Page interaction method, system, computer device, and storage medium

Abstract

The present disclosure provides a page interaction method and system, computer device and storage medium, and belongs to the technical field of artificial intelligence. The technical scheme provided by the present disclosure can trigger a target task through the interaction operation of a terminal, and a target intelligent agent can determine to use a model of at least one side of a cloud AI service cluster and an end-side model according to the specific situation and running state of the target task, for example, a complex task is handed over to the model of the cloud end in the case of good network state, the powerful generation capability of the cloud model guarantees the quality and efficiency of the generated interaction intention information, thereby improving the page interaction efficiency, and handing over a simple task to the end-side model can realize instant response and improve the page interaction efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a page interaction method, system, computer device, and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, model-based inference for interactive web pages has gained widespread attention in e-commerce marketing, advertising landing page design, and personalized content display. Currently, terminals can upload user interaction information to cloud models. After the cloud model infers the results, these results can be returned to the terminal for rendering. However, network transmission between the terminal and the cloud causes inference latency. Alternatively, models can be deployed on the terminal, where they infer based on user interaction information to obtain rendering results. Terminal model inference can eliminate network latency between the terminal and the cloud. However, terminal models have limited computing power, making it difficult to handle complex inference tasks, resulting in sluggish responses and impacting interaction efficiency. In summary, current methods of implementing page interaction through models suffer from inference latency or insufficient computing power, which ultimately translates into page response latency, leading to low page interaction efficiency. Summary of the Invention

[0003] This disclosure provides a page interaction method, system, computer device, and storage medium, which can improve page interaction efficiency. The technical solution of this disclosure is as follows: According to one aspect of the embodiments of this disclosure, a page interaction method is provided, including: In response to interactive operations on the terminal, the target agent determines at least one target model from at least one side of the terminal model and the cloud AI service cluster based on the target task and running status corresponding to the interactive operation. The running status indicates at least one of the terminal running status, the terminal model running status, and the network status. Interactive intent information is generated based on the target task using at least one target model; Based on the interaction intent information, a component tree of the page to be generated is determined, wherein the components in the component tree are used to indicate the page structure of the page to be generated; Based on the component tree and the first response data of the target task, the first page is displayed.

[0004] According to one aspect of the present disclosure, a page interaction system is provided, including: an end-to-cloud scheduler, a page description module, and a page generation module; The edge-cloud scheduler is used to respond to a terminal triggering a target task, and based on the target task and the running status, determine at least one target model from at least one side of the edge-side model and the cloud AI service cluster, and generate interactive intent information based on the target task through the at least one target model. The running status indicates at least one of the terminal running status, the edge-side model running status and the network status. The page description module is used to determine the component tree of the page to be generated based on the interaction intent information, wherein the components in the component tree are used to indicate the page structure of the page to be generated; The page generation module is used to display a first page based on the component tree and the first response data of the target task.

[0005] In some embodiments, the system further includes an intelligent agent gateway, which is configured to perform rule matching between the target task and multiple first rules through a rule engine in response to a terminal triggering a target task; The page generation module is used to obtain second response data based on the matching rules and display a second page when the target task successfully matches the plurality of first rules. The edge-cloud scheduler is used to determine at least one target model from at least one side of the edge-side model and the cloud-based AI service cluster, based on the target task corresponding to the interaction operation and its running status, when the target task fails to match the plurality of first rules.

[0006] In some embodiments, the edge-cloud scheduler is used for: The evaluation score of the terminal model is determined based on the task complexity, running status, and terminal evaluation strategy of the target task corresponding to the interactive operation. Based on the task complexity, running status, and cloud evaluation strategy of the target task corresponding to the interactive operation, the evaluation score of the cloud AI service cluster is determined. If the evaluation score of the edge model is higher than the evaluation score of the cloud AI service cluster, the edge model is determined as the target model; if the evaluation score of the cloud AI service cluster is higher than the evaluation score of the edge model, the target model is determined from the cloud AI service cluster.

[0007] In some embodiments, the edge-cloud scheduler is used for: Based on the target task corresponding to the interactive operation, the session context of the target task, and the running state, at least one target model is determined from at least one side of the terminal model and the cloud AI service cluster. Based on the target task and the session context, the at least one target model generates interaction intent information.

[0008] In some embodiments, the edge-cloud scheduler is used for: Based on the target task and the running status, multiple sub-tasks are determined, and the target model corresponding to each sub-task is determined from at least one side, the edge model and the cloud AI service cluster. Each sub-task corresponds to an inference stage of the target task. Each target model performs a corresponding inference phase for its respective sub-task to generate interactive intent information.

