Task processing method and device, electronic equipment and storage medium

Abstract

The invention provides a task processing method and device, electronic equipment and a storage medium, relates to the technical field of agents, and is used for improving the accuracy of task processing. Checking whether an effective task session corresponding to the user request exists or not; the effective task session refers to a task session which exists in a cache, is associated with the user request and is not completed; under the condition that the effective task session does not exist, intention recognition processing is conducted on the user request, and at least one intention category and the confidence coefficient corresponding to each intention category are generated; determining a target agent from a plurality of agents based on the intention category and the corresponding confidence; the target agent comprises a first agent used for processing a specific task or a second agent used for processing a complex task; and calling the target agent to process the user request.

Description

Technical Field

[0001] This application relates to the field of intelligent agent technology, and in particular to a task processing method, apparatus, electronic device, and storage medium. Background Technology

[0002] In the field of artificial intelligence technology, especially with the maturity of large language models (LLMs) technology, its application to enterprise-level business scenarios (such as intelligent customer service and automated business processing) has become an important development trend.

[0003] However, when dealing with highly specialized and knowledge-intensive business queries, large models suffer from inaccurate content generation, making it difficult for their output to meet the high rigor requirements of enterprise-level applications.

[0004] Therefore, improving the accuracy of task processing has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a task processing method, apparatus, electronic device, and storage medium for improving data transmission rates.

[0006] Firstly, this application provides a task processing method, which includes: receiving a user request; checking whether a valid task session corresponding to the user request exists; a valid task session refers to a cached task session that is associated with the user request and is not yet completed; if no valid task session exists, performing intent recognition processing on the user request to generate at least one intent category and a confidence level corresponding to each intent category; determining a target intelligent agent from multiple intelligent agents based on the intent category and the corresponding confidence level; the target intelligent agent includes a first intelligent agent for processing a specific task or a second intelligent agent for processing a complex task; and invoking the target intelligent agent to process the user request.

[0007] In one possible implementation, the target agent is determined from multiple agents based on the intent category and its corresponding confidence level, including: if the confidence level of a single intent category is higher than a preset threshold, the first agent corresponding to the single intent category is determined as the target agent according to a preset mapping relationship; wherein, the preset mapping relationship is the mapping relationship between intent categories and agents.

[0008] In one possible implementation, the target agent is determined from multiple agents based on the intent category and its corresponding confidence level, including: generating a selection list when the confidence levels of multiple intent categories are higher than a preset threshold, the selection list including multiple intent categories for the user to select the target intent from the multiple intent categories; and determining the agent corresponding to the target intent as the target agent according to the target intent and a preset mapping relationship.

[0009] In one possible implementation, the target agent is determined from multiple agents based on the intent category and its corresponding confidence level. This includes: if the confidence level of all intent categories is lower than a preset threshold, querying the cache for historical session records associated with the user request and marked as ended; if historical session records exist, determining the agent that was most recently successfully invoked as the target agent based on the historical session records; if no historical session records exist, determining the second agent as the target agent.

[0010] In one possible implementation, invoking the target agent to process the user request includes: if the target agent is a second agent, performing task planning on the user request through the second agent to generate a task sequence containing at least one sub-task; for each sub-task in the task sequence, determining the corresponding execution agent from multiple agents; and invoking the execution agent to process the corresponding sub-task.

[0011] In one possible implementation, the method further includes: if a valid task session exists, obtaining the agent associated with the valid task session from the valid task session and identifying the agent as the target agent.

[0012] In one possible implementation, when the user request is in a multi-turn dialogue, the method further includes: when the user request's task session meets the preset completion conditions, updating the status identifier to the normal completion status, and determining whether to perform intent extension processing on the user request; intent extension processing refers to the process of re-parsed the deep intent of the user request; if intent extension processing is performed, the user request is re-processed for intent recognition; if intent extension processing is not performed, the user request's task session is terminated.

[0013] In one possible implementation, the method further includes: updating the status identifier of the user request based on the execution status of the user request; the status identifier is used to characterize the processing stage of the user request; wherein the status identifier includes unprocessed state, processing state, suspended state, invalid state, abnormal termination state, and normal termination state.

[0014] In one possible implementation, the status flag of the user request is updated based on the execution status of the user request, including: if the user request is in a multi-turn dialogue, updating the status flag to a suspended state.

[0015] Secondly, this application provides a task processing apparatus, comprising: a processing unit and a communication unit; the communication unit is configured to receive a user request; the processing unit is configured to check whether a valid task session corresponding to the user request exists; a valid task session refers to a cached task session associated with the user request that has not yet been completed; the processing unit is further configured to perform intent recognition processing on the user request in the absence of a valid task session, generating at least one intent category and a confidence level corresponding to each intent category; the processing unit is further configured to determine a target intelligent agent from multiple intelligent agents based on the intent category and the corresponding confidence level; the target intelligent agent includes a first intelligent agent for processing a specific task or a second intelligent agent for processing a complex task; the processing unit is further configured to invoke the target intelligent agent to process the user request.

[0016] Thirdly, this application provides an electronic device comprising a processor and a memory. The memory stores processor-executable instructions, and when the processor is configured to execute the instructions, the electronic device implements the method described in the first aspect above.

[0017] Fourthly, this application provides a readable storage medium comprising software instructions. When the software instructions are executed in an electronic device, they cause the electronic device to implement the method described in the first aspect above.

[0018] Fifthly, this application provides a computer program product comprising computer instructions. When the computer instructions are executed on an electronic device, the electronic device performs the method described in the first aspect.

[0019] The beneficial effects of the second to fifth aspects mentioned above can be referred to the first aspect, and will not be repeated here.

[0020] The technical solution provided in this application offers at least the following beneficial effects: This technical solution achieves efficient, accurate, and resource-optimized intelligent task processing through a progressive series of technical steps. First, by prioritizing the checking of valid task sessions upon receiving a request, duplicate processing of the same request is avoided, significantly improving system response efficiency. Next, intent recognition is initiated when no valid session exists, ensuring that each new request is accurately parsed, laying the foundation for subsequent intelligent routing.

