned weight coefficient. The sum of the weight coefficients for the three dimensions is 1. For example, when a user initiates a complaint through the highest-priority intelligent telephone customer service channel, the dynamic feature priority is: intent dimension > time dimension > service entry dimension. The corresponding weight coefficients can be set as: α=0.2 (channel consistency), β=0.3 (temporal proximity), γ=0.5 (intent coherence).

[0082] When a user initiates a progress query through the regular priority mini-program customer service channel, the dynamic feature priority is: service entry dimension > time dimension > intent dimension, and the corresponding weight coefficients can be set as: α=0.5 (channel consistency), β=0.3 (temporal proximity), γ=0.2 (intent coherence).

[0083] Based on the channel consistency, temporal proximity, and intent coherence characteristics of each historical service message, and their respective dynamic weighting coefficients, the contextual relevance of each historical service message to the current session is calculated. The formula for calculating the contextual relevance is: Contextual relevance = α × Channel consistency feature + β × Temporal proximity feature + γ × Intent coherence feature Wherein, α, β, and γ are weight coefficients determined based on the priority of dynamic features in this round of conversation.

[0084] Based on the calculated context relevance, a preset number of target historical service information entries are selected from each historical service information entry. Specifically, all historical service information entries can be sorted in descending order based on context relevance, and the preset number of historical service information entries with the highest context relevance are selected as the target historical service information entries. The preset number can be flexibly configured according to the service characteristics of different application scenarios. For example, in high-priority complaint and refund scenarios, the preset number can be set to 10 entries to ensure complete historical service context is obtained; in low-priority ordinary consultation scenarios, the preset number can be set to 3-5 entries to ensure sufficient information while avoiding information redundancy and improving system processing efficiency.

[0085] In this way, based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information, after selecting the target historical service information that has a contextual relationship with the current session, the following steps 125 can be performed.

[0086] Step 125: Integrate the target historical service information and user intent in a structured manner to obtain the decision context for this round of the session. The decision context can be stored in a key-value pair structure, which may include: user basic information field, current intent and confidence level field, target service entry field, session initiation time field, target historical service information list field, and current service status field, etc.

[0087] The structured and integrated decision context will be written into the work order record module, serving as the core decision-making basis for this session and also as historical service information for subsequent sessions. This structured approach significantly improves the processing efficiency and accuracy of the service decision model, avoiding semantic ambiguity issues caused by unstructured text.

[0088] In this way, the multi-feature weighted approach can accurately filter relevant historical service information, ensuring the integrity and coherence of the decision-making context while effectively filtering out irrelevant information, significantly improving the accuracy and response speed of service decisions, and further optimizing the user experience.

[0089] In some optional embodiments, when the master control intelligent module 110 plans the first service node based on the decision context, it may do so by following these steps: Step 131: Perform dimensional vectorization processing on the decision context to obtain vector matrices in the service entry dimension, initiation time dimension, and intent dimension; the vector matrix in any dimension includes the current vector of the current session in that dimension and the historical vector of historical service information in that dimension.

[0090] When vectorizing service entry points, the interactive characteristics of different service entry points, such as information density, interaction methods, and user preferences, can be encoded into vectors of fixed length. For example, the mini-program customer service channel can be encoded as a vector [0.9, 0.7, 0.2] (representing high information density, text-based interaction, and low urgency, respectively), and the intelligent telephone customer service channel can be encoded as a vector [0.5, 0.3, 0.9] (representing medium information density, voice-based interaction, and high urgency, respectively).

[0091] When vectorizing the initiation time dimension, the initiation time of the current session, the initiation time of each historical service message, and the time interval between them can be encoded into a time feature vector.

[0092] When vectorizing the intent dimension, a pre-trained semantic embedding model can be used to convert the user intent of the current session and the historical intent corresponding to each historical service information into a dense vector with semantic expression capabilities, so that semantically similar intents are closer in the vector space.

[0093] Dimensional vectorization can convert contextual information of different types and modalities into a numerical form that can be processed by cross-attention networks, while preserving the independent semantic features of each dimension, laying the foundation for subsequent cross-dimensional feature fusion.

[0094] Step 132: Concatenate the vector matrices of the service entry dimension, initiation time dimension, and intent dimension into a joint vector matrix.

[0095] Specifically, the vector matrices of the service entry dimension, initiation time dimension, and intent dimension can be concatenated in any form according to a preset order, such as vertical stacking, horizontal concatenation, or dimension-weighted fusion concatenation, to generate a joint vector matrix containing all contextual information. This joint vector matrix can completely preserve the original feature information of each dimension, while establishing the positional correspondence between features of different dimensions, enabling the cross-attention network to learn the interaction relationships between features of different dimensions.

[0096] Step 133: Input the joint vector matrix into the cross-attention network, and use different attention heads to extract service entry consistency features, temporal proximity features and intent semantic similarity features respectively. After fusion, the current session state representation is obtained.

[0097] The cross-attention network can be configured with the following attention heads: Service entry consistency attention head: used to calculate the attention weight between the current service entry vector and the historical service entry vector, extract service entry consistency features, and measure the continuity of user behavior across different channels.

[0098] Temporal proximity attention head: used to calculate the attention weights between the current time vector and historical time vectors, extract temporal proximity features, and distinguish the importance of recent historical information from distant historical information.

[0099] Intent semantic similarity attention head: used to calculate the attention weight between the current intent vector and the historical intent vector, extract the semantic similarity features of intent, and accurately capture the evolution logic and jump relationship of user intent.

[0100] Cross-attention networks can learn the weight allocation between features of different dimensions, enabling them to prioritize information that is more important to the current service node's planning and suppress interference from irrelevant information. For example, when a user repeatedly rejects a standardized solution within a short period of time, the temporal proximity attention head and the intent semantic similarity attention head will automatically assign higher weights to this historical information in order to more accurately identify the user's needs.

[0101] Step 134: Input the current session state representation into the node classifier or reinforcement learning policy network, and output the first service node that matches the current session state representation.

[0102] Among them, the appropriate decision network can be automatically selected from the node classifier and reinforcement learning policy network according to the characteristics of the target application scenario and the complexity of the service process. It can also support the hybrid deployment and dynamic switching of node classifier and reinforcement learning policy.

[0103] The node classifier can be built on a supervised learning framework, learning a static mapping from session state representation to service nodes. Specifically, the current session state representation output by the cross-attention network can be used as input features. This is then processed through a multi-layer feedforward neural network to perform forward propagation calculations, outputting the probability distribution of all predefined service nodes. Finally, the service node with the highest probability is selected as the first service node. By employing two differentiated decision networks, the system balances response speed for standardized scenarios with decision-making flexibility for complex scenarios.

[0104] Based on steps 131-134 above, the deep fusion of multi-dimensional contextual information and intelligent feature extraction are achieved based on the cross-attention network. It has a stronger ability to understand complex contexts and capture intent evolution, and can provide users with more accurate and personalized service path planning.

[0105] In some optional embodiments, when the main control intelligent module 110 schedules the scheme decision module 1201 corresponding to the scheme generation node among multiple task intelligent modules 120, it can be implemented based on the following steps 141-143: Step 141: Read the target skill requirements of the solution generation node, and select the solution decision module 1201 from the multiple task intelligence modules 120 based on the target skill requirements and the predefined skill constraint descriptions of the multiple task intelligence modules 120.

[0106] The target skill requirements are a set of skill requirements imposed by the solution generation node on the task intelligence module 120, which may include four core dimensions: task type requirements, access permissions for callable tools, output format requirements, and compliance level requirements. The main control intelligence module 110 can be configured with a global skill registry to store the core features of the predefined skill constraint descriptions of each task intelligence module.

[0107] During the skill matching process, a combination of exact matching and priority matching can be used. In the exact matching stage, the target skill requirements of the solution generation node can be compared field by field with the skill constraint descriptions of each task intelligent module 120 in the global skill registry to filter out candidate task intelligent modules that meet the target skill requirements.

[0108] If there are multiple candidate task intelligent modules that meet the requirements, they will be comprehensively scored and ranked according to quantitative indicators such as historical execution success rate, average response speed, and historical user satisfaction. The task intelligent module with the highest score will be selected as the solution decision module for this scheduling.

[0109] Step 142: Based on the decision context and the skill constraint description of the solution decision module, generate structured prompts for the solution decision module.

[0110] Structured prompts are standardized carriers for the master control intelligence module 110 to transmit instructions to the task intelligence module 120. Their content follows the skill constraint descriptions of the corresponding task intelligence module 120 and will not contain instructions exceeding skill boundaries. Structured prompts can adopt a fixed field-based format, and for example, may include the following fields: Context Information Field: Contains core decision context such as user intent for this session, target historical service information, and current service status. Task Instruction Field: Explicitly specifies the specific task to be performed (e.g., generating an after-sales response plan). Tool Invocation Rules Field: Explicitly limits the types and methods of tools that can be invoked for this task (e.g., a dual-engine scheduling mechanism of synchronous rule engine + asynchronous inference engine must be used). Output Format Field: Explicitly specifies the output structure, language style, and word count limits of the response plan. Compliance Constraint Field: Explicitly lists the service rules and prohibitions that must be followed for this task.

[0111] In this embodiment, the structured prompts can be used to transform skill constraint descriptions from static configurations into dynamic execution instructions, further defining the execution scope of the task intelligence module 120. Even if the basic model of the task intelligence module 120 has generation capabilities, it can only perform operations within the scope defined by the structured prompts, further preventing hallucination output at the instruction level.

[0112] It should be noted that in this embodiment, the skill constraint description is transmitted to the solution decision module 1201 through the skill constraint field of the structured prompt words, specifically including the following two implementable methods: Method 1: Skill Constraint Identification Transmission Method Each task intelligence module 120 has a pre-configured unique identifier for its skill constraint description. This unique identifier can be a skill name, skill number, or a unique storage address of the skill constraint description in the global skill registry. When generating structured prompts, this unique identifier is added to the skill constraint field of the prompt. After receiving the structured prompt, the solution decision module 1201 can read the corresponding complete skill constraint description from the global skill registry based on this unique identifier and execute the task according to the read skill constraint description. The advantage of this method is that the skill constraint description can be centrally managed and dynamically updated without modifying the task intelligence module itself.