[0009] In some embodiments, the intelligent agent gateway is configured to perform rule matching between the target task and multiple second rules through a rule engine when the at least one target model call fails; The page generation module is used to obtain third response data based on the matched second rules and display a third page when the target task and multiple second rules are successfully matched. The edge-cloud scheduler is used to, when the target task does not match any of the second rules and the at least one target model is an edge model, invoke a model in the cloud AI service cluster to generate interactive intent information based on the target task; and when the target task does not match any of the second rules and the at least one target model is a model in the cloud AI service cluster, invoke the edge model to generate interactive intent information based on the target task.

[0010] In some embodiments, the system further includes an inference engine module; The inference engine module is configured to respond to a model invocation command for a target model, determine a first adapter corresponding to the target model, translate the model invocation command into an invocation command for the target inference engine on which the target model depends through the first adapter, and drive the target inference engine to run the target model.

[0011] In some embodiments, the inference engine module includes a post-processor; The post-processor is used to verify the inference results generated by the target model. If the verification fails, a verification failure event is reported; if the verification succeeds, the inference results are corrected to generate interactive intent information, and an inference completion event is reported.

[0012] In some embodiments, the system further includes an intelligent agent framework module; The agent framework module is configured to respond to an agent invocation command for a target agent, determine a second adapter corresponding to the target agent, translate the agent invocation command into an invocation command of the target agent framework on which the target agent depends through the second adapter, and drive the target agent framework to run the target agent.

[0013] In some embodiments, the system further includes an agent gateway. The intelligent agent gateway is used to respond to a terminal triggering a target task by querying the session context associated with the target task in memory based on the current terminal window.

[0014] In some embodiments, the agent gateway is further configured to: The task execution log of the target task is written to the memory to update the session context, and the session context is stored in memory in independent partitions according to the corresponding session identifier.

[0015] In some embodiments, the system further includes a model management module, the model management module being used for: In response to a model update event that detects a new model, the new model is preloaded; After the new model is loaded, the service pointer is switched from the old model to the new model through an atomic operation; The old model is marked as drained, and the system waits for the requests being processed on the old model to be completed before unloading the old model.

[0016] In some embodiments, the model management module is used to mark the old model as being in a drained state, and then, if the processing time of the old model for the request meets a preset timeout condition, unload the old model.

[0017] In some embodiments, the model management module is used for: In response to a rollback instruction, the service pointer is switched back from the new model to the old model via an atomic operation.

[0018] In some embodiments, the model management module is used for: In response to the completion of downloading the main model resource package obtained from the main content delivery network, a first integrity check is performed on the main model resource package; if the first integrity check passes, the main model resource package is loaded into memory; if the first integrity check fails, a backup model resource package of the main model resource package is obtained from the backup content delivery network, and the model resource check failure event is reported.

[0019] In some embodiments, the system further includes a component management module, which is configured to perform a second integrity check on the main component resource package in response to the completion of downloading the main component resource package obtained from the main content delivery network; if the second integrity check passes, load the main component resource package into memory; if the second integrity check fails, obtain a backup component resource package of the main component resource package from the backup content delivery network, and report the component resource check failure event.

[0020] In some embodiments, the system further includes a platform abstraction module, which is configured to perform at least one of the following: In response to a call request to the storage factory interface, execute the storage instance corresponding to the runtime environment; In response to a call request to the background task factory interface, execute the background task instance corresponding to the runtime environment; In response to a call request to the event bus factory interface, execute the event communication instance corresponding to the runtime environment.

[0021] In some embodiments, the system further includes a management platform for performing at least one of the following: Check whether the tenant or scenario has the permission to call the interface. If yes, generate an instruction to open the interface function; otherwise, generate an instruction to close the interface function. In response to the model release command, the new model is sent to the tenant that belongs to the grayscale group or the experimental group according to the preset grayscale ratio or experimental group identifier. Collect data corresponding to alarm indicators, and check whether the data corresponding to the alarm indicators meets the corresponding alarm conditions. If so, generate an alarm event. The alarm indicators include at least one of the following: response latency during task execution, task execution success rate, and resource utilization rate. In response to a query request based on a target link identifier, the target query task corresponding to the target link identifier is determined, and a waterfall chart of the target query task is generated based on the task execution log of the target query task. The waterfall chart displays the task processing stages of the target query task on the vertical axis and the time consumed by each task processing stage on the horizontal axis. Check whether the data in the task execution log meets the risk conditions; if so, generate a risk warning event. Resource consumption is calculated according to resource statistics dimensions, which include at least one of tenant, model, and scenario. Provides debugging functionality based on the edge runtime environment.