[0021] Furthermore, by dynamically selecting the target agent based on intent confidence, simple tasks are quickly processed by the efficient first agent, while complex tasks are handled by the more capable second agent, thus achieving on-demand allocation and optimal utilization of computing resources.

[0022] Ultimately, the target intelligent agent is invoked to execute the request, forming a closed loop from intent understanding to resource scheduling to task execution. This ensures both immediate response in simple scenarios and effective resolution of complex problems, thereby improving the overall adaptability and processing capabilities of the system. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the architecture of a task processing system provided in an embodiment of this application; Figure 2 This is a schematic diagram of the architecture of another task processing system provided in an embodiment of this application; Figure 3 A flowchart illustrating a task processing method provided in an embodiment of this application; Figure 4 A flowchart illustrating another task processing method provided in an embodiment of this application; Figure 5 A flowchart illustrating another task processing method provided in an embodiment of this application; Figure 6 A flowchart illustrating another task processing method provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a task processing device provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0027] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0028] In the field of artificial intelligence technology, especially with the maturity of large language models (LLMs) technology, its application to enterprise-level business scenarios (such as intelligent customer service and automated business processing) has become an important development trend.

[0029] However, when dealing with highly specialized and knowledge-intensive business queries, large models suffer from inaccurate content generation, making it difficult for their output to meet the high rigor requirements of enterprise-level applications.

[0030] To address these issues, intelligent systems in related technologies typically employ a loosely integrated approach, combining multiple functional modules (e.g., intent recognition, knowledge base, and business process engine). This high degree of coupling and complex interfaces between modules leads to high costs for overall system integration, development, and maintenance, making it difficult to support rapid iterations and changes in business requirements.

[0031] In open, conversational interactive environments, ensuring the protection of sensitive enterprise data, implementing granular user permission verification, and guaranteeing that AI behavior complies with security and compliance standards are key obstacles preventing large-scale models from deeply integrating into core business scenarios. Related intelligent chat systems lack end-to-end security protection designs that span the entire business process. Current solutions such as intelligent customer service robots, task-oriented dialogue systems, and retrieval-augmented generation (RAG)-based knowledge question-answering systems still have limitations in their overall architecture. They fail to effectively integrate the general understanding capabilities of large models with the precise execution capabilities of smaller models or specialized systems. Furthermore, they lack a unified, global intelligent capability coordination and scheduling mechanism, resulting in insufficient overall efficiency, accuracy, and adaptability in handling complex business processes.

[0032] Therefore, how to improve the accuracy of task processing is a problem that needs to be studied.

[0033] Based on this, this application provides a task processing method that, upon receiving a request, prioritizes checking for valid task sessions to avoid duplicate processing of the same requests, significantly improving system response efficiency. Then, when no valid session exists, intent recognition is initiated to ensure that each new request is accurately parsed, laying the foundation for subsequent intelligent routing.

[0034] Furthermore, by dynamically selecting the target agent based on intent confidence, simple tasks are quickly processed by the efficient first agent, while complex tasks are handled by the more capable second agent, thus achieving on-demand allocation and optimal utilization of computing resources.

[0035] Ultimately, the target intelligent agent is invoked to execute the request, forming a closed loop from intent understanding to resource scheduling to task execution. This ensures both immediate response in simple scenarios and effective resolution of complex problems, thereby improving the overall adaptability and processing capabilities of the system.

[0036] Figure 1 This is a schematic diagram of the architecture of a task processing system provided in an embodiment of this application, such as... Figure 1 As shown, the task processing system includes a top layer, an access layer, a core layer, a support layer, and a platform layer.

[0037] In some embodiments, the top layer is the interface layer for interaction between the task processing system and the user, which is responsible for receiving user requests in multiple modalities and presenting the results in a rich format.

[0038] Specifically, the top layer includes front-end touchpoints, assistant interaction tools, UI card tools, and digital humans.

[0039] It should be understood that the assistant interaction tool handles voice interaction, converting user speech to text through speech recognition, and then converting the system's text responses back to speech through voice broadcasting. The UI card tool handles graphical interaction, understanding user intent through feature extraction and generating cards (such as buttons, lists, and charts) as responses. The digital human provides human-like interaction, communicating with the core system through an interaction API, and its expressions and movements are controlled by the image management module.

[0040] In some embodiments, the access layer serves as a unified gateway and scheduling layer, acting as the entry point and central scheduling center for all user requests, ensuring the standardization and secure control of requests.

[0041] Specifically, the access layer includes a central control intelligent service for touchpoints and a unified capability development platform.

[0042] For example, the Touchpoint Control Intelligent Service receives requests from front-end touchpoints, performs compliance checks and filters on the input content of the requests, standardizes requests of different formats from different front-end touchpoints into a unified format within the system, and routes them to the corresponding Business Control Intelligent Service based on the request content.

[0043] The unified capability development platform is used to provide reusable general AI capabilities (such as general intent recognition and information extraction) for upper-layer applications and to support the rapid development of new businesses.

[0044] In some embodiments, the core layer is responsible for understanding user intent and intelligently scheduling appropriate resources to process business.

[0045] Specifically, the core layer includes a business overall control intelligent service and business capability units. The business overall control intelligent service is used to parse the intent of user requests, decompose complex tasks into processes and plan steps, and can dynamically select to call the first or second intelligent agent for user requests based on the confidence level of the intent.

[0046] A business capability unit can be multiple, referring to a dedicated processing unit for each specific business area (such as customer service, marketing, and operations). For example, there are business capability units A, B, C, and D. Business capability unit A can be used for intent recognition, prompt word engineering, service standardization, and security verification; business capability unit B can be used for information extraction, business parameters, access control, and service standardization; business capability unit C can be used for service standardization and security verification; and business capability unit D can be used for service standardization and security verification.

[0047] In some embodiments, the support layer is used to provide the intelligent agents required for the intelligent decision-making and execution of the core layer.

[0048] The support layer includes a first intelligent agent, a second intelligent agent, and a knowledge base. The first intelligent agent is designed to handle specific tasks, the second intelligent agent is designed to handle complex tasks, and the knowledge base provides accurate and up-to-date professional knowledge to both agents to ensure accurate responses.