[0113] Method 2: Skill Constraint Activation Command Transmission Method Skill constraint descriptions are hard-coded into the corresponding solution decision module 1201, with each skill constraint description corresponding to a unique activation command or activation code. When generating structured prompts, the corresponding activation command or activation code is added to the skill constraint field of the prompt. Upon receiving the structured prompt, the solution decision module 1201 can automatically activate the corresponding internal skill constraint based on the activation command or activation code and execute the task according to that skill constraint. The advantages of this method are higher execution efficiency, no tampering of skill constraints during transmission, and stronger security.

[0114] It should be noted that the skill constraint description of the solution decision module 1201, transmitted in two ways as described above, includes at least two core components: an available toolset and concurrent execution logic. This skill constraint description predefines the execution boundaries and operating rules of the solution decision module, ensuring that it can only call tools and execute tasks within a limited scope. Optionally, the skill constraint description may also include: a skill metadata section, a trigger condition constraint section, and an input / output contract section. The skill metadata section contains the skill name, version number, and compatibility declaration, used to identify the solution decision module 1201 and for discovery and indexing in the skill registry. The trigger condition constraint section contains a trigger intent description and disabled scenario information. The trigger intent description defines under what service intent the solution decision module 1201 should be called, and the disabled scenario information defines the reverse intent and boundary conditions under which the skill should not be called, thus constraining the calling decisions of the main control intelligent module 110. The input / output contract section can adopt a predefined data structure specification to declare the set of input parameters required to call the scheme decision module 1201, as well as their parameter types, mandatory and value ranges, and the output structure format of the first response scheme returned synchronously and the second response scheme written in the background, so that the input / output behavior of the scheme decision module 1201 can be deterministically verified.

[0115] The available toolset is a collection of tools authorized to be invoked by the solution decision module 1201, including at least: a rule engine, an inference engine, and a work order writing tool. Each tool has clearly defined functional boundaries and invocation permissions.

[0116] Step 143: Based on the structured prompt word scheduling scheme decision module 1201, the scheme decision module 1201 is triggered to generate a first response scheme and return it based on its skill constraint description, and the asynchronous scheduling inference engine generates a second response scheme and writes it into the work order record module 130.

[0117] The main control intelligent module 110 can package the generated structured prompts and decision context into standardized scheduling instructions and send them to the selected solution decision module 1201. Upon receiving the scheduling instructions, the solution decision module 1201 first loads the corresponding complete skill constraint description based on the unique identifier or activation instruction in the skill constraint field of the structured prompts, and verifies whether the scheduling instructions conform to the skill constraint description. If the verification is successful, the dual-engine scheduling logic can be executed. When the rule engine generates the first response solution, the solution decision module 1201 can return it to the main control intelligent module 110. When the inference engine generates the second response solution, the solution decision module writes it to the work order record module 130 and sends an execution completion notification to the main control intelligent module 110.

[0118] If the solution decision module 1201 verifies that the scheduling instruction exceeds its own skill boundary, it can return a skill mismatch exception to the main control intelligent module 110. The main control intelligent module 110 will then re-perform skill matching and schedule other qualified task intelligent modules 120 to ensure that the task can be executed correctly.

[0119] In some optional implementations, when the main control intelligent module 110 processes the problem information according to the second response scheme, if it receives confirmation information from the user after outputting the second response scheme, such as "Okay, let's do it this way" or "Agree to this scheme," it schedules the execution engine corresponding to the second response scheme to execute the corresponding response action to process the problem information. Optionally, it may also execute the response action corresponding to the second response scheme directly after outputting the second response scheme without consulting the user, i.e., without receiving confirmation information from the user. For example, it may perform a refund or return.

[0120] Specifically, based on the service type of the second response scheme, the corresponding execution engine can be scheduled to perform the appropriate response action to complete the response processing of the problem information. For example: a refund execution engine for the refund scheme, a remote assistance execution engine for the troubleshooting scheme, a service center appointment scheduling execution engine for the after-sales repair scheme, and a voucher upload guidance execution engine for the voucher re-transmission scheme, etc. The execution engine can have the ability to execute preset standardized actions.

[0121] Optionally, the work order recording module 130 can also be used to record the end-to-end information generated in this round of session. The end-to-end information may include at least one of the following service information: problem information, user intent, first response plan, process of negotiating the first response plan with the user, user rejection information of the first response plan, process of negotiating the second response plan with the user, and user confirmation information of the second response plan.

[0122] Specifically, the system records the user's original input text, speech-to-text content, and attachments, preserving the original expression of the user's request. It also records the user intent classification and confidence level output by the intent recognition model. The system records the standardized first response solution generated by the rules engine and its generation time. The process of negotiating the first response solution with the user records the user's interaction content, interaction time, and interaction channel. Information on the user's rejection of the first response solution records the specific reasons for rejection and the time of rejection. The process of negotiating the second response solution with the user records the user's interaction content, interaction time, and interaction channel. Finally, it records the time, method, and additional requirements for the user's confirmation of the second response solution.

[0123] The aforementioned service information can be stored in a structured format in the work order record module 130, forming a complete service end-to-end action record chain. In this way, complete contextual data can be provided for subsequent cross-session and cross-channel service connections, ensuring users do not need to repeat their questions; original evidence can be provided for compliance auditing and quality traceability throughout the service process, meeting the regulatory requirements of various industries; and high-quality labeled training data can be provided for the continuous optimization of node classifiers and reinforcement learning strategy networks, continuously improving the system's decision-making accuracy.

[0124] In some optional embodiments, if a user's confirmation is received after the first response scheme is output, the problem information is responded to according to the first response scheme. Specifically, based on the service type of the first response scheme, the execution engine corresponding to that first response scheme can be scheduled to execute the appropriate response action, thus completing the response processing for the problem information.

[0125] If a rejection message is received from the user after outputting the second response plan, the intelligent outbound calling module among multiple task intelligent modules is scheduled to conduct telephone negotiations with the user through the intelligent outbound calling module. Among them, the main control intelligent module 110 can plan intelligent outbound calling service nodes and schedule the pre-configured intelligent outbound calling modules among multiple task intelligent modules 120 to conduct telephone negotiations with the user through the intelligent outbound calling module.

[0126] As a task-oriented intelligent module with clearly defined skill constraints, the intelligent outbound calling module is limited to executing predefined after-sales negotiation outbound calling processes. The outbound calling content can be generated based on the decision context stored in the work order record module 130. Before the outbound call, the intelligent outbound calling module can retrieve the historical service information of the order from the work order record module 30, including the user's original request, the two provided response plans, and the user's reasons for refusal, and generate a targeted outbound calling script. During the outbound call, the voice interaction content is transcribed into text in real time and appended to the work order record module 130. After the outbound call, the intelligent outbound calling module can synchronize the negotiation result (such as the user agreeing to the new plan, the user still refusing, or the user requesting to be transferred to a human agent) to the main control intelligent module 110. The main control intelligent module 110 plans subsequent service nodes according to the negotiation result: if the negotiation is successful, the corresponding execution engine is scheduled to complete the processing; if the negotiation fails, the user is directly transferred to a human customer service representative for processing.

[0127] Optionally, after receiving confirmation from the user regarding the first or second response scheme, a second service node can be planned based on the first or second response scheme. If the second service node includes a credential verification node, then the credential verification module corresponding to the credential verification node among multiple task intelligent modules can be scheduled.

[0128] As a task-oriented intelligent module with clearly defined skill boundaries, the voucher review module's behavior is governed by predefined skill constraints and it lacks the ability to make autonomous decisions or execute beyond its authority. These skill constraints include: available toolset constraints (allowing the use of tools such as evidence card distribution, OCR recognition, voucher compliance verification, and work order writing, but not unauthorized tools); execution process constraints (following the process of "issuing evidence cards → receiving voucher materials → intelligent review → outputting review results"); output format constraints (results are output in a structured format, including fields such as pass / fail indicators, reasons for failure, and required voucher types); and compliance constraints (review standards adhere to a predefined service rule base, preventing review decisions beyond the rule scope).

[0129] The credential review module is configured to: based on its skill constraint description, sequentially execute the following steps: issue evidence cards, receive credential materials submitted by users through the evidence cards, perform intelligent review of the credential materials, and output the review results. If the review result is satisfactory, the issue information is addressed according to either the first or second response plan. If the review result is unsatisfactory, the response is routed according to the type of response plan currently confirmed by the user: if the received information is confirmation of the first response plan (i.e., the user previously rejected the first plan), the main control intelligent module does not need to reschedule the inference engine for real-time generation, but directly obtains the second response plan from the work order record module and performs response processing according to the second response plan. This mechanism reuses the asynchronous pre-generated results of the dual engines, reducing latency; if the received information is rejection of the second response plan (i.e., both the first and second plans are rejected), the main control intelligent module schedules the comprehensive adjudication module to generate a third response plan and performs response processing according to the third response plan.

[0130] The comprehensive adjudication module is the final adjudication module of the question-and-answer information processing system. It serves as the unified intelligent processing entry point when all preceding intelligent tools (including the solution decision module, credential verification module, and intelligent outbound call module) fail. The comprehensive adjudication module possesses more comprehensive service processing capabilities and more flexible rule adaptation capabilities than other task intelligent modules 120, enabling it to handle complex, special, and edge scenarios that preceding intelligent modules cannot cover. Furthermore, the scheduling process of the comprehensive adjudication module is consistent with other task intelligent modules 120, adhering to standardized skill matching and constraint control mechanisms. The comprehensive adjudication module also has skill constraint descriptions and does not have the ability to execute beyond its authority. After the third response solution is generated, the main control intelligent module 110 can output it to the user and continue to plan subsequent service nodes based on the user's feedback; if the solution generated by the comprehensive adjudication module still cannot solve the user's problem, the main control intelligent module 110 can directly transfer the user to human customer service for backup processing.