[0022] According to another aspect of the embodiments of the present disclosure, a computer device is provided, the computer device comprising: One or more processors; Memory used to store the executable program code of the processor; The processor is configured to execute the program code to implement the aforementioned page interaction method.

[0023] According to another aspect of the present disclosure, a computer-readable storage medium is provided that, when program code in the computer-readable storage medium is executed by a processor of a computer device, enables the computer device to perform the above-described page interaction method.

[0024] According to another aspect of the present disclosure, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the above-described page interaction method.

[0025] The technical solution provided in this disclosure allows the terminal's interactive operations to trigger a target task. The target intelligent agent can determine whether to use a model from either the cloud AI service cluster or the terminal model based on the specific circumstances and operating status of the target task. For example, when the network is in good condition, complex tasks can be handled by the cloud model. The powerful generation capabilities of the cloud model ensure the quality and efficiency of the generated interactive intent information, thereby improving page interaction efficiency. Handling simple tasks to the terminal model can achieve instant response and improve page interaction efficiency. Furthermore, a large number of simple tasks are carried by the terminal's computing power, effectively reducing the frequency of calls to the cloud model and reducing resource consumption.

[0026] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0028] Figure 1 This is a schematic diagram illustrating the implementation environment of a page interaction method according to an exemplary embodiment.

[0029] Figure 2 This is a flowchart illustrating a page interaction method according to an exemplary embodiment.

[0030] Figure 3 This is a flowchart illustrating yet another page interaction method according to an exemplary embodiment.

[0031] Figure 4 This is an architecture diagram of a page interaction system according to an exemplary embodiment.

[0032] Figure 5 This is an architecture diagram of yet another page interaction system illustrated according to an exemplary embodiment.

[0033] Figure 6 This is a data flow diagram illustrating a page interaction according to an exemplary embodiment.

[0034] Figure 7 This is a block diagram illustrating a terminal according to an exemplary embodiment. Detailed Implementation

[0035] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0036] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0037] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this disclosure are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0038] Figure 1 This is a schematic diagram illustrating an implementation environment for a page interaction method according to an exemplary embodiment. See also... Figure 1 The implementation environment specifically includes: terminal 101 and server 102. Terminal 101 and server 102 can be connected directly or indirectly through wired or wireless communication, which is not limited herein.

[0039] Terminal 101 is one implementation of a computer device, and can be at least one of the following: a smartphone, a smartwatch, a desktop computer, a laptop, an MP3 player, an MP4 player, and a laptop computer. Terminal 101 runs an application that supports page interaction. Users can log in to the application through Terminal 101 to access the services provided by the application. For example, users can interact with the application on Terminal 101.

[0040] Terminal 101 generally refers to one of a plurality of terminals; this embodiment uses terminal 101 as an example. Those skilled in the art will understand that the number of terminals can be more or less. For example, there may be several terminals, or dozens or hundreds of terminals, or even more. This disclosure does not limit the number of terminals or the type of device.

[0041] Server 102 can be at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. Server 102 provides backend services for applications that support page interaction. Server 102 can connect to terminal 101 via a wireless or wired network, and server 102 can provide the terminal with inference services such as cloud-based AI (Artificial Intelligence) service clusters. In some embodiments, the number of servers may be more or fewer, and this disclosure does not limit this. Of course, server 102 may also include other functional servers to provide more comprehensive and diversified services.

[0042] Figure 2 This is a flowchart illustrating a page interaction method according to an exemplary embodiment, see [link / reference]. Figure 2 The method can be executed by a terminal, and the method includes the following steps: In step 201, in response to the interactive operation on the terminal, the target agent determines at least one target model from at least one side of the terminal model and the cloud AI service cluster based on the target task and running status corresponding to the interactive operation. The running status indicates at least one of the terminal running status, the terminal model running status, and the network status.

[0043] Interactive actions performed by an object on a terminal can include clicking, swiping, and input. For example, if an object enters "I want to check my recent orders" on the platform's smart conversation page and clicks the send control, the corresponding target task is "checking recent orders." If an object enters "item A" in the search box and clicks the search control, the corresponding target task is "searching for item A." If an object scrolls up and down through the item list, the corresponding target task is "browsing more items."