[0049] In some embodiments, the platform layer includes a data annotation platform and a dataset management platform for preparing data, as well as a model training and inference platform and an AI application development platform for creating and deploying models.

[0050] Figure 2 This is a schematic diagram of the architecture of another task processing system provided in an embodiment of this application, such as... Figure 2As shown, this task processing system adopts a front-end and back-end separation architecture. Through clear boundaries of responsibility, it decouples the front-end interaction layer from the back-end business logic, building an efficient and flexible multimodal intelligent system. This task processing system includes a front-end and a back-end.

[0051] In some embodiments, the front end serves as a unified interaction entry point, mainly including modules such as authentication and authorization, multimodal input, input parsing, security management, session management, and card display, and is integrated and uniformly distributed through a front end scheduling service.

[0052] For example, when a user requests "Check this month's bill" via voice, the front end first verifies the user's identity through the authentication module. Then, the voice recognition submodule converts the voice stream into the text "Check this month's bill" in real time, completing the initial input parsing. The built-in conversion and normalization processing module transforms heterogeneous data into a unified text stream, that is, it organizes input data of different formats from different channels (such as APP, mini-program, webpage) into a standardized request message with standard fields (such as user ID, session ID, intent text). And it calls the backend service based on the standardized interface protocol.

[0053] Furthermore, standardized interface protocols can be implemented based on modern remote procedure call protocols, ensuring communication efficiency.

[0054] In some embodiments, the backend is centered around a central control service, responsible for the intelligent scheduling and execution of business processes. The central control service includes core components such as central control planning and decision-making, security permission verification, intelligent capability invocation, card matching, and data security verification. This central control service can invoke different standardized business workflows (e.g., Business 1 workflow, Business 2 workflow, Business 3 workflow) based on the request type. Each business workflow includes steps such as semantic understanding, task planning and decision-making, business checks, slot matching, service invocation (e.g., calling the service list for Task 1), and organization return. For example, Business 1 workflow includes semantic understanding, task planning and decision-making, business checks, calling the service list for Business 1, and organization return; Business 2 workflow includes security verification, slot matching, business checks, calling the service list for Business 2, and organization return; Business 3 workflow includes slot matching, process matching, business checks, calling the service list for Business 3, and organization return.

[0055] For example, for complex services such as "applying for international roaming," the central control service can dynamically orchestrate multiple business workflow nodes such as "qualification verification, package inquiry, and real-name authentication." Through intelligent scheduling strategies, the central control service can flexibly organize and trigger corresponding capability combinations based on the context of front-end requests to complete complex business logic processing.

[0056] For example, simple requests with high confidence (such as "check phone bill") are directly routed to a dedicated small-model agent for processing; while fuzzy or complex requests are routed to a general-purpose large model for deep planning. In this architecture, the front-end focuses on the development of the user interface and user controls, while the back-end focuses on the implementation, scheduling, and operational support of business capabilities. This architectural design reduces the coupling and integration complexity between system modules, lowers development and operational costs during system integration and iteration, enhances the system's adaptability to changes in business requirements, significantly shortens the development cycle, and improves response speed.

[0057] It should be noted that the task processing system provided in this application embodiment has security protection capabilities, end-to-end session management capabilities, intelligent scheduling mechanisms, business capability support capabilities, and data training and knowledge management capabilities.

[0058] The security protection capabilities adhere to the "principle of least privilege," implementing refined permission management. Specifically, the system implements security controls at multiple levels: strong authentication (such as two-factor authentication) is performed when users log in; access permissions are verified when a session is established; and during continuous interaction, for each business operation (such as "querying detailed records"), the system verifies in real time whether the user has the corresponding functional and data permissions (e.g., only able to query local detailed records). By deeply embedding security elements such as identity authentication, permission verification, and data protection into the business logic, a complete operation traceability mechanism is established. The system generates an immutable security log for each critical operation, recording the operator, time, action, and object, forming a complete audit traceability chain. Throughout the entire process from user login, session establishment, continuous interaction to session termination, the system dynamically verifies the user's functional permissions, data permissions, and session permissions to ensure system operation security, data access security, and information display security, providing users with a safe and reliable intelligent interactive environment.

[0059] For example, when a user with low privileges requests access to data with high privileges, the system interrupts the process during the business check and returns a permission denied message to ensure data security.

[0060] In some embodiments, end-to-end session management capabilities establish a comprehensive session management system covering user business inquiries, problem-solving, and business processing. A unified session ID is used throughout the entire process. This session ID is generated when a session is created and uniquely associates with a specific user's continuous interaction sequence within its effective lifecycle (e.g., 30 minutes). This enables closed-loop management of the entire lifecycle, from session access, intent recognition, planning decisions to service invocation.

[0061] Specifically, such as Figure 2As shown, the system receives user requests and associates or creates sessions through a front-end scheduling service. The planning and decision-making module in the central control service performs intent recognition and task decomposition, and then calls various services in the business workflow (such as calling the service list for Business 1). Finally, the results are organized and returned to the user, forming a closed loop. This system ensures that users can obtain professional and accurate business services through natural language interaction throughout the entire process of initiating, interrupting, and terminating a session, guaranteeing the continuity and integrity of the business experience.

[0062] For example, if a user temporarily interrupts a multi-round conversation and re-enters, the system can restore the context through the session identifier, and the user can seamlessly resume the conversation simply by saying "Continue with the business you just did".

[0063] In some embodiments, the intelligent scheduling mechanism decomposes complex problems through a central intelligent service, classifying them hierarchically according to organizational and business architectures. For example, for a complex request such as "complaining about package charges," the central service can decompose it into three sub-task sequences: "querying package orders," "verifying charges rules," and "generating a complaint ticket."

[0064] The front end uses the touchpoint control service to quickly distribute business requests according to the organizational structure. For example, requests from VIP channels are prioritized and routed to a dedicated service queue. The back end uses the business control service to route requests to the corresponding specialized processing units based on user intent.