[0131] This embodiment constructs a progressive service process of "first response plan → second response plan → comprehensive adjudication module → human customer service", which realizes hierarchical and intelligent processing of user problems. While maximizing the processing capacity of the intelligent system and reducing labor costs, it ensures that user problems can be properly resolved.

[0132] In some optional embodiments, if the second service node includes a merchant negotiation node, the main control intelligent module 110 can adopt a standardized scheduling process consistent with the solution decision module 1201 to schedule the solution negotiation module corresponding to the merchant negotiation node among multiple task intelligent modules 120. The solution negotiation module can be implemented as a PB (Platform Business) negotiation tool, a dedicated intelligent tool for automated after-sales negotiation between the platform and its merchants, responsible for driving standardized information interaction and negotiation process execution between the platform and the merchants.

[0133] The solution negotiation module, as a task intelligence module with clearly defined skill boundaries, is governed by predefined skill constraint descriptions. These skill constraint descriptions include at least: Available toolset constraints: Only the merchant information query tool, negotiation strategy model calling tool, work order writing tool, and status notification tool may be used. Calling any other unauthorized tools is prohibited. Execution process constraints: The fixed process of "writing asynchronous negotiation initiation action → driving negotiation strategy model → obtaining negotiation execution result → writing to work order record module" must be strictly followed. No execution steps may be skipped or changed. Output format constraints: All content written to work orders must be in a standardized structured format, including fixed fields such as negotiation initiation time, negotiation participants, negotiation progress, negotiation result, and next action. Compliance constraints: Negotiation strategies must comply with predefined platform merchant negotiation rules and after-sales policies. No promises or concessions exceeding the scope of the rules may be made.

[0134] The solution negotiation module can be configured to: based on its skill constraint description, first write an asynchronous negotiation start action to the work order record module 130 and synchronize it to the main control intelligent module 110, and then drive the built-in negotiation strategy model in the background to execute the complete merchant negotiation process; after the negotiation process is completed, the solution negotiation module writes the execution result of the merchant negotiation process (including negotiation success / failure indicators, the specific solution reached through negotiation, merchant feedback, etc.) into the work order record module 130 in a standardized format, and sends a negotiation completion notification to the main control intelligent module 110.

[0135] After receiving the negotiation completion notification, the main control intelligent module 110 automatically reads the negotiation result from the work order record module 130 and plans subsequent service nodes based on the negotiation result: if the negotiation is successful, the corresponding execution engine is scheduled to execute the negotiated solution; if the negotiation fails, a comprehensive adjudication node or a human customer service node is automatically planned for subsequent processing. The entire merchant negotiation process is executed asynchronously in the background, without affecting the real-time interaction between the user and the system, and the user can continue other operations without waiting for the negotiation to be completed.

[0136] This embodiment achieves automation, asynchronization, and controllability of the merchant negotiation process by reusing the unified standardized scheduling and constraint architecture of the entire system, thereby improving the processing efficiency of cross-entity negotiations and reducing the cost of manual intervention.

[0137] In addition to the above-mentioned question-and-answer information processing system, this application embodiment also provides the following question-and-answer information processing method, applied to the main control intelligent module, which is associated with multiple task intelligent modules in the target application scenario. For example... Figure 2 As shown, the method may include: Step 11: Receive the user's problem information in this session and determine the corresponding work order record module.

[0138] Step 12: Identify user intent based on problem information, and construct the decision context for this round of conversation based on user intent and historical service information recorded in the work order record module.

[0139] Step 13: Plan the first service node based on the decision context. If the first service node is a solution generation node, schedule the solution decision module corresponding to the solution generation node from among the multiple task intelligence modules.

[0140] Step 14: The solution decision module is configured as follows: based on its skill constraint description, the synchronous scheduling rule engine generates a first response solution and returns it, and the asynchronous scheduling inference engine generates a second response solution and writes it to the work order record module.

[0141] Step 15: If a user's rejection message is received after the first response solution is output, then the second response solution is obtained from the work order record module, and the problem information is responded to according to the second response solution.

[0142] In this method, the master control intelligent module associates multiple task intelligent modules in the target application scenario, receives problem information and determines the corresponding work order record module, identifies user intent and constructs a decision context, and plans the first service node based on the decision context. If the first service node is a solution generation node, the corresponding solution decision module is scheduled, and based on its skill constraint description, the synchronous scheduling rule engine generates a first response solution and returns it, while the asynchronous scheduling inference engine generates a second response solution and writes it to the work order record module. If a user rejection message is received after outputting the first response solution, the second response solution is retrieved from the work order record module for response. A balance between rapid response and accurate calculation is achieved through a dual-engine asynchronous collaboration mechanism; the cross-channel state persistence mechanism based on the work order record module avoids context memory loss and service solution inconsistency.

[0143] Optionally, determining the work order record module corresponding to the problem information includes: identifying the resource object to which the problem information belongs, where the resource object is the smallest resource unit in the target application scenario that can independently create a work order; using the identifier of the resource object as the query key, querying the work order mapping table to see if the resource object is already bound to a work order record module; if it is already bound, then determining the work order record module bound to the resource object as the work order record module corresponding to the problem information; if it is not bound, then creating and binding a work order record module for the resource object in the work order mapping table, and determining the created work order record module as the work order record module corresponding to the problem information; wherein, the same resource object corresponds to the same work order record module under each service entry point and in each round of sessions in the target application scenario.

[0144] Optionally, the work order mapping table has a work order access interface, and each service entry point in the target application scenario is uniformly mapped to the work order access interface; wherein, creating and binding a work order record module for a resource object in the work order mapping table includes: establishing a binding relationship between the identifier of the resource object and the work order access interface, and using the identifier of the resource object as a unique primary key to create a blank work order entry in the work order mapping table as the work order record module of the resource object; wherein, the binding relationship is used to record the full-link service information corresponding to the problem information from any service entry point and belonging to the resource object into the work order record module of the resource object.

[0145] Optionally, based on user intent and historical service information recorded in the work order record module, a decision context for the current session is constructed, including: obtaining the target service entry point and initiation time corresponding to the current session, wherein the target service entry point is any one of multiple service entry points in the target application scenario; inputting the target service entry point, initiation time, and user intent corresponding to the current session into the context to construct a model: based on the target service entry point, initiation time, and user intent corresponding to the current session, calculating the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information recorded in the work order record module; selecting target historical service information that has a contextual association with the current session based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information; and structurally integrating the target historical service information and user intent to obtain the decision context for the current session.

[0146] Optionally, based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information, target historical service information with contextual relevance to the current session is selected, including: obtaining the dynamic feature priority corresponding to the current session, whereby the dynamic feature priority indicates the relative importance of the service entry point, initiation time, and user intent in the current session; assigning corresponding weight coefficients to the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information based on the dynamic feature priority; calculating the contextual relevance of each historical service information to the current session based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information and their respective weight coefficients; and selecting a preset number of target historical service information from each historical service information based on the contextual relevance.

[0147] Optionally, planning the first service node based on the decision context includes: performing dimensional vectorization processing on the decision context to obtain vector matrices in the service entry dimension, initiation time dimension, and intent dimension; the vector matrix in any dimension includes the current vector of the current session in that dimension and the historical vector of historical service information in that dimension; concatenating the vector matrices in the service entry dimension, initiation time dimension, and intent dimension into a joint vector matrix; inputting the joint vector matrix into a cross-attention network, using different attention heads to extract service entry consistency features, temporal proximity features, and intent semantic similarity features respectively, and fusing them to obtain the current session state representation; inputting the current session state representation into a node classifier or reinforcement learning policy network, and outputting the first service node that matches the current session state representation.

[0148] Optionally, scheduling a solution decision module corresponding to a solution generation node among multiple task intelligence modules includes: reading the target skill requirements of the solution generation node; selecting a solution decision module from multiple task intelligence modules based on the target skill requirements and predefined skill constraint descriptions of multiple task intelligence modules; generating structured prompts for the solution decision module based on the decision context and the skill constraint descriptions of the solution decision module; scheduling the solution decision module according to the structured prompts to trigger the solution decision module to synchronously schedule the rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule the inference engine to generate a second response solution and write it to the work order record module.

[0149] Optionally, the skill constraint description of the solution decision module includes at least: the available toolset and concurrent execution logic of the solution decision module; the available toolset includes at least: a rule engine, an inference engine, and a work order writing tool; the concurrent execution logic is used to instruct the solution decision module to perform the following operations in parallel: synchronously schedule the rule engine to perform keyword matching with a predefined rule base based on the decision context in the structured prompts to generate a first response solution and return it; asynchronously schedule the inference engine to call the solution inference model to perform semantic understanding and comprehensive inference based on the decision context in the structured prompts to generate a second response solution, and call the work order writing tool to write the second response solution into the work order record module.

[0150] Optionally, according to the second response scheme, the problem information is processed to respond, including: if the user's confirmation information is received after the second response scheme is output, the execution engine corresponding to the second response scheme is scheduled to perform the corresponding response action to process the problem information.

[0151] Optionally, the work order recording module is also used to record at least one of the following service information generated in this round of conversation: problem information, user intent, first response plan, process of negotiating the first response plan with the user, user rejection information of the first response plan, process of negotiating the second response plan with the user, and user confirmation information of the second response plan.

[0152] Optionally, it also includes: if a user's confirmation is received after outputting the first response scheme, then the problem information is responded to according to the first response scheme; or if a user's rejection is received after outputting the second response scheme, then the intelligent outbound call module among multiple task intelligent modules is scheduled to conduct telephone negotiation with the user through the intelligent outbound call module.