[0044] An edge model refers to an artificial intelligence model deployed on a terminal. This method determines at least one target model from at least one side—the edge model or the cloud AI service cluster—based on the target task and its operational status. This can be achieved by: determining at least one target model from at least one side—the edge model or the cloud AI service cluster—based on the task complexity, privacy sensitivity, and operational status of the target task. For example, the task complexity of a target task can be determined as follows: match the target task with preset task complexity rules using keywords, and determine the task complexity based on the matching results. If the target task is "query orders," the preset task complexity rules are: tasks containing "generate" and "batch" → high complexity, complexity score of 0.8, number of steps greater than 20; tasks containing "query" and "orders" → medium complexity, complexity score of 0.5, number of steps not exceeding 5. The task complexity can be determined based on the number of execution steps; the more execution steps, the higher the task complexity. The preset task complexity rules can be determined by referring to the actual number of execution steps of historical tasks. Of course, the method of determining task complexity is not limited to determining it based on preset task complexity rules. For example, the privacy sensitivity of a target task can be determined by keyword matching based on the target task and preset sensitivity rules. For instance, if the target task matches a preset sensitivity rule: the input information of the task contains an ID card → high sensitivity, sensitivity score is 0.8.

[0045] The operational status indicates at least one of the following: terminal operational status, edge model operational status, and network status. Terminal operational status can be measured based on at least one of the following metrics: CPU (Central Processing Unit) utilization, memory utilization, and disk utilization. These metrics indicate the terminal's resource usage; higher resource usage indicates a worse operational status and a lower score, while lower resource usage indicates a better operational status and a higher score. Network status can be measured based on at least one of the following metrics: network latency and network bandwidth. Lower latency indicates a better network status and a higher score, as does higher bandwidth. Edge model operational status can be evaluated based on the edge model's loading and occupancy status. The closer the edge model is to completion, the better its operational status and a higher score; higher resource usage indicates a worse operational status and a lower score.

[0046] As an optional approach, this method can determine the evaluation score of the edge model based on the task complexity, runtime status, and edge evaluation strategy of the target task corresponding to the interactive operation; and determine the evaluation score of the cloud AI service cluster based on the task complexity, runtime status, and cloud evaluation strategy of the target task corresponding to the interactive operation. If the evaluation score of the edge model is higher than the evaluation score of the cloud AI service cluster, the edge model is determined as the target model; if the evaluation score of the cloud AI service cluster is higher than the evaluation score of the edge model, the target model is determined from the cloud AI service cluster. In other words, this method determines a target model from either the edge model or the cloud AI service cluster, and this target model is either the edge model or a model from the cloud AI service cluster.

[0047] For example, an edge-side evaluation strategy can include the task complexity of the target task and the edge-side weight coefficient of the running state. The evaluation score of the edge model is obtained by weighted summation of the target task's task complexity score and the corresponding edge-side weight coefficient, as well as the running state score and the corresponding edge-side weight coefficient. A cloud-based evaluation strategy can include the target task's task complexity and the cloud-based weight coefficient of the running state. The evaluation score of the cloud AI service cluster is obtained by weighted summation of the target task's task complexity score and the corresponding cloud-based weight coefficient, as well as the running state score and the corresponding cloud-based weight coefficient. Since higher task complexity is more suitable for using large cloud models to execute tasks, the cloud-based weight coefficient of task complexity in the cloud-based evaluation strategy can be greater than the edge-side weight coefficient of task complexity in the edge-side evaluation strategy. This significantly improves the score of high-complexity tasks in cloud evaluation, while low-complexity tasks have a relatively better edge-side evaluation score due to limited cloud score gains. The operational status can include at least one of the following: terminal operational status, edge model operational status, and network operational status. The cloud evaluation strategy can include a cloud weight coefficient corresponding to at least one of these three operational statuses. The edge evaluation strategy can include an edge weight coefficient corresponding to at least one of these three operational statuses. In the terminal evaluation strategy, the edge weight coefficient for the terminal operational status can be greater than the cloud weight coefficient for the terminal operational status in the cloud evaluation strategy. This ensures that a better terminal operational status significantly improves the score in the edge evaluation, while a worse terminal operational status results in limited gain in the terminal score, leading to a relatively better score in the cloud evaluation. Similarly, in the terminal evaluation strategy, the edge weight coefficient for the edge model operational status can be greater than the cloud weight coefficient for the edge model operational status in the cloud evaluation strategy. This ensures that a better edge model operational status significantly improves the score in the edge evaluation, while a worse edge model operational status results in limited gain in the edge score, leading to a relatively better score in the cloud evaluation. In cloud-based evaluation strategies, the cloud-based weighting coefficient for network operation status can be greater than the edge-based weighting coefficient for network operation status in edge-based evaluation strategies. This means that better network operation status results in a significantly higher score in cloud-based evaluation, while worse network operation status leads to a relatively better score in edge-based evaluation due to limited gains in cloud-based scoring. Furthermore, edge-based evaluation strategies can include edge-based evaluation coefficients for the privacy sensitivity of the target task, and cloud-based evaluation strategies can also include cloud-based evaluation coefficients for the privacy sensitivity of the target task. In edge-based evaluation strategies, the edge-based weighting coefficient for privacy sensitivity can be greater than the cloud-based weighting coefficient for privacy sensitivity in cloud-based evaluation strategies. This means that higher privacy sensitivity results in a significantly higher score in edge-based evaluation, while lower privacy sensitivity leads to a relatively better score in cloud-based evaluation due to limited gains in edge-based scoring.