[0065] like Figure 2 As shown, the overall business control service routes requests to workflows 1, 2, or 3 based on semantic understanding results. For example, a request with the intent of "query" is routed to workflow 1, while a request with the intent of "processing" is routed to workflow 2. A secure asynchronous communication protocol and a high-concurrency asynchronous call algorithm are employed to schedule collaborative work among various professional application capabilities. For instance, a message queue (MQ) is used to achieve asynchronous decoupling communication between services, improving system throughput. Returned data is integrated according to decision-making strategies to ensure timely response in simple scenarios and accurate processing in complex scenarios, ensuring the correctness and timeliness of business processing across the entire domain.

[0066] In some embodiments, business capability support breaks down the "planning, decision-making, execution, and optimization" business loop into reusable long and short process meta-capabilities. A meta-capability is the smallest schedulable unit encapsulating specific business logic; for example, "user authentication" is a short process meta-capability, while "generating business reports" is a long process meta-capability. End-to-end business processing is completed within milliseconds to seconds through the collaboration of multiple intelligent capabilities. Long and short process meta-capabilities represent the smallest granularity of planning and decision-making task decomposition, and model planning is based on this for the collaboration of multiple intelligent capabilities. Based on a high-precision industry model, supplemented by an industry knowledge base, user input is received and a task list is generated. A dynamic confidence optimization strategy is used to execute the workflow.

[0067] Specifically, the system calculates a confidence score for each identified intent. When the score is higher than a preset threshold (e.g., 90%), a pre-defined standardized workflow is directly triggered; when it is lower than the threshold, a large model is triggered to make in-depth decisions or the user is asked to clarify. This is the dynamic confidence optimization strategy, enabling efficient processing of business processes.

[0068] In some embodiments, the data training and knowledge management capabilities employ a hybrid annotation mode of "human annotation and machine pre-annotation." First, a machine learning model performs preliminary intent labeling on the original corpus, followed by review, correction, and refinement by human annotators, significantly improving annotation efficiency and consistency. Intent categories and sentiment tendencies are labeled according to unified standards, and cross-validation ensures annotation quality.

[0069] For example, the phrase "I want to complain" is uniformly labeled with the intent "complaint" and the sentiment "negative," and the labeling results are cross-validated by different labelers. This model provides high-quality data support for model training, ensuring an error rate of less than 2%, a data quality compliance rate consistently above 95%, and an intent recognition rate of over 95% for small models. Simultaneously, through knowledge preprocessing (such as denoising and deduplication), vectorized storage, and hierarchical weighted management, an efficient knowledge retrieval solution is constructed. Combined with large models and Retrieval-Augmented Generation (RAG) technology, it effectively supplements domain knowledge and improves the accuracy and precision of problem-solving.

[0070] For example, when a user inquires about the latest policies, the system uses RAG technology to first retrieve relevant policy points from the latest knowledge base, and then provides them to the large model to generate an accurate answer, avoiding the large model from generating outdated or false information.

[0071] Figure 3 This is a flowchart illustrating a task processing method provided in an embodiment of this application, as shown below. Figure 3 As shown, the method includes the following steps: S101, Receive user request.

[0072] In some embodiments, user requests may originate from various front-end interaction touchpoints, such as intelligent voice assistants, graphical user interfaces, or digital human interfaces.

[0073] The requested content can be multimodal input, including text, voice, images, or a combination thereof. The system receives user requests through a unified API gateway and performs preliminary standardization processing. For example, voice input is converted into text using automatic speech recognition technology, or image input is subjected to optical character recognition to extract key text information.

[0074] For example, a user says "I want to check my international roaming charges from last month" through the voice assistant of a mobile app. The system microphone captures the voice signal and transmits it to the backend service through the audio interface. The voice recognition module converts it into structured text data and encapsulates it into a standard data packet containing user identifier, session identifier, request timestamp, and request text for subsequent processing.

[0075] S102. Check if there is a valid task session corresponding to the user request.

[0076] A valid task session refers to a cached task session that is associated with a user request and is not yet completed. Specifically, "not completed" means that the session's status code is in the initial, active, or suspended state, rather than in the normal or abnormal termination state; "associated" means that the user identifier, device identifier, or session identifier of the current request matches the key identifier of the session record in the cache.

[0077] In some embodiments, the inspection process is implemented by querying the session management cache. The system extracts key identifiers from the current request, such as the user ID or device ID, and uses these as keys to look up the data in the distributed cache. Criteria for determining a valid task session include that the session is in a non-terminated state and has not timed out.

[0078] For example, the user ID is extracted from the received user request data packet, and a query is performed using that user ID as the key. If a record is found with a status field of "active" and a creation time within its 30-minute validity period, a valid task session is determined to exist. If no record with this key is found in the cache, or the record status is "completed," or the record has exceeded its 30-minute validity period, a valid task session is determined not to exist.

[0079] S103. In the absence of a valid task session, perform intent recognition processing on the user request to generate at least one intent category and a confidence level corresponding to each intent category.

[0080] In some embodiments, intent recognition processing can be performed by a dedicated intent recognition model, which is a classifier trained on a deep neural network. The process includes segmenting and vectorizing the request text, then inputting it into the model for inference, and the output layer using a Softmax function to calculate the probability of belonging to each preset intent category, which is the confidence level.

[0081] For example, the user request text "I want to check my international roaming charges from last month" is input into the intent recognition model. The model first performs preprocessing, such as word segmentation into ["I", "want", "check", "last month", "of", "international roaming", "charges"], and then converts it into a sequence of word vectors. After model calculation, the output might be: {"Check bill": 0.85, "Apply for international roaming": 0.10, "Complain": 0.05}. This indicates that the system recognizes "Check bill" as the primary intent with a confidence level of 85%; and "Apply for international roaming" as the secondary intent with a confidence level of 10%. The confidence threshold can be preset to 70%.

[0082] S104. Based on the intent category and the corresponding confidence level, determine the target agent from multiple agents.