[0153] Optionally, after receiving confirmation from the user regarding the first or second response scheme, the method further includes: planning a second service node according to the first or second response scheme; if the second service node includes a credential verification node, scheduling the credential verification module corresponding to the credential verification node from among multiple task intelligent modules; the credential verification module is configured to: based on its skill constraint description, sequentially execute the issuance of evidence cards, receive the credential materials submitted by the user through the evidence cards, perform intelligent verification of the credential materials, and output the verification result; if the verification result is passed, then perform the operation of responding to the problem information according to the first or second response scheme.

[0154] Optionally, it also includes: if the review result is unsuccessful and the received information is the user's confirmation of the first response plan, then the second response plan is obtained from the work order record module, and the problem information is responded to according to the second response plan; if the review result is unsuccessful and the received information is the user's confirmation of the second response plan, then the comprehensive adjudication module in the multiple task intelligent modules is scheduled to generate a third response plan, and the problem information is responded to according to the third response plan.

[0155] Optionally, it also includes: if the second service node includes a merchant negotiation node, then scheduling the solution negotiation module corresponding to the merchant negotiation node among multiple task intelligence modules; the solution negotiation module is configured to: write an asynchronous negotiation start action to the work order record module based on its skill constraint description, drive the negotiation strategy model to execute the merchant negotiation process, and write the execution result of the merchant negotiation process to the work order record module.

[0156] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0157] The following section, in conjunction with the accompanying drawings, will provide a detailed explanation of the application of the aforementioned question-and-answer information processing system and method in real-world scenarios.

[0158] like Figure 3 As shown, in an exemplary embodiment, the question-and-answer information processing system can be divided into a user interaction layer, a main control intelligence layer, a task intelligence layer, and a basic support layer from top to bottom in a real-world scenario, achieving complete control from user request access, intelligent decision-making, task execution to end-to-end recording.

[0159] User interaction layer: Includes online instant messaging consultation portal and refund page customer service intervention portal, used to receive user after-sales consultation, evidence materials, negotiation feedback, and complete user-side request input and result output.

[0160] The master control intelligence layer includes a master control intelligence module, which has a built-in cross-attention network and decision network; the decision network includes a node classifier, a reinforcement learning policy network, etc.

[0161] Task Intelligence Layer: Includes solution decision-making module, voucher verification module, solution negotiation module, intelligent outbound calling module, and comprehensive adjudication module.

[0162] Basic support layer: includes work order record module, global skill registry, etc.

[0163] For example, a user buys a dress priced at 269 yuan on an e-commerce platform, and after signing for it, finds that the dress has a large area of ​​damage. The user initiates an after-sales dispute through the customer service portal on the refund page, requesting a full refund of 269 yuan plus 15 yuan for actual shipping costs. The merchant refuses to pay compensation, and this system initiates a tiered intelligent processing procedure.

[0164] Users can submit after-sales requests through the customer service portal on the refund page, uploading photos of the damaged dress and filling in order information and compensation requests. The user interaction layer collects data such as user request text, evidence images, and order numbers, and transmits them to the main control intelligent layer. The main control intelligent module calls a cross-attention network to integrate multimodal data such as user requests, order information, evidence images, and the merchant's historical after-sales records to complete information feature extraction and semantic alignment. The decision network identifies the dispute as a product quality-related after-sales dispute and determines the full-process task nodes: standardized solution generation → evidence review → merchant negotiation → user online negotiation → timeout outbound call negotiation.

[0165] The main control intelligent module queries the global skill registry, verifies the calling permissions of each task module, and issues scheduling instructions to the task intelligence layer. First, it calls the solution decision module to dynamically plan a processing strategy based on order risk, user credit rating, and platform after-sales rules: synchronously generating a first response solution (full refund of 269 yuan, compensation of 12 yuan for standard platform shipping), and asynchronously generating a second response solution (full refund + 15 yuan actual shipping compensation) and writing it to the work order record module. Then, it calls the voucher verification module to verify the voucher. The voucher verification module uses multimodal recognition capabilities to verify the voucher uploaded by the user, confirming that the dress is indeed damaged and determining the voucher to be valid.

[0166] After the voucher is approved, the main control intelligent module calls the solution negotiation module. First, it sends a compensation request to the merchant, who confirms and agrees to cover the quality issue refund and shipping cost compensation. Then, it sends the first response solution to the user and initiates online negotiation. If the user does not respond within the 24-hour SLA deadline, the main control intelligent module receives a timeout event, updates the strategy via the reinforcement learning strategy network, and dispatches the intelligent outbound calling module. The intelligent outbound calling module automatically dials the user's number, retrieves the second response solution from the work order record module, and sends it (full refund + 15 yuan actual shipping cost compensation). After the user agrees to the solution over the phone, the comprehensive adjudication module generates the final adjudication result and issues an execution instruction.

[0167] The work order recording module retains a complete record of all actions, including user request access, voucher verification results, merchant negotiation records, user online negotiation SLA timeouts, and outbound call negotiation results, generating traceable after-sales work orders. It also synchronously triggers asynchronous events such as timeouts and outbound call completions, providing real-time feedback to the main control intelligent layer. The execution instructions from the comprehensive adjudication module are sent to the capability supply center (execution engine), and the system automatically completes the 269 yuan refund and 15 yuan shipping fee compensation for the user, pushing a compensation receipt notification to the user through the user interaction layer.

[0168] In another exemplary embodiment, the question-and-answer information processing system can seamlessly integrate online real-time consultation with offline work order application processes. For example... Figure 4 As shown, a question-and-answer information processing system may include: User interaction layer: includes the online customer service portal (real-time chat) on the user's end, offline dispute channels (after-sales application on the order page, complaint center), and the customer service interaction interface that the user finally sees.

[0169] Message and Session Layer: Includes an instant messaging module, a session memory module, and an offline message push system, responsible for real-time sending and receiving of online messages, session context storage, and asynchronous notification of offline messages.

[0170] The core layer of the intelligent agent includes an intelligent agent workflow SDK (Software Development Kit), a master control intelligent module, a planner, and multiple task intelligent modules. These task intelligent modules include, but are not limited to: a decision-making tool (corresponding to a solution decision module), an online / offline evidence submission tool, a credential verification module, a platform-vendor negotiation module (a specific implementation of the solution negotiation module), and a platform-user negotiation module (a specific implementation of the solution negotiation module). The master control intelligent module receives requests from online or offline channels, calls the planner to decompose tasks, and schedules the task intelligent modules according to their skill constraints.

[0171] Execution and Closed-Loop Layer: This layer includes an execution engine (capability supply center), a work order recording module, a distributed message queue, and an action recording message listening module. The execution engine is responsible for executing the final solution; the work order recording module records the service status across the entire chain; the distributed message queue supports asynchronous event processing; and the action recording message listening module captures the execution logs of each module and writes them to the work order recording module.

[0172] Basic support layer: includes robot question answering module, intelligent agent gateway, external event access module, etc.

[0173] Suppose a user purchases a 20-inch suitcase for 329 yuan on an e-commerce platform. After receiving the suitcase, they discover that one of the casters is broken and unusable. At 23:00 that same day, they submit an after-sales application through the after-sales portal on the order page (offline dispute channel), requesting a "full refund of 329 yuan" and uploading a photo of the broken wheel. The system initiates a fully intelligent processing flow. At 09:30 the next day, when the merchant confirms and agrees to the compensation, the user happens to be browsing the platform application, and the system automatically switches to the online channel to push the processing result. Based on this, the question-and-answer information processing system can execute the following process: Phase 1: The user submits an after-sales application through an offline dispute channel, uploading photos of the broken suitcase wheels, filling in order information and a refund request. The application request enters the system via the intelligent agent gateway. After identity verification and parameter parsing, the gateway stores the order information, user request, and supporting documentation in the session memory module. The intelligent agent workflow SDK retrieves the complete context from the session memory module and calls the main control intelligent module to initiate the decision-making process.

[0174] Phase 2: The main control intelligent module calls the planner to analyze the user's request and context, identifying it as a "product quality dispute," and formulates an execution plan: first, verify the evidence, then negotiate compensation with the merchant. The planner schedules the evidence verification module to intelligently verify the uploaded photos, confirming that the wheel breakage is real and the evidence is valid. Next, it schedules the platform-merchant negotiation module to send a processing notification to the merchant's backend: "Your order XXX has a product quality issue. The platform has determined to agree to the user's full refund request. Please confirm within 24 hours; otherwise, the system will automatically execute." At 09:30 the next day, the merchant's backend replies "Agree to the processing plan," and this reply is integrated into the system as an external event. After receiving the merchant's agreement event, the planner calls the judgment plan tool to generate the final processing plan: "Agree to a full refund of 329 yuan, which will be returned to the original payment method within 1-3 business days." The intelligent agent workflow SDK converts the processing plan into executable actions and asynchronously writes all intermediate results to the work order record module. The work order record module writes the action record "Merchant has agreed, plan generated" to a distributed message queue.

[0175] Phase 3: The action log message listening module captures the "Solution generated" event in the queue in real time and proceeds to channel judgment. The channel judgment module identifies that the work order comes from an offline channel and continues to determine the user's online status. The user online status judgment module detects that the user is currently browsing the platform APP and is online, so it automatically switches to the online channel to push the result.

[0176] Phase 4: The Intelligent Agent Workflow SDK pushes the final processing solution to the chatbot Q&A module. The chatbot Q&A module proactively initiates a conversation with the user's online customer service interface via the instant messaging module, pushing the processing result: "Hello, regarding your refund request for the XXX suitcase, the merchant has agreed to a full refund of 329 yuan. The amount will be returned to your payment account via the original payment method within 1-3 business days." The user replies "Okay, thank you" on the online customer service interface, confirming their agreement to the solution. The Intelligent Agent Workflow SDK receives the user's confirmation message and calls the execution engine to complete the refund operation. The execution engine writes the execution result to the work order record module, officially closing the process loop.