[0048] As an alternative approach, this method can determine a target model from either the edge model or the cloud AI service cluster based on the target task, operational status, and preset evaluation rules. For example, the preset evaluation rules may include at least one of the following: if the network status score is lower than a first preset threshold, the edge model is determined as the target model; if the edge model's operational status score is lower than a second preset threshold, the target model is determined from the cloud AI service cluster; if the target task's task complexity score is lower than a third preset threshold, the edge model is determined as the target model; if the terminal's operational status score is lower than a fourth preset threshold, the target model is determined from the cloud AI service cluster; if the task has strong real-time requirements, the edge model is determined as the target model. Of course, the preset evaluation rules are not limited to these.

[0049] The above-mentioned optional methods all involve determining a target model from either the edge model or the cloud AI service cluster.

[0050] As another alternative approach, this method can utilize different target models to execute different inference stages of the target task. Therefore, the method can include: determining multiple sub-tasks based on the target task and its running state, and determining the target model corresponding to each sub-task from at least one side—the edge model and the cloud AI service cluster—with each sub-task corresponding to a inference stage of the target task. For example, when the input data volume of the target task is large and the edge model is running well, the target task can be divided into a feature extraction task and an intent inference task. The edge model performs feature extraction on the input data of the target task, such as noise reduction, cropping, key information extraction, or intent classification. The target model in the cloud AI service cluster then performs intent inference based on the feature-extracted data to generate interactive intent information. This alternative approach uses a hybrid model approach from different sides to achieve inference. Utilizing a lightweight edge model to suppress noise and remove redundant information from the input data effectively reduces the feature dimension and sequence length of the input data, thereby reducing the computational overhead and response latency of the large cloud model. Furthermore, the large cloud model can focus its computational resources on contextual modeling and logical deduction of high-value semantic units, ultimately ensuring the accuracy and reliability of the inferred interactive intent information.

[0051] In this step, the target model is determined by considering factors such as network status, device status, and model capabilities. When the network is stable and high-speed and device resources are limited, cloud-based models are preferred to leverage their powerful computing capabilities to complete complex inferences and save local resources. When the network is poor, device computing power is sufficient, and the task has high real-time requirements, edge-side models are preferred to ensure low latency, offline availability, and data privacy and security.

[0052] In step 202, interactive intent information is generated based on the target task using at least one target model.

[0053] When there is only one target model, this method can complete the entire reasoning process of generating interactive intents based on the target task using a single target model. When there are multiple target models, each used to execute a subtask of the corresponding target task, this method can generate interactive intent information by having each target model perform its respective reasoning stage for its subtask.

[0054] The interaction intent information can be in JSON (JavaScript Object Notation) format, and can include the interaction intent name and the entity parameters on which the interaction intent is based. For example, if the target task is "Help me check the orders that are pending shipment today", the interaction intent information generated by the target model can be as follows: { "intent": "query_orders", "confidence": 0.96, "params": { "date": "2026-06-28", "status": "pending_delivery" } } The interaction intent information includes the interaction intent name "query_orders" and the entity parameters on which the interaction intent is based, namely today's date "2026-06-28" and the order status "pending_delivery".

[0055] In step 203, based on the interaction intent information, the component tree of the page to be generated is determined, and the components in the component tree are used to indicate the page structure of the page to be generated.