[0083] The target intelligent agent includes a first intelligent agent for handling a specific task or a second intelligent agent for handling a complex task. For example, the first intelligent agent may be a miniaturized, high-efficiency dedicated model or rule engine optimized for a specific business scenario, such as a model specifically for querying phone bills and applying for packages; the second intelligent agent may be a large language model with powerful general semantic understanding and task planning capabilities.

[0084] In some embodiments, the determination of the target agent can be any of the following four cases.

[0085] Case (1): If the confidence level of a single intent category is higher than the preset threshold, the first agent corresponding to the single intent category is determined as the target agent according to the preset mapping relationship.

[0086] It should be understood that this situation (1) is for scenarios with high confidence and clear intent.

[0087] In this embodiment, the preset threshold is a configurable system parameter, such as 0.8 (or 80%), used to distinguish the degree of clarity in intent recognition. A confidence level higher than this threshold indicates that the model has a high degree of certainty in judging the intent.

[0088] The preset mapping relationship is the mapping relationship between intent categories and agents. This mapping relationship can be stored in a configuration library or database, and its core is the key-value pair of "intent-agent". For example, the intent "query bill" is mapped to "bill query agent", and the intent "recharge phone credit" is mapped to "recharge service agent".

[0089] For example, the user request is "query this month's traffic", and the intent recognition model outputs the highest confidence intent as "query traffic", with a confidence of 0.95. The system determines that 0.95 > 0.8 (preset threshold), which meets condition (1). Subsequently, the system queries the preset mapping table and finds the agent corresponding to the intent "query traffic" as "traffic query dedicated module" (first agent), and identifies it as the target agent.

[0090] In one scenario, a user says "I want to recharge my phone credit by 50 yuan" via voice. The intent recognition result is a single intent "recharge phone credit" with a confidence level of 0.92. Based on the mapping relationship, the system directly routes the request to the efficient and low-cost "online recharge dedicated service" (the first intelligent agent), thereby quickly completing the recharge operation, avoiding unnecessary resource overhead and latency caused by calling large models, and improving processing efficiency.

[0091] Case (2): When there are multiple intent categories with confidence levels higher than the preset threshold, a selection list is generated, and the agent corresponding to the target intent is determined as the target agent based on the target intent and the preset mapping relationship.

[0092] It should be understood that this situation (2) refers to a scenario where the processing intent is ambiguous but the confidence level is high. By introducing human-computer interaction to clarify the user's true intent, it is a manifestation of the system's intelligence and flexibility.

[0093] The selection list includes multiple intent categories, allowing users to choose their desired intent from these categories. This selection list can be presented to the user through a graphical interface (such as a group of buttons) or voice interaction (such as "Do you want to check your bill or subscribe to a package?").

[0094] For example, the user's request is "My Package". The intent recognition outputs two high-confidence intents: "Query Package Details" (confidence 0.85) and "Change Package" (confidence 0.82). The system generates a selection list and displays the two options through the APP interface: "Query Current Package" and "Change Package". After the user clicks "Query Current Package", the system, based on the selected target intent "Query Package Details", queries the mapping relationship and determines the target agent as "Package Query Core Service" (first agent).

[0095] In one scenario, a user inputs "I want to resolve a data usage issue." The system identifies two intents: "data top-up" (confidence 0.88) and "data package complaint" (confidence 0.84). The system returns options for the user to choose from. After the user explicitly selects "data package complaint," the request is accurately routed to the "complaint ticket processing system" (the first intelligent agent), effectively avoiding incorrect routing caused by ambiguous intents and improving processing accuracy and user experience.

[0096] Case (3): If the confidence level of all intent categories is lower than the preset threshold, query the cache for historical session records associated with the user request and marked as ended. If historical session records exist, determine the most recently successfully invoked agent as the target agent based on the historical session records.

[0097] It should be understood that this situation (3) is an intelligent degradation strategy. When the system cannot clearly understand the current intention, it prioritizes the user's recent behavior patterns and strives to provide continuous or related services, which reflects the system's adaptive capabilities.

[0098] For example, a user inputs a vague request, "Let's see that." All intents identified have a confidence level below 0.8 (the highest confidence level is only 0.6). The system then uses the user ID as the key to query the cache for the user's most recent session record with a status of "Normal End," and finds that the user successfully used the "Monthly Bill Inquiry" service 2 minutes ago. Therefore, the system identifies the "Bill Inquiry Agent" (the first agent) as the target agent and executes the query operation, assuming the user's intent is "Let's see that bill again."

[0099] In one scenario, a user initially inquired about "5G plans" and immediately followed up with, "What about the Gold Edition?". The confidence level for the intent recognition of "What about the Gold Edition?" was very low. By reviewing the query history, the system discovered that the previous session intent was "inquire about plans," and thus routed the current request to the "Plan Inquiry Agent," correctly understanding the user's intent as "inquire about details of the Gold Edition 5G plan," ensuring the continuity of the dialogue context.

[0100] Case (4): If there is no historical session record, the second agent is identified as the target agent.

[0101] It should be understood that this situation (4) is the ultimate safeguard mechanism, which enables the most capable and versatile second agent (such as a large model) to handle complex, open or novel requests when the system is unable to accurately identify the intent and has no historical clues to follow.

[0102] For example, after the system queries the cache and finds that the user has no valid historical session records (e.g., the user is a new user or the session has expired), the current ambiguous request "see that" will be directly routed to the "General Large Language Model Agent" (the second agent). This second agent will attempt to understand the user's intent and, through its powerful generative capabilities, engage in multiple rounds of interaction with the user to clarify the request, or directly invoke its built-in knowledge and reasoning abilities to provide a solution.

[0103] In one scenario, a user, using the service for the first time, makes a complex request: "Help me choose a mobile phone plan that's most suitable for frequent business travelers and offers good value for money, and compare it with my current plan." This request is complex, has low confidence, and has no history. The system routes it to a second intelligent agent. The second agent can break down this task: first, it understands the needs of "frequent business travelers" and "good value for money," then it searches the plan database for intelligent recommendations, and finally, it uses the price comparison function to generate a structured comparison suggestion and returns it to the user. This demonstrates the core advantage of the second intelligent agent in handling complex, multi-step tasks.