[0177] Optionally, if the user is still offline when the merchant agrees to compensation, the user's online status judgment module determines them as "offline," and proceeds to the "Does planning need to be done?" judgment logic. The system determines that a final solution has already been generated and no further planning is needed; therefore, it calls the offline message push system to send the user an SMS and an app push notification: "Hello, your refund application for order XXX has been approved. 329 yuan will be refunded to the original payment method within 1-3 business days. Click to view details." The user can view the notification after logging back in.

[0178] The following will continue to combine Figure 5 This illustrates a specific implementation process of a question-and-answer information processing method in a real-world scenario. For example... Figure 5 As shown, the following steps may be included: Step 1: The user initiates an after-sales inquiry through the online instant messaging consultation portal, enters the request text "The headphones frequently disconnect, requesting a replacement", and provides the order number XXX. The main control intelligent module receives the consultation request.

[0179] Step 2: The main control intelligent module calls the cross-attention network to fuse multimodal data such as user request text, order information, product model, and user's historical after-sales records to complete the confirmation of user request and extract the core elements: the problem type is product function abnormality, and the request is to replace the product.

[0180] Step 3: The main control intelligent module writes the "confirm user request" action record to the work order record module and persists it.

[0181] Step 4: After the main control intelligent module queries the global skill registry to verify permissions, it sends a capability call request to the solution decision module, passing in the parsed request parameters.

[0182] Step 5: Based on the preset rule base, the solution decision module synchronously returns the first response solution, namely the ISO (Industry Standard Operation) standardized solution: "We agree to process a full refund of 399 yuan for you. You only need to send the goods back, and the platform will bear the shipping cost."

[0183] Step 6: While synchronously returning the first response solution, the solution decision module asynchronously executes the generation logic of the second response solution (a solution outside of ISO rules): calls the merchant inventory interface to verify the inventory of the same model of headphones, calls the logistics interface to estimate the delivery time, calls the compensation rule interface to calculate the replacement cost, and generates the second response solution: "Agreed to replace your headphones with brand new ones of the same model free of charge. It is expected to be shipped within 3 working days, and the round-trip shipping costs will be borne by the platform."

[0184] Step 7: The solution decision module pushes the generated second response solution to the work order record module, and the work order record module persists the "second response solution".

[0185] Step 8: The main control intelligent module writes the "access first response scheme" action record to the work order record module and persists it.

[0186] Step 9: The main control intelligent module calls the built-in reinforcement learning policy network to select a model planning scheme. Based on the user's historical behavior data (the user has previously chosen to exchange goods instead of refunding multiple times), it determines to prioritize pushing the first response scheme and preloads the second response scheme.

[0187] Step 10: The main control intelligent module initiates online negotiation with the user through the user interaction layer, pushing the first response solution: "Hello, after verification, your headphones do indeed have a disconnection problem. We can process a full refund of 399 yuan for you. You only need to send the goods back, and we will reimburse your shipping costs." Step 11: The main control intelligent module writes the "Initiate online solution negotiation" action record to the work order record module and persists it.

[0188] Step 12: The user replies, "I don't want a refund, I want a new one," clearly rejecting the first response plan.

[0189] Step 13: The main control intelligent module writes the "buyer online negotiation rejection" action record to the work order record module and persists it.

[0190] Step 14: The main control intelligent module parses the second response scheme (outside the rules) pre-stored in the work order record module and extracts the core parameters and constraints of the scheme.

[0191] Step 15: The main control intelligent module calls the built-in reinforcement learning strategy network to plan the model. Based on the user's rejection reasons and the parameters of the second response scheme, it generates the final secondary negotiation scheme without having to call the backend interface again to verify the information.

[0192] Step 16: The main control intelligent module initiates a second online negotiation with the user through the user interaction layer, pushing a second response solution: "We fully understand your needs. We have verified that the same model of headphones is in stock and can replace them with brand new headphones for you free of charge. We expect to ship within 3 business days, and we will bear the round-trip shipping costs. Is that acceptable to you?" Step 17: The main control intelligent module writes the "Initiate online negotiation with buyer (outside the rules)" action record to the work order record module and persists it.

[0193] Step 18: The user replies "Yes", agreeing to the second response plan.

[0194] Step 19: The main control intelligent module writes the "buyer agrees online" action record to the work order record module and persists it.

[0195] Step 20: The main control intelligent module sends a plan execution request to the capability supply center (execution engine), passing in the exchange order parameters, logistics information, etc.

[0196] Step 21: The capacity supply center executes the replacement order generation and warehouse shipment instructions, and displays the execution result to the user through the user interaction layer: "Hello, your replacement request has been confirmed. The new headphones will be shipped within 3 business days. The tracking number will be sent to you via SMS. Please check your mail."

[0197] Through the above steps, a dual-track parallel processing mechanism of "synchronously returning the first response scheme + asynchronously generating the second response scheme" is implemented, which solves the response delay problem caused by the need to regenerate the scheme after the user rejects the standardized scheme. The following will combine Figure 6 This document explains the specific application of question-and-answer information processing methods in product quality dispute resolution scenarios where users are required to provide physical proof of purchase. A user purchased a moisturizing face cream priced at 168 yuan from an e-commerce platform. After signing for the package, they discovered a noticeable crack in the bottle and leakage of the cream. They initiated an after-sales inquiry through the online instant messaging system, requesting a full refund of 168 yuan. This scenario falls under the after-sales dispute type where the platform rules explicitly require users to provide physical proof of purchase. In this case, if... Figure 6 As shown, the question-and-answer information processing system can execute the following process: Step 1: The user initiates an after-sales inquiry through the online instant messaging consultation portal, enters the request text "The face cream cracked upon receipt and leaked half of it, requesting a full refund", and provides the order number XXX. The main control intelligent module receives the consultation request.

[0198] Step 2: The main control intelligent module calls the cross-attention network to fuse multimodal data such as user request text, order information, product category, and user's historical after-sales records. It identifies the data as "after-sales dispute related to product damage" through a node classifier. After querying the global skill registry to verify permissions, it sends a capability call request to the solution judgment module, passing in the parsed request parameters.

[0199] Step 3: The solution judgment module calls the solution decision module, and based on the preset "After-sales Dispute Evidence Rules", it determines that the user needs to provide evidence first for this type of dispute, and returns the "Evidence is required as a prerequisite for the solution" processing solution to the main control intelligent module.

[0200] Step 4: The solution decision module returns the evidence example ID "cosmetic_crack_002" corresponding to the scenario to the main control intelligent module.

[0201] Step 5: The main control intelligent module sends a voucher example reasoning request to the voucher review module, passing in the voucher example ID "cosmetic_crack_002".

[0202] Step 6: The voucher verification module generates a personalized voucher example for this scenario based on a pre-trained multimodal model: "Upload a clear photo showing the cracked part of the bottle, leakage marks, product label and express waybill", along with 3 standard reference images, and returns it to the main control intelligent module.

[0203] Step 7: The main control intelligent module sends a request to the work order record module to start the evidence collection task, and sets the evidence collection deadline to 24 hours.

[0204] Step 8: The work order record module generates and persists the "Initiate Buyer Evidence" action record, and pushes the record to the distributed log collection system simultaneously.

[0205] Step 9: The main control intelligent module parses the action records returned by the work order record module, combines them with the voucher example data, and assembles and generates standardized evidence card parameters, including evidence requirements, example images, upload entry, and deadline.

[0206] Step 10: The main control intelligent module sends an evidence card to the user through the user interaction layer, displaying the following content: "Hello, in order to help you process your refund application as soon as possible, please upload a clear photo of the cracked part of the bottle, the leakage marks, the product label and the express waybill within 24 hours."

[0207] Step 11: The user uploads three compliant photos of the supporting documents via the evidence card and submits them to the main control intelligent module.

[0208] Step 12: The main control intelligent module receives the credentials submitted by the user and forwards them to the credentials verification module.

[0209] Step 13: The main control intelligent module writes the "user completes evidence submission" action record to the work order record module and persists it, and pushes it synchronously to the distributed log collection system.

[0210] Step 14: The main control intelligent module sends an intelligent audit request to the voucher audit module.

[0211] Step 15: The voucher verification module calls the multimodal recognition model to intelligently verify the vouchers submitted by the user: First photo verification: Confirms a crack approximately 2cm long in the bottle, with obvious signs of cream leakage. Second photo verification: Confirms the product label matches the order information. Third photo verification: Confirms the recipient information on the courier waybill matches the order. Verification result: "Valid voucher."

[0212] Step 16: The voucher review module returns a "Smart review completed" status notification to the work order record module.

[0213] Step 17: The voucher review module pushes the detailed intelligent review results (review passed + verification details of each voucher) to the main control intelligent module.

[0214] Step 18: After receiving the review results, the main control intelligent module starts the intelligent agent decision-making process.

[0215] Step 19: The main control intelligent module calls the built-in reinforcement learning strategy network, plans the model based on the voucher review results, order amount, and user credit rating, and generates the final processing solution: "Agree to a full refund of 168 yuan, the amount will be returned to the original payment method within 1-3 business days".

[0216] Step 20: The main control intelligent module sends a solution execution request to the capability supply center (execution engine), passing in parameters such as the refund amount and the user's payment account.

[0217] Step 21: After the capacity supply center completes the refund operation, it will play the solution execution result to the user through the user interaction layer: "Hello, your refund application has been approved. 168 yuan will be refunded to your payment account within 1-3 business days. Please check your account." Optionally, if the voucher verification module returns an "invalid voucher" result (e.g., the photo only shows the whole body and cannot clearly identify the crack and leakage marks), step 22 can be executed. That is, the main control intelligent module will send a notification to the user to submit evidence again, clearly stating the reason for invalidity and the correction requirements: "We are very sorry, the photo you uploaded cannot clearly identify the crack and leakage marks on the bottle. Please re-upload a clear close-up photo of the cracked area. Thank you for your cooperation." The process then returns to step 11, waiting for the user to resubmit the voucher.