[0056] In this method, the target agent can determine what components are needed to realize the interaction intent and what response data the components need based on the interaction intent information and the preset component protocol, and generate a JSON component tree containing data binding placeholders. This JSON component tree can describe the page to be generated. The component protocol contains the names, attribute types, event interfaces and data structures of all available components. In other words, the component protocol includes a component list and component description information.

[0057] In step 204, the first page is displayed based on the component tree and the first response data of the target task.

[0058] This method can retrieve the first response data based on the data binding placeholders in the component tree, fill the corresponding data binding placeholders with the first response data, render the first page, and display the first page.

[0059] In this embodiment of the disclosure, the terminal's interactive operation can trigger the target task. The target intelligent agent can determine whether to use a model from either the cloud AI service cluster or the terminal model based on the specific situation and running status of the target task. For example, if the network is in good condition, complex tasks can be handled by the cloud model. The powerful generation capability of the cloud model ensures the quality and efficiency of the generated interactive intent information, thereby improving the page interaction efficiency. Handling simple tasks to the terminal model can achieve instant response and improve page interaction efficiency. Furthermore, a large number of simple tasks are carried by the terminal's computing power, which effectively reduces the frequency of calling the cloud model and reduces resource consumption.

[0060] Figure 3 This is a flowchart illustrating yet another page interaction method according to an exemplary embodiment, see [link to flowchart]. Figure 3 The method can be executed by a terminal, and the method includes the following steps: In step 301, in response to the interactive operation on the terminal, the target agent performs rule matching on the target task and multiple first rules based on the rule engine.

[0061] This method can use keywords, regular expressions, or semantic similarity to perform rule matching between the target task and multiple first rules. For example, a first rule can include a condition part and an execution part. The condition part includes preset matching features, and the execution part defines the response when the condition is met. For instance, if the target task is "Who are you?", and the matched first rule is "If it's an identity query, reply 'I am XX'", then "Who are you?" matches the condition part "If it's an identity query" in this first rule.

[0062] In step 302, if a match is successful, the second response data is obtained based on the first rule of the match, and the second page is displayed.

[0063] When a match is successful, meaning the target task hits any of the first rules, this method determines the second response data corresponding to the matched first rule. Based on this second response data, a second page is rendered and displayed. This second response data can be data from the execution portion of the matched first rule, such as "I am XX". Alternatively, the second response data can be data retrieved from the execution portion of the matched first rule. For example, if the matched first rule is "If querying personal balance, return account, params:{userId}", the second response data could be the user's account balance retrieved based on the userId identifier.

[0064] In step 303, if the matching is unsuccessful, the target agent determines at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running status corresponding to the interactive operation.

[0065] The method for determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running status corresponding to the interactive operation in this step is described in step 201 above.

[0066] In step 304, interactive intent information is generated based on the target task using at least one target model.

[0067] The specific implementation of this step is the same as step 202 above, and will not be repeated here.

[0068] In step 305, based on the interaction intent information, the component tree of the page to be generated is determined, and the components in the component tree are used to indicate the page structure of the page to be generated.

[0069] The specific implementation of this step is the same as step 203 above, and will not be repeated here.

[0070] In step 306, the first page is displayed based on the component tree and the first response data of the target task.

[0071] The specific implementation of this step is the same as step 204 above, and will not be repeated here.

[0072] In this embodiment, the terminal's interactive operation can trigger a target task. The target intelligent agent can perform rule matching according to the target task. If the rule matching is successful, the method can directly render and generate a second page based on the first response data. The model is only called when the rules are completely mismatched, which can significantly reduce the cost of calling high-performance computing power. The target intelligent agent can determine whether to use a model from the cloud AI service cluster or the terminal model based on the specific situation and running status of the target task. For example, when the network is in good condition, complex tasks can be handled by the cloud model. The powerful generation capability of the cloud model ensures the quality and efficiency of the generated interactive intent information, thereby improving the page interaction efficiency. Handling simple tasks to the terminal model can achieve instant response and improve page interaction efficiency. Moreover, a large number of simple tasks are carried by the terminal computing power, which effectively reduces the frequency of calling the cloud model and can reduce resource consumption.