[0104] S105, Invoke the target agent to process the user request.

[0105] In some embodiments, when the target agent is a second agent, the second agent performs task planning on the user request, generates a task sequence containing at least one sub-task, and for each sub-task in the task sequence, determines the corresponding execution agent from among multiple agents, and then calls the execution agent to process the corresponding sub-task.

[0106] The task sequence can be a directed acyclic graph (DAG), where nodes represent subtasks and edges represent execution dependencies between subtasks (such as sequential execution or parallel execution).

[0107] For example, the system queries the mapping relationship between the subtask type and the agent's capabilities based on the subtask type (such as "query", "calculation", "approval"), and assigns a suitable agent (usually the first agent) that specializes in this type to each subtask to achieve optimal processing.

[0108] It should be understood that if the target intelligent agent is a second intelligent agent, it means that complex user requests need to be handled.

[0109] For example, a user requests, "I want to apply for international roaming and check the tariff rates." This request is complex and is routed to a second agent (the larger model). The second agent performs task planning, generating a task sequence: [Subtask A: Verify if the user's plan supports international roaming -> Subtask B: Obtain the roaming tariff for the user's destination -> Subtask C: Activate roaming for the user]. Subsequently, the system assigns an executing agent to each subtask: Subtask A is assigned to the "User Plan Verification Service" (first agent A), Subtask B to the "Tariff Query Engine" (first agent B), and Subtask C to the "Business Processing Core System" (first agent C). The system calls these executing agents in parallel or sequentially according to dependencies, and finally summarizes the results of each subtask (such as "Plan Supported," "Destination Tariff List," "Activation Successful"), with the second agent generating a natural language response: "International roaming has been successfully activated for you. Your tariff rate in country A is...".

[0110] In one scenario, a user submits a complex request: "Please plan a two-day trip to Beijing next weekend, including budget and attraction reservations." The second intelligent agent first breaks down the request into a sequence of sub-tasks: 1. Identify user interests (history, nature, etc.) -> 2. Search for Beijing attractions open on weekends -> 3. Generate a draft itinerary -> 4. Estimate transportation and accommodation costs -> 5. Simulate reservation feasibility. Subsequently, the system schedules different specialized intelligent agents: sub-tasks 1 and 3 are completed by the second intelligent agent itself (leveraging its powerful understanding and generation capabilities); sub-task 2 calls the "real-time information retrieval agent" (the first intelligent agent) to obtain the latest open information; sub-task 4 calls the "cost calculation agent" (the first intelligent agent) to access real-time price data; and sub-task 5 calls the "reservation interface testing agent" (the first intelligent agent) to check reservation policies. After all results are returned, the second intelligent agent performs data fusion and logical verification, ultimately generating a complete and accurate personalized travel plan and returning it to the user. This process demonstrates the core value of the second intelligent agent in complex task decomposition, cross-system scheduling, and result synthesis.

[0111] The technical solution provided in this application provides at least the following beneficial effects: First, by prioritizing the checking of valid task sessions after receiving a request, the system avoids duplicate processing of the same request, significantly improving system response efficiency. Next, when no valid session exists, intent recognition is initiated to ensure that each new request is accurately parsed, laying the foundation for subsequent intelligent routing. Furthermore, the target agent is dynamically selected based on intent confidence, allowing simple tasks to be quickly processed by an efficient first agent, while complex tasks are handled by a more capable second agent, achieving on-demand allocation and optimal utilization of computing resources. Finally, the target agent is invoked to execute the request, forming a closed loop from intent understanding to resource scheduling to task execution, ensuring both immediate response in simple scenarios and effective resolution of complex problems, thus improving the overall adaptability and processing capacity of the system.

[0112] It should be understood that the above description assumes the absence of a valid task session. The following provides a detailed overview of the scenario where a valid task session exists.

[0113] Figure 4 A flowchart illustrating another task processing method provided in this application embodiment is shown below. Figure 4 As shown, the method includes the following steps: S201. If a valid task session exists, obtain the agent associated with the valid task session from the valid task session and identify the agent as the target agent.

[0114] It should be understood that this scenario refers to a scenario where a valid task session exists, meaning that the user request is a follow-up interaction to an ongoing multi-turn dialogue. Its core logic is to skip time-consuming intent re-identification and directly utilize the processing units already determined in the historical session context to ensure dialogue continuity and response efficiency.

[0115] For example, when a user first asks "I want to apply for international roaming" in the current session, the system, after steps S101-S105, ultimately identifies the target agent as the "International Roaming Service Processing Engine" and records it in the session cache. When the user subsequently says "Help me activate roaming to the United States," the system checks for the existence of this valid session in step S102 and directly reads the "agent_id: International Roaming Service Processing Engine" field from the session record in S201, directly identifying it as the target agent for this request. This allows for seamless continuation of the previous application process without the user needing to repeat their overall intent.

[0116] It should be understood that in the case of multi-turn dialogues, this indicates that a valid task session related to the user's request exists in the cache, and thus... Figure 5 A flowchart illustrating another task processing method provided in this application embodiment is shown below. Figure 5 As shown, the method includes the following steps: S301. When the task session requested by the user meets the preset completion conditions, the status flag is updated to the normal completion status.

[0117] In some embodiments, the preset completion condition is a configurable set of rules, and the main judgment criteria include, but are not limited to: the planning engine determines that the current task sequence has been successfully executed and the planning engine determines, based on the dialogue context and task logic, that there is no need to generate new subsequent tasks (for example, after a user queries a bill and the system returns the result, the dialogue goal has been achieved).

[0118] For example, a user engages in multiple rounds of dialogue with the system to process a "package change." After a series of sub-tasks, including "confirming identity, selecting a new package, and confirming the change," the system successfully processes the new package for the user. The planning engine determines that all necessary steps have been completed, satisfying the preset completion conditions. The system then updates the session status from "active" to "normally ended" and records the reason for termination as "task successfully completed."

[0119] S302. Determine whether to perform intent extension processing on the user request.