[0218] This approach, combining evidence example IDs with a multimodal model, enables the automatic generation of personalized evidence examples for different dispute scenarios, solving the problems of inconsistent manual guidance and users' lack of knowledge about how to provide evidence. From receiving credentials and multimodal recognition to result determination, no manual intervention is required throughout the entire process. All key operational nodes are persisted to the work order record module, and a distributed log collection system ensures full-process log backup, providing complete data support for after-sales dispute arbitration and platform compliance audits.

[0219] The following will combine Figure 7This paper details another improved process for question-and-answer information processing. Specifically, this improved process is a hierarchical review based on intelligent initial review and manual segmentation verification, combined with nested processing of user online status and real-time feedback. For example, suppose a user purchases an electric pressure cooker priced at 499 yuan from an e-commerce platform. After 10 days of use, they claim that the non-stick coating on the inner pot has peeled off due to non-human error and initiates after-sales service through online instant messaging, requesting a free replacement of the inner pot. This scenario falls under the category of AI-reviewed ambiguity (unable to 100% distinguish between factory defects and wear caused by improper user use), requiring the initiation of a human-machine collaborative segmentation review process. Figure 7 As shown, this improved process can be implemented based on the following steps: Step 1: The user initiates an after-sales inquiry through the online instant messaging consultation portal, enters the request text "The coating of the inner pot of the electric pressure cooker has fallen off, and I request a new one", and provides the order number XXX. The main control intelligent module receives the consultation request.

[0220] Step 2: The main control intelligent module calls the cross-attention network to fuse multimodal data such as user request text, order information, product model, and user's historical after-sales records, and identifies it as "home appliance quality after-sales dispute" through the node classifier; after querying the global skill registry to verify permissions, it initiates a capability call request to the solution judgment module, passing in the parsed request parameters.

[0221] Step 3: The solution judgment module calls the solution decision module, determines based on preset rules that the user needs to provide evidence first for this type of dispute, and returns a "proof required" processing solution to the main control intelligent module.

[0222] Step 4: The main control intelligent module initiates a solution consultation with the user, informing them of the initial handling intention: "Hello, in order to help you verify the inner liner coating issue, please provide relevant photos, and we will handle it for you as soon as possible." Step 5: The user replies "OK" to agree to the evidence submission plan.

[0223] Step 6: The main control intelligent module parses and assembles the evidence card parameters, including evidence requirements, sample images, upload entry, and 24-hour evidence submission deadline.

[0224] Step 7: The main control intelligent module sends standardized evidence cards to the user through the user interaction layer.

[0225] Step 8: The main control intelligent module writes the "Initiate Buyer Evidence" action record to the work order record module and persists it.

[0226] Step 9: The user uploads three photos of the inner liner coating peeling off using the evidence card and submits them to the main control intelligent module.

[0227] Step 10: The main control intelligent module writes the "user completes evidence submission" action record to the work order record module and persists it.

[0228] Step 11: The main control intelligent module forwards the user-submitted credentials to the credential verification module.

[0229] Step 12: The voucher verification module starts the AI ​​intelligent verification process.

[0230] Step 13: The AI ​​model conducts a preliminary review of the voucher and finds that the inner liner does indeed have coating peeling. However, it also detects obvious scratches from use and high-temperature discoloration. It cannot be 100% determined to be a factory quality defect and is therefore judged to "require manual slice verification".

[0231] Step 14: The voucher review module returns a "Smart review completed" status notification to the work order record module.

[0232] Step 15: The voucher review module returns the review result "slicing required" to the main control intelligent module.

[0233] Step 16: The main control intelligent module sends a "Pending Inquiry" message to the user: "Hello, your voucher has been received and is undergoing manual review. Please wait a moment; we expect to reply to you within 1-2 hours." Step 17: The voucher review module starts the manual segmentation review task, and assigns the vouchers and AI preliminary review results to professional human reviewers specializing in home appliances.

[0234] Step 18: After comparing the quality inspection report of the same batch of goods with the photos uploaded by the user, the manual reviewer confirms that the coating peeling is consistent with the characteristics of the factory quality defect and returns the manual slice review result of "approved".

[0235] Step 19: The voucher review module writes the manual review results into the work order record module and persists them.

[0236] Step 20: The work order recording module sends a "manual review completed" event notification to the main control intelligent module through the message queue.

[0237] Step 21: The main control intelligent module queries the work order record module to obtain the user's current online status.

[0238] Step 22: The main control intelligent module calls the built-in reinforcement learning strategy network to plan the model, and combines the results of manual review with the user's online status to generate the final solution: "Replace the inner liner for the user free of charge and bear the round-trip shipping costs".

[0239] The following will explain in detail Figure 7 Nested branch content: Branch 1, Approval + User Online + User Agreement (Main Process), includes: Step 23: The main control intelligent module initiates a pre-request for solution execution to the capacity supply center. Step 24: The main control intelligent module plays a solution confirmation card to the user through the user interaction layer: "Hello, your application has been approved. We will replace the inner liner for you free of charge and cover the round-trip shipping costs. Is that acceptable?" Step 25: The user replies "Yes," agreeing to the solution. Step 26: The main control intelligent module initiates a confirmation execution request to the capacity supply center. The capacity supply center generates a replacement order and arranges delivery, completing the process loop.

[0240] Branch 2: Approval + User Online + User Change of Mind, including: If the user replies "No, I don't want it anymore, I want a full refund" indicating a change of mind: Step 26: The main control intelligent module selects a higher-level solution based on the preset strategy: "We understand your feelings. We can process a full refund for you. You only need to send the goods back, and we will bear the shipping costs," and negotiate with the user again. Step 27: After the user agrees, the refund process is executed.

[0241] Branch 3: If the user is offline when the manual review is completed, proceed with the following steps: Step 27: The main control intelligent module sends a solution confirmation card to the user: "Hello, your electric pressure cooker inner pot issue has been approved. We will replace it with a brand new inner pot free of charge. Click the link to confirm and we will arrange delivery immediately." Step 28: The user clicks the link to confirm the solution after logging in. Step 29: The main control intelligent module initiates a solution execution request to the capacity supply center.

[0242] Branch 4: If the review fails, meaning the human reviewer determines that the coating peeling is caused by improper user operation and returns a "Review Failed" result, then proceed to step 30. The main control intelligent module will inform the user that the solution is unavailable and the reason: "We are very sorry, after manual verification, your inner liner coating peeling is due to improper use and does not meet the conditions for free replacement. We can offer you a 20% discount on purchasing a new inner liner. Would you like that?" Simultaneously, the next round of negotiation and progressive solutions will be pushed to the user.

[0243] This approach ensures efficiency while improving the accuracy of complex dispute resolution. It automatically selects the appropriate interaction method based on the user's online status: online users confirm in real time, while offline users confirm via asynchronous notification, avoiding user waiting and information omissions.

[0244] The following will combine Figure 8 This document details the implementation of the user-side negotiation process within the platform of the Q&A information processing system. This process seamlessly integrates online SLA timeout detection with outbound telephone negotiation, resolving the issue of long-term work order stagnation caused by users viewing solutions online but not receiving responses.

[0245] Suppose a user purchases a dress priced at 229 yuan on an e-commerce platform. The merchant promises to ship within 48 hours, but the shipment is actually delayed by 3 days. The user initiates an after-sales inquiry through the online instant messaging portal, requesting a 10 yuan no-minimum-spend coupon as compensation. This scenario falls under the category of unilateral after-sales service where the platform can make a direct decision without needing to negotiate with the merchant, and this process applies. In this case, the Q&A information processing system can execute the following platform-user unilateral negotiation process: Step 1: The user initiates an after-sales inquiry through the online instant messaging consultation portal, enters the request text "Your shipment was delayed by 3 days, I request a 10 yuan coupon as compensation", and provides the order number XXX. The main control intelligent module receives the consultation request.

[0246] Step 2: The main control intelligent module calls the cross-attention network to fuse multimodal data such as user request text, order information, shipping records, and user historical after-sales records. It identifies the dispute as "delayed delivery compensation dispute" through the node classifier. After querying the global skill registry to verify permissions, it sends a capability call request to the solution judgment module, passing in the parsed request parameters.

[0247] Step 3: The solution judgment module calls the solution decision module, and based on the preset "Delayed Delivery Compensation Rules", generates a standardized processing solution: "Agree to compensate with a 5 yuan no-threshold coupon, which is valid for all items and is valid for 30 days", and returns to the main control intelligent module.

[0248] Step 4: The main control intelligent module invokes the online negotiation capability of the platform's user negotiation module to initiate a solution negotiation with the user: "Hello, we are very sorry for the inconvenience caused. We will compensate you with a 5 yuan no-threshold coupon. Is that acceptable?" Step 5: The main control intelligent module writes the "Initiate online negotiation with buyer" action record to the work order record module and persists it, while starting a 24-hour SLA timeout timer.

[0249] Step 6: The user viewed the online message but did not reply within the 24-hour preset time limit due to being busy. The work order record module automatically triggered the Service Level Agreement timeout event.

[0250] Step 7: The work order recording module sends an "SLA timeout event notification" to the main control intelligent module through the message queue.

[0251] Step 8: The main control intelligent module writes the "Buyer Online Negotiation Timeout" action record to the work order record module and persists it.

[0252] Step 9: The main control intelligent module calls the built-in reinforcement learning strategy network to plan the model. Analyzing the user's historical behavior data, it finds that the user rarely uses online chat functions and prefers telephone communication. Moreover, the user's demands are not much different from the platform's standard solution. Therefore, an outbound negotiation plan is generated: "Inform the user of the compensation plan by phone and increase the compensation amount to 8 yuan to improve the user's consent rate."