[0073] In some embodiments, in addition to the target task and its running state, the method can also determine at least one target model from at least one side of the edge model and the cloud AI service cluster based on the context of the target task. Step 303 above, determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task corresponding to the interactive operation and its running state, can include: determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task corresponding to the interactive operation, the session context of the target task, and its running state. Specifically, the method can determine the task complexity of the target task based on the target task and the session context, and determine at least one target model from at least one side of the edge model and the cloud AI service cluster based on the task complexity and its running state. For example, the task complexity can be determined based on a preset task complexity rule, based on the target task and the session context. For example, the task complexity score can be calculated as: task complexity score = base weight × target task base score + incremental weight × session context score. For example, the target task might match the task complexity rule: the task contains "query" and "order" → medium complexity, with a complexity score of 0.5, and the target task's base score is 0.5. The target task's session context matches the task complexity rule "the token consumed by the context is in the first preset range → medium complexity, with a complexity score of 0.4". The session context score is 0.4, thus the task complexity score of the target task can be calculated. This disclosed embodiment can overcome the limitations of the surface semantics of a single target task, taking historical tasks into consideration, thereby achieving a three-dimensional assessment of the true difficulty of the target task.

[0074] Step 304 above, generating interaction intent information based on the target task using at least one target model, may include: generating interaction intent information based on the target task and session context using at least one target model. The session context may include task execution logs of historical tasks belonging to the same session ID as the target task. In this embodiment of the disclosure, the target model jointly generates interaction intent information based on the target task and session context. By introducing historical interaction trajectories, it significantly enhances the coherent understanding of complex intents and the ability for multi-step logical reasoning, reduces illusions, and ensures the consistency and accuracy of the output interaction intent information in terms of temporal logic.

[0075] If the target model call fails, the method can perform a fallback, switching from inference through the target model to rule matching for the target task. In some embodiments, the method may further include: if at least one target model call fails, performing rule matching between the target task and multiple second rules using a rule engine; if the match is successful, obtaining third response data based on the matched second rule and displaying a third page; if the target task does not match any second rule and at least one target model is an edge model, calling a model in a cloud AI service cluster to generate interactive intent information based on the target task; if the target task does not match any second rule and at least one target model is a model in a cloud AI service cluster, calling an edge model to generate interactive intent information based on the target task. The second rule is different from the first rule described above. The method can determine the second response data corresponding to the matched second rule when the target task successfully matches any second rule (i.e., the target task hits any second rule), render the third page based on the second response data, and switch to a cloud model to generate interactive intent information based on the target task when the edge model call fails and the target task fails to match the second rule. In cases where a model call in the cloud fails and the target task fails to match the second rule, the system switches to the edge model to generate interactive intent information based on the target task. In this embodiment, when a model call fails, rule matching is performed first, and the other model is only switched when the rules completely mismatch. This significantly reduces the high-frequency call cost of high-performance computing power, and switching the model when rules mismatch occurs minimizes the risk of the task failing to progress.

[0076] Figure 4 This is a schematic diagram illustrating a page interaction system according to an exemplary embodiment. See also: Figure 4 The page interaction system may include an end-to-cloud scheduler 401, a page description module 402, and a page generation module 403.

[0077] The edge-cloud scheduler 401 is used to respond to a terminal triggering a target task. Based on the target task and its running status, it determines at least one target model from at least one of the edge-side model and the cloud-based AI service cluster. It then generates interactive intent information based on the target task using the at least one target model. The running status indicates at least one of the terminal's running status, the edge-side model's running status, and the network status. The edge-cloud scheduler 401 can be used to determine whether to use a model from the edge-side or cloud-based AI service cluster. If the target model is determined to be from the cloud-based AI service cluster, it can invoke the model from the cloud-based AI service cluster through the cloud gateway.

[0078] The page description module 402 is used to determine the component tree of the page to be generated based on the interaction intent information. The components in the component tree are used to indicate the page structure of the page to be generated. The page description module 402 can be an A2UI (Agent-to-UI) protocol module. The page description module 402 is used to generate a JSON component tree conforming to the A2UI specification based on the interaction intent information and a preset component protocol. This JSON component tree is used to describe the page to be generated.

[0079] Page generation module 403 is used to display the first page based on the component tree and the first response data of the target task.

[0080] like Figure 5 The diagram illustrates the architecture of another page interaction system. This system may also include a component protocol module 404, which provides basic component protocols for fundamental components such as Button and Select, as well as scene component protocols for scene components such as DataCard and Chat. The page description module 402 can generate a JSON component tree conforming to the A2UI specification based on the interaction intent information and the component protocols provided by the component protocol module 404. This JSON component tree is used to describe the page to be generated.