[0120] Intent extension processing refers to the process of re-analyzing the deeper intent of a user's request. Intent extension processing is a remedial and deepened understanding mechanism, typically involving the invocation of a more powerful second agent to re-analyze the user's initial request or historical dialogue at a broader intent space or deeper semantic level, in an attempt to uncover potential or unexpressed intents.

[0121] For example, a user requests "Help me with this," but the planning engine cannot understand the specific meaning of "handle" or generate a task based on the current context. However, the intent extension flag for the current session is "not extended." The system determines that there is a possibility of clarifying the user's goal through intent extension, and therefore decides to perform intent extension processing. Conversely, if the flag is already "extended," it decides not to perform the process to prevent an infinite loop.

[0122] S303. If intent extension processing is performed, the user request will be reprocessed for intent recognition.

[0123] For example, the system determines the intention to extend. Subsequently, the system creates a new task session, the ID of which is associated with the original session. The user's current ambiguous request "Please help me with this" and the previous dialogue context are sent to the second agent for intention extension identification. The second agent performs deep analysis and may output a more accurate intention hypothesis, such as "The user may want to check the bill" or "The user may want to file a complaint," along with a new confidence level. Based on this result, the system will enter a new round of the S104 decision process.

[0124] S304. If intent extension processing is not performed, the user-requested task session is terminated.

[0125] For example, the system detects that the intent extension flag has already been set, and therefore determines that no extension will be performed. The system then performs a termination operation. Specifically, the current session status is updated to "Normal Termination," and a session termination log is recorded, noting "Intent extension has been performed once, no further extensions will be made." Simultaneously, a generic closure message is returned to the user, such as "Sorry, we cannot process this; please describe your problem again for assistance," and related data for this session is cleared from the cache. This mechanism ensures the efficient utilization of system resources.

[0126] Figure 6 A flowchart illustrating another task processing method provided in this application embodiment is shown below. Figure 6 As shown, the method includes the following steps: S401. Based on the execution status of the user request, update the status identifier of the user request.

[0127] The status identifier is used to characterize the processing stage of the user request; the status identifier includes, but is not limited to, unprocessed status, processing status, suspended status, invalid status, abnormal termination status, and normal termination status.

[0128] Specifically, the unprocessed state indicates that the user request has been received by the system but has not yet started processing; it is the initial state of the task.

[0129] The processing status indicates that the task is being executed by the target intelligent agent and system resources have been allocated.

[0130] The suspended state indicates that the task execution has been actively paused, usually waiting for external events (such as user input or the resumption of dependent services).

[0131] An invalidation status indicates that the session has been automatically terminated due to timeout (e.g., no interaction for more than 30 minutes).

[0132] An abnormal termination status indicates that the task was terminated due to abnormal circumstances such as system errors, permission verification failures, or business rule conflicts.

[0133] A normal completion status indicates that the task has successfully completed all processing steps according to the expected process.

[0134] In some embodiments, when a user request is in a multi-turn dialogue, the update status is marked as suspended.

[0135] It should be understood that the suspension state transition occurs in scenarios where the current task execution process needs to be temporarily interrupted to await subsequent user input. When setting the suspension state, the system starts a timer. If no user response is received within the preset time, the state is automatically changed from suspended to invalid to avoid resource hoarding.

[0136] For example, when a user requests "international roaming service," during a multi-round dialogue with the system, the system prompts "Please enter your passport number." At this point, because the system needs to wait for the user to enter key information before continuing the process, it updates the current session status from "processing" to "suspended." The system also records the reason for the suspension (waiting for passport number) and starts a 10-minute countdown timer. If the user provides their passport number within 10 minutes, the system restores the status to "processing" and continues the operation; if there is no response within the timeout, the system automatically updates the status to "invalid," releases relevant resources, and sends a session timeout notification to the user. This mechanism ensures business continuity while effectively avoiding the unnecessary use of system resources.

[0137] In an exemplary embodiment, this application also provides a task processing device that can be applied to the first traffic transmission control device at the data receiving end described above. Figure 7 This is a schematic diagram of a task processing device provided in an embodiment of this application. Figure 7 As shown, the task processing device 700 includes a processing unit 701 and a communication unit 702.

[0138] In some embodiments, the communication unit 702 is configured to receive user requests; Processing unit 701 is configured to check whether a valid task session corresponding to the user request exists; a valid task session refers to a task session that exists in the cache, is associated with the user request, and is not yet completed; processing unit 701 is further configured to perform intent recognition processing on the user request in the absence of a valid task session, generating at least one intent category and a confidence level corresponding to each intent category; processing unit 701 is further configured to determine a target agent from multiple agents based on the intent category and the corresponding confidence level; the target agent includes a first agent for handling a specific task or a second agent for handling a complex task; processing unit 701 is further configured to invoke the target agent to handle the user request.

[0139] In some embodiments, the processing unit 701 is specifically configured to determine the first agent corresponding to the single intent category as the target agent according to a preset mapping relationship when the confidence level of a single intent category is higher than a preset threshold; wherein, the preset mapping relationship is the mapping relationship between intent categories and agents.

[0140] In some embodiments, the processing unit 701 is specifically configured to generate a selection list when the confidence of multiple intent categories is higher than a preset threshold. The selection list includes multiple intent categories for the user to select a target intent from the multiple intent categories. Based on the target intent and a preset mapping relationship, the agent corresponding to the target intent is determined as the target agent.

[0141] In some embodiments, the processing unit 701 is specifically configured to, when the confidence level of all intent categories is lower than a preset threshold, query the cache for historical session records associated with the user request and whose status is marked as ended; if historical session records exist, determine the agent that was most recently successfully invoked as the target agent based on the historical session records; if no historical session records exist, determine the second agent as the target agent.

[0142] In some embodiments, the processing unit 701 is specifically configured to, when the target agent is a second agent, perform task planning on the user request through the second agent to generate a task sequence containing at least one sub-task; for each sub-task in the task sequence, determine the corresponding execution agent from multiple agents; and call the execution agent to process the corresponding sub-task.

[0143] In some embodiments, the processing unit 701 is further configured to, in the presence of a valid task session, obtain the agent associated with the valid task session from the valid task session and identify the agent as the target agent.