[0253] Step 10: The main control intelligent module provides the platform-user negotiation module with the generated outbound call negotiation script, user reserved information and key context information.

[0254] Step 11: The main control intelligent module writes the "Initiate buyer outbound call negotiation" action record to the work order record module and persists it.

[0255] Step 12: The platform's user negotiation module invokes the outbound call capability to automatically dial the user's number and play the intelligent voice message: "Hello, this is the after-sales customer service of XX platform. Regarding your request for compensation for the delayed shipment of your order XXX, we can offer you an 8 RMB no-minimum-spend coupon, valid for all items and for 30 days. Do you agree to this offer?" Step 13: After the outbound call is completed, the platform-user negotiation module will write the detailed outbound call results (connected / not connected, user feedback, call duration) into the work order record module and persist them.

[0256] Step 14: The work order record module sends an "outbound call completed" event notification to the main control intelligent module.

[0257] Step 15: The main control intelligent module calls the built-in reinforcement learning strategy network to plan subsequent processes based on the outbound call results.

[0258] like Figure 8 As shown, in Branch 1, where the outbound call is connected and the user agrees (main process), the user can reply "agree" on the phone. The main control intelligent module initiates a plan execution request to the capability supply center. The capability supply center automatically issues an 8 yuan no-threshold coupon to the user's account and notifies the user via SMS.

[0259] like Figure 8 As shown in Branch 2, if the outbound call is connected but the user refuses, for example, if the user replies on the phone "No, I want the 10 yuan coupon", the main control intelligent module generates an upgrade plan according to the preset service strategy: "Agree to compensate with a 10 yuan no-threshold coupon", confirm with the user again, and execute the distribution operation after the user agrees; if the user still refuses, it will automatically transfer to human customer service for follow-up.

[0260] like Figure 8As shown in Branch 3, in the case of unanswered outbound calls, for example, if the platform-user negotiation module makes 3 consecutive calls to the user and the call is not answered, the main control intelligent module can generate a follow-up plan: "Send the user an SMS containing an 8 yuan coupon compensation plan and a confirmation link. If the user does not confirm within 48 hours, automatically issue a 5 yuan standard coupon" and execute the corresponding operation.

[0261] By employing an automatic SLA timeout detection mechanism, we promptly identify work orders that users have viewed online but have not responded to, preventing long-term work order stagnation and shortening the average processing time for one-sided after-sales work orders. We have built a multi-channel negotiation system encompassing online messaging, outbound phone calls, and SMS notifications, selecting appropriate outreach methods based on different user preferences to improve effective user reach rates.

[0262] In addition to the platform-user one-sided negotiation process described above, this embodiment also provides a complete two-sided negotiation process for a question-and-answer information processing system, such as... Figure 9 As shown, the complete bilateral negotiation process may include the following steps: Step 1: The user initiates an after-sales inquiry through the online instant messaging consultation portal, enters the request text "The backpack shoulder strap came undone as soon as I received it, I request a full refund", and provides the order number XXX. The main control intelligent module receives the consultation request.

[0263] Step 2: The main control intelligent module calls the cross-attention network to fuse multimodal data such as user request text, order information, product category, and user's historical after-sales records. It identifies the dispute as "product quality-related after-sales dispute" through a node classifier. After querying the global skill registry to verify permissions, it sends a capability call request to the solution decision module, passing in the parsed request parameters.

[0264] Step 3: Based on the preset rule base, the solution decision module synchronously returns the first response solution: "We will compensate you with a 20 yuan no-threshold coupon and provide free on-site repair service."

[0265] Step 4: The main control intelligent module invokes the online negotiation capability of the platform's user negotiation module to initiate a solution negotiation with the user: "Hello, we are very sorry for the inconvenience caused. We can compensate you with a 20 yuan coupon and provide free backpack repair. Is that acceptable?" Step 5: The main control intelligent module writes the "Initiate online negotiation with buyer" action record to the work order record module and persists it, while starting a 24-hour SLA timeout timer.

[0266] Step 6: The user replies "No, I don't want repairs, I want a full refund," clearly rejecting the first response plan.

[0267] Step 7: The main control intelligent module writes the "buyer online negotiation rejection" action record to the work order record module and persists it.

[0268] Step 8: The main control intelligent module parses the second response scheme generated asynchronously in advance in the work order record module: "Full refund of 359 yuan, user does not need to return the goods".

[0269] Step 9: The main control intelligent module calls the judgment logic to determine that the execution of the second solution requires the platform and the merchant to negotiate first and confirm that the merchant agrees to assume full refund responsibility.

[0270] Step 10: The main control intelligent module provides the negotiation objective to the platform-merchant negotiation module: "Request the merchant to assume full refund responsibility for order XXX, totaling 359 yuan."

[0271] Step 11: The platform's merchant negotiation tool writes the "Initiate seller negotiation" action record to the work order record module and persists it.

[0272] Step 12: The platform's merchant negotiation tool automatically sends a negotiation notification to the merchant's backend: "Your order XXX has a product quality issue, and the user requests a full refund. Please confirm within 24 hours. Otherwise, the system will automatically execute the platform's judgment."

[0273] Step 13: After 12 hours, the merchant replies "Agree to full refund". The platform's merchant negotiation tool writes the "Completed seller negotiation" action record to the work order record module and persists it.

[0274] Step 14: The work order recording module sends a "seller negotiation completed" event notification to the main control intelligent module through the message queue.

[0275] Step 15: The main control intelligent module queries the work order record module to obtain the user's current online status and determines that the user is currently offline.

[0276] Step 16: The main control intelligent module calls the built-in reinforcement learning policy network to plan the model and generate a secondary negotiation plan for the user's offline scenario: "Send the full refund plan to the user via SMS or APP push, and wait for the user's confirmation."

[0277] Step 17: The main control intelligent module calls the offline notification capability of the platform's user negotiation tool to initiate negotiation with the user: "Hello, regarding your order XXX backpack refund application, we have reached an agreement with the merchant to refund you a full amount of 359 yuan. You do not need to return the product. After clicking the link to confirm, the payment will arrive in your account within 1-3 business days."

[0278] Step 18: The main control intelligent module writes the "Initiate online negotiation with buyer" action record to the work order record module and persists it, while starting a new round of 24-hour SLA timeout timer.

[0279] Step 19: The user viewed the push message, but did not reply within the 24-hour preset time limit due to being busy. The work order record module automatically triggered the Service Level Agreement timeout event.

[0280] Step 20: The work order record module sends a "buyer negotiation timeout" event notification to the main control intelligent module through the message queue.

[0281] Step 21: The main control intelligent module calls the built-in reinforcement learning strategy network to plan the model. Analyzing the user's historical behavior data, it finds that the user prefers telephone communication. Therefore, it generates an outbound negotiation plan: "Inform the user of the full refund plan by phone and confirm whether the user agrees."

[0282] Step 22: The main control intelligent module provides the outbound call negotiation target and script parameters to the platform-user negotiation tool.

[0283] Step 23: The platform's user negotiation tool writes the "Initiate buyer outbound call negotiation" action record to the work order record module and persists it.

[0284] Step 24: The platform's user negotiation tool invokes the outbound call capability to automatically dial the user's number and play an intelligent voice message: "Hello, this is the after-sales customer service of XX platform. Regarding your refund request for the backpack in your order XXX, we have reached an agreement with the merchant to provide you with a full refund of 359 yuan, without the need to return the product. Do you agree to this solution?" Step 25: The user replies "agree" over the phone, and the platform's user negotiation tool submits the negotiation result of the "user agreement plan" to the main control intelligent module.

[0285] Step 26: The platform's user negotiation tool writes the "Completed buyer outbound call negotiation" action record to the work order record module and persists it.

[0286] Step 27: The work order record module sends an "outbound call completed" event notification to the main control intelligent module.

[0287] Step 28: The main control intelligent module calls the built-in reinforcement learning policy network to perform final model planning, confirms that all preconditions have been met, and generates execution instructions.

[0288] Step 29: The main control intelligent module sends a solution execution request to the capacity supply center. The capacity supply center automatically completes the refund of 359 yuan for the user and notifies the user via SMS.

[0289] like Figure 9 As shown, this includes process branches: Branch 1, User Online After Merchant Negotiation: If step 15 determines that the user is currently online, the main control intelligent module will directly push a full refund plan to the user through the online chat window. The refund will be executed immediately after the user confirms online.

[0290] Branch 2, User Refuses Second Online Negotiation: If the user replies "No, I want extra compensation for shipping" during the second online negotiation, the main control intelligent module will generate an upgrade plan based on the preset service strategy: "Extra compensation of 10 yuan shipping coupon", and negotiate with the user again; if the user still refuses, it will automatically transfer to human customer service for follow-up.

[0291] Branch 3, Outbound Calls Not Connected: If the platform-user negotiation tool makes 3 consecutive calls to the user and the call is not connected, the main control intelligent module will generate a follow-up plan: "Send the user an SMS containing a full refund plan and an automatic confirmation explanation. If the user does not raise any objection within 48 hours, the system will automatically execute the refund."

[0292] In this way, the platform and the merchant can automatically negotiate after the user rejects the standard solution, without the need for manual intervention, thus reducing the rate of manual intervention in complex disputes.

[0293] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 11 to 15 can be device A; or the execution subject of steps 11 and 12 can be device A, and the execution subject of steps 13 to 15 can be device B; and so on.

[0294] Furthermore, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 11, 12, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0295] Figure 10 This is a schematic diagram of the structure of a question-and-answer information processing device provided as another exemplary embodiment of this application. For example... Figure 10As shown, the device includes: a receiving module 1001, configured to: receive user question information in the current session and determine the work order record module corresponding to the question information; a construction module 1002, configured to: identify user intent based on the question information, and construct a decision context for the current session based on the user intent and historical service information recorded in the work order record module; a first scheduling module 1003, configured to: plan a first service node based on the decision context, and if the first service node is a solution generation node, schedule the solution decision module corresponding to the solution generation node among the multiple task intelligence modules; a configuration module 1004, configured to: configure the solution decision module to: synchronously schedule a rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule a reasoning engine to generate a second response solution and write it into the work order record module; and a response module 1005, configured to: if a user's rejection information is received after outputting the first response solution, obtain the second response solution from the work order record module, and perform response processing on the question information according to the second response solution.