[0081] In some embodiments, the page interaction system may further include an agent gateway 405. The agent gateway 405 may include a rule engine, which can be used to perform rule matching between the target task and multiple first rules in response to a target task triggered by a terminal.

[0082] The page generation module 403 is used to obtain second response data based on the matching rules and display the second page when the target task successfully matches multiple first rules.

[0083] The edge-cloud scheduler 401 is used to determine at least one target model from at least one side of the edge model and the cloud AI service cluster, based on the target task corresponding to the interactive operation and the running status, when the target task fails to match multiple first rules.

[0084] In some embodiments, the edge-cloud scheduler 401 can be used to: determine the evaluation score of the edge model based on the task complexity, running status, and edge evaluation strategy of the target task corresponding to the interactive operation; determine the evaluation score of the cloud AI service cluster based on the task complexity, running status, and cloud evaluation strategy of the target task corresponding to the interactive operation; if the evaluation score of the edge model is higher than the evaluation score of the cloud AI service cluster, determine the edge model as the target model; if the evaluation score of the cloud AI service cluster is higher than the evaluation score of the edge model, determine the target model from the cloud AI service cluster.

[0085] In some embodiments, the agent gateway 405 can be used to query the session context associated with the target task in memory based on the current terminal window in response to the terminal triggering the target task. The agent gateway 405 can query the session context with the same session identifier as the target task from memory based on the session identifier of the current terminal window corresponding to the target task; that is, query the session context belonging to the same terminal window as the target task. Alternatively, the session identifier of the target task can also be the session identifier of the target task's current tab. The agent gateway 405 can then query the session context with the same session identifier as the target task from memory based on the session identifier of the target task's tab in response to the terminal triggering the target task; that is, query the session context belonging to the same tab as the target task.

[0086] In some embodiments, the agent gateway 405 can be used to write the task execution log of the target task into memory to update the session context. The agent gateway 405 can write the execution log into the context storage space associated with the session identifier in memory, based on the session identifier corresponding to the target task. This session context is stored in memory in independent partitions according to the corresponding session identifier. This embodiment enables multi-session isolation, avoiding data crosstalk between different tasks in concurrent scenarios, ensuring that each target task can only access the session context information of its own session, effectively improving data security and logical correctness. Furthermore, the agent gateway 405 can be used to orchestrate the tasks of each stage of the target task decomposed by the agent, sending each stage task to the corresponding tool, such as sending the target task to the end-cloud scheduler 401 to trigger the end-cloud scheduler 401 to obtain interaction intent information. The agent gateway 405 can also be responsible for the lifecycle management of tools and skills.

[0087] In some embodiments, the edge-cloud scheduler 401 can be used to determine at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task corresponding to the interactive operation, the session context of the target task, and the running state; and generate interactive intent information based on the target task and the session context through at least one target model.

[0088] In some embodiments, the edge-cloud scheduler 401 can be used to determine multiple sub-tasks based on the target task and the running status, and to determine the target model corresponding to each sub-task from at least one side of the edge model and the cloud AI service cluster, with each sub-task corresponding to an inference stage of the target task; and to generate interactive intent information by executing the corresponding inference stage for each sub-task through each target model.

[0089] In some embodiments, the agent gateway 405 can be used to perform rule matching between the target task and multiple second rules through a rule engine in the event that at least one target model call fails.

[0090] The page generation module 403 is used to obtain third response data based on the matched second rules and display the third page when the target task and multiple second rules are successfully matched.

[0091] The edge-cloud scheduler 401 can be used to, when the target task does not match any of the second rules and at least one target model is an edge model, call the model in the cloud AI service cluster to generate interactive intent information based on the target task; and when the target task does not match any of the second rules and at least one target model is a model in the cloud AI service cluster, call the edge model to generate interactive intent information based on the target task.

[0092] In some embodiments, the page interaction system may include a terminal platform 406 for end users or tenants, such as a merchant platform or content creator platform. This terminal platform 406 can be used to integrate and provide various services, such as product services, order services, store services, and content services. The terminal platform 406 supports multiple operating environments, including PC Web, PC desktop applications, and mobile clients (such as Apps, Applications).

[0093] In some embodiments, the page interaction system may include an SDK (Software Development Kit) module 407. The SDK module 407 supports multiple interface calls, such as the inference interface inference.run(), the A2UI interface a2ui.generate(), the Agent interface agent.run(), and the tool management interface agent.registerTool(). For example, agent.run() can encapsulate the capabilities of inference.run() and a2ui.generate(). The terminal platform 406 can respond to the terminal's