[0144] In some embodiments, the processing unit 701 is further configured to update the status identifier to a normal completion status when the user-requested task session meets the preset completion conditions, and determine whether to perform intent extension processing on the user request; intent extension processing refers to the processing procedure for re-parsed the deep intent of the user request; if intent extension processing is performed, the user request is re-processed for intent recognition; if intent extension processing is not performed, the user request task session is terminated.

[0145] In some embodiments, the processing unit 701 is further configured to update the status identifier of the user request based on the execution status of processing the user request; the status identifier is used to characterize the processing stage of the user request; wherein, the status identifier includes unprocessed state, processing state, suspended state, invalid state, abnormal termination state, and normal termination state.

[0146] In some embodiments, the processing unit 701 is specifically configured to update the status identifier to a suspended state when the user request is in a multi-turn dialogue.

[0147] Of course, the task processing device 700 includes, but is not limited to, the unit modules listed above. Furthermore, the specific functions that the aforementioned functional units can implement include, but are not limited to, the functions corresponding to the method steps in the above embodiments. For detailed descriptions of other modules of the task processing device 700, please refer to the detailed descriptions of their corresponding method steps; these descriptions will not be repeated here.

[0148] In an exemplary embodiment, this application also provides a computer program product that, when run on a computer, causes the computer to execute the aforementioned related method steps to implement the task processing method described above.

[0149] In an exemplary embodiment, this application also provides an electronic device. Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device may include a processor 1101 and a memory 1102; the memory 1102 stores instructions executable by the processor 1101; when the processor 1101 is configured to execute the instructions, the electronic device implements the method as described in the foregoing method embodiments.

[0150] In an exemplary embodiment, this application also provides a computer-readable storage medium storing computer program instructions thereon; when the computer program instructions are executed by an electronic device, the electronic device performs the method as described in the foregoing embodiments. The computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0151] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer-executable instructions. When these computer-executable instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer-executable instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer-executable instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks, SSDs).

[0152] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0153] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

[0154] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A task processing method characterized by, The method includes: Receive user requests; Check if there is a valid task session corresponding to the user request; the valid task session refers to a task session that exists in the cache, is associated with the user request, and is not yet completed. In the absence of a valid task session, the user request is processed for intent recognition, generating at least one intent category and a confidence level corresponding to each intent category; Based on the intent category and the corresponding confidence level, a target agent is determined from multiple agents; the target agent includes a first agent for handling a specific task or a second agent for handling a complex task. The target intelligent agent is invoked to process the user request.

2. The method of claim 1, wherein, The step of determining the target agent from multiple agents based on the intent category and the corresponding confidence level includes: If the confidence level of a single intent category is higher than a preset threshold, the first agent corresponding to the single intent category is determined as the target agent according to a preset mapping relationship. The preset mapping relationship is the mapping relationship between intent categories and intelligent agents.

3. The method of claim 2, wherein, The step of determining the target agent from multiple agents based on the intent category and the corresponding confidence level includes: If the confidence level of multiple intent categories is higher than the preset threshold, a selection list is generated, which includes the multiple intent categories, for the user to select a target intent from the multiple intent categories; Based on the target intent and the preset mapping relationship, the agent corresponding to the target intent is determined as the target agent.

4. The method according to claim 1, characterized in that, The step of determining the target agent from multiple agents based on the intent category and the corresponding confidence level includes: If the confidence level of all intent categories is lower than the preset threshold, query the cache for historical session records associated with the user request that are marked as ended. If the historical session record exists, the agent that was most recently successfully invoked is identified as the target agent based on the historical session record. If the historical session record does not exist, the second agent will be identified as the target agent.

5. The method according to claim 4, characterized in that, The step of invoking the target intelligent agent to process the user request includes: When the target agent is the second agent, the second agent performs task planning on the user request and generates a task sequence containing at least one sub-task. For each subtask in the task sequence, a corresponding execution agent is determined from the plurality of agents; The execution agent is invoked to process the corresponding subtask.

6. The method according to claim 1, characterized in that, The method further includes: If a valid task session exists, the agent associated with the valid task session is obtained from the valid task session, and the agent is identified as the target agent.

7. The method according to claim 6, characterized in that, In the case where the user request involves a multi-turn dialogue, the method further includes: When the user-requested task session meets the preset completion conditions, the status identifier is updated to the normal completion status, and it is determined whether to perform intent extension processing on the user request; the intent extension processing refers to the processing procedure used to re-parse the deep intent of the user request. If the intent extension process is performed, the user request will be reprocessed for intent recognition. If the intent extension process is not performed, the user-requested task session is terminated.

8. The method according to claim 1, characterized in that, The method further includes: Based on the execution status of the user request, update the status identifier of the user request; the status identifier is used to characterize the processing stage of the user request. The status identifiers include unprocessed status, processed status, suspended status, failed status, abnormal termination status, and normal termination status.

9. The method according to claim 8, characterized in that, The step of updating the status identifier of the user request based on the execution status of processing the user request includes: If the user request is in a multi-turn dialogue, update the status identifier to the suspended state.

10. A task processing device, characterized in that, The device includes: a processing unit and a communication unit; The communication unit is used to receive user requests; The processing unit is used to check whether there is a valid task session corresponding to the user request; the valid task session refers to a task session that exists in the cache, is associated with the user request, and is not yet completed; The processing unit is also configured to perform intent recognition processing on the user request in the absence of a valid task session, and generate at least one intent category and a confidence level corresponding to each intent category. The processing unit is further configured to determine a target agent from multiple agents based on the intent category and the corresponding confidence level; the target agent includes a first agent for handling a specific task or a second agent for handling a complex task; The processing unit is also used to invoke the target intelligent agent to process the user request.

11. An electronic device, characterized in that, The electronic device includes: a processor and a memory; The memory stores instructions that the processor can execute; When the processor is configured to execute the instructions, it causes the electronic device to implement the method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes: computer software instructions; When the computer software instructions are executed in an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1 to 9.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed on the processing device, the method of any one of claims 1 to 9 is implemented.

Images