[0296] For detailed implementation methods and technical effects of the above modules, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0297] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 11 As shown, in practice, this electronic device includes a memory 1101 and a processor 1102.

[0298] Memory 1101 is used to store computer programs and can be configured to store various other data to support operation on the computing platform. Examples of this data include instructions for any application or method operating on the computing platform, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0299] Processor 1102, coupled to memory 1101, is used to execute computer programs in memory 1101 for: receiving user question information in the current session and determining the work order record module corresponding to the question information; identifying user intent based on the question information and constructing a decision context for the current session based on the user intent and historical service information recorded in the work order record module; planning a first service node based on the decision context; if the first service node is a solution generation node, scheduling a solution decision module corresponding to the solution generation node among the multiple task intelligence modules; the solution decision module is configured to: synchronously schedule a rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule an inference engine to generate a second response solution and write it to the work order record module; if a user's rejection information is received after outputting the first response solution, obtaining the second response solution from the work order record module and processing the question information according to the second response solution.

[0300] For a detailed description of the above operations and their technical effects, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0301] Furthermore, such as Figure 11 As shown, the electronic device also includes: a communication component 1103, a display 1104, a power supply component 1105, an audio component 1106, and other components. Figure 11 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 11 The components shown.

[0302] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0303] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0304] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0305] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0306] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0307] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0308] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0309] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A question-and-answer information processing method, characterized in that, The method, applied to a main control intelligent module, which is associated with multiple task intelligent modules in a target application scenario, includes: Receive the user's problem information in this round of conversation and determine the work order record module corresponding to the problem information; Based on the problem information, the user intent is identified, and based on the user intent and the historical service information recorded in the work order record module, the decision context for this round of conversation is constructed. Based on the decision context, a first service node is planned. If the first service node is a solution generation node, the solution decision module corresponding to the solution generation node among the multiple task intelligence modules is scheduled. The solution decision module is configured to: generate a first response solution and return it based on its skill constraint description, and generate a second response solution and write it into the work order record module based on the asynchronous scheduling inference engine; If a user's rejection message is received after the first response solution is output, the second response solution is obtained from the work order record module, and the problem information is processed according to the second response solution.

2. The method according to claim 1, characterized in that, The module for determining the work order record corresponding to the problem information includes: Identify the resource object to which the problem information belongs, where the resource object is the smallest resource unit in the target application scenario that can independently create a work order; Using the identifier of the resource object as the query key, query the work order mapping table to see if the resource object has been bound to the work order record module; If it is already bound, then the work order record module that the resource object is bound to is determined as the work order record module corresponding to the problem information; If not bound, a work order record module is created and bound for the resource object in the work order mapping table, and the created work order record module is determined as the work order record module corresponding to the problem information; In this context, the same resource object corresponds to the same work order record module under each service entry point and in each round of sessions in the target application scenario.

3. The method according to claim 2, characterized in that, The work order mapping table has a work order access interface, and each service entry point in the target application scenario is uniformly mapped to the work order access interface. The module for creating and binding work order records for the resource object in the work order mapping table includes: Establish a binding relationship between the identifier of the resource object and the work order access interface, and use the identifier of the resource object as a unique primary key to create a blank work order entry in the work order mapping table as the work order record module of the resource object; The binding relationship is used to record the full-link service information corresponding to the problem information from any service entry point and belonging to the resource object into the work order record module of the resource object.

4. The method according to claim 1, characterized in that, Based on the user intent and the historical service information recorded in the work order record module, a decision context for the current session is constructed, including: Obtain the target service entry point and initiation time corresponding to this round of session, wherein the target service entry point is any one of multiple service entry points in the target application scenario; The model is constructed by inputting the target service entry point, initiation time, and user intent into the context of this round of conversation: Based on the target service entry point, initiation time, and user intent corresponding to this round of conversation, calculate the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information recorded in the work order record module. Based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information, target historical service information that has a contextual relationship with the current session is selected. The target historical service information and the user intent are structured and integrated to obtain the decision context for this round of the session.

5. The method according to claim 4, characterized in that, Based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information entry, target historical service information that has a contextual relationship with the current session is selected, including: Obtain the dynamic feature priority corresponding to the current session, whereby the dynamic feature priority is used to indicate the relative importance of the service entry point, initiation time, and user intent in the current session. Based on the dynamic feature priority, corresponding weight coefficients are assigned to the channel consistency feature, temporal proximity feature, and intent coherence feature of each historical service information. Based on the channel consistency characteristics, temporal proximity characteristics, and intent coherence characteristics of each historical service information and their respective weight coefficients, the contextual relevance of each historical service information to the current session is calculated. Based on the aforementioned contextual relevance, a preset number of target historical service information entries are selected from each historical service information entry.

6. The method according to claim 1, characterized in that, Planning the first service node based on the aforementioned decision context includes: The decision context is vectorized in different dimensions to obtain vector matrices in the service entry dimension, initiation time dimension, and intent dimension; the vector matrix in any dimension includes the current vector of the current session in that dimension and the historical vector of historical service information in that dimension. The vector matrices of the service entry dimension, initiation time dimension, and intent dimension are concatenated into a joint vector matrix; The joint vector matrix is ​​input into the cross-attention network, and different attention heads are used to extract service entry consistency features, temporal proximity features and intent semantic similarity features, which are then fused to obtain the current session state representation. The current session state representation is input into a node classifier or reinforcement learning policy network, and the first service node that matches the current session state representation is output.

7. The method according to any one of claims 1-6, characterized in that, Scheduling the solution decision module corresponding to the solution generation node among the multiple task intelligence modules includes: Read the target skill requirements of the solution generation node, and based on the target skill requirements and the skill constraint descriptions predefined by the multiple task intelligence modules, select the solution decision module from the multiple task intelligence modules; Based on the decision context and the skill constraint description of the solution decision module, structured prompt words are generated for the solution decision module; The solution decision module is scheduled according to the structured prompt words, which triggers the solution decision module to synchronously schedule the rule engine to generate a first response solution and return it based on its skill constraint description, and asynchronously schedule the inference engine to generate a second response solution and write it into the work order record module.

8. The method according to claim 7, characterized in that, The skill constraint description of the solution decision module includes at least: the available toolset and concurrent execution logic of the solution decision module; The available toolset includes at least: the rule engine, the inference engine, and the work order writing tool; The concurrent execution logic is used to instruct the scheme decision module to perform the following operations in parallel: The rule engine is synchronously scheduled to perform keyword matching based on the decision context in the structured prompt words and the predefined rule base to generate a first response scheme and return it. The inference engine is asynchronously scheduled to perform semantic understanding and comprehensive reasoning based on the decision context in the structured prompt words, thereby generating a second response scheme. The work order writing tool is then called to write the second response scheme into the work order record module.

9. The method according to any one of claims 1-6, characterized in that, According to the second response scheme, the question information is processed to provide a response, including: If a confirmation message is received from the user after the second response scheme is output, the execution engine corresponding to the second response scheme is scheduled to perform the corresponding response action to process the problem information.

10. The method according to claim 9, characterized in that, The work order recording module is also used to record at least one of the following service information generated in this round of session: The problem information, the user intent, the first response plan, the process of negotiating the first response plan with the user, the user's rejection of the first response plan, the process of negotiating the second response plan with the user, and the user's confirmation of the second response plan.

11. The method according to any one of claims 1-6, characterized in that, Also includes: If a confirmation message is received from the user after the first response solution is output, then the question information is responded to according to the first response solution. or If a rejection message is received from the user after the second response scheme is output, the intelligent outbound call module among the multiple task intelligent modules is scheduled to conduct telephone negotiation with the user through the intelligent outbound call module.

12. The method according to claim 11, characterized in that, After receiving confirmation from the user regarding the first response scheme or the second response scheme, the method further includes: The second service node is planned according to the first response scheme or the second response scheme. If the second service node includes a credential verification node, the credential verification module corresponding to the credential verification node among the multiple task intelligent modules is scheduled. The credential review module is configured to: based on its skill constraint description, sequentially issue credential cards, receive credential materials submitted by the user through the credential cards, perform intelligent review of the credential materials, and output the review results; If the audit result is passed, then the operation of responding to the problem information according to the first response scheme or the second response scheme will be performed.

13. The method according to claim 12, characterized in that, Also includes: If the review result is unsuccessful and the received information is the user's confirmation of the first response plan, then the second response plan is obtained from the work order record module, and the problem information is responded to according to the second response plan. If the review result is unsuccessful, and the received information is the user's confirmation of the second response plan, then the comprehensive adjudication module in the multiple task intelligent modules is scheduled to generate a third response plan, and the problem information is processed according to the third response plan.

14. The method according to claim 12, characterized in that, Also includes: If the second service node includes a merchant negotiation node, then the solution negotiation module corresponding to the merchant negotiation node among the multiple task intelligence modules is scheduled; The solution negotiation module is configured to: write an asynchronous negotiation start action to the work order record module based on its skill constraint description, drive the negotiation strategy model to execute the merchant negotiation process, and write the execution result of the merchant negotiation process to the work order record module.

15. An electronic device, characterized in that, include: A memory and a processor; wherein the memory is used to store a computer program, and the processor is coupled to the memory for executing the computer program to perform the method as described in any one of claims 1 to 14.

16. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the method as described in any one of claims 1 to 14.

17. A computer program product, characterized in that, include: A computer program / instruction that, when executed by a processor, causes the processor to perform the steps of the method according to any one of claims 1 to 14.

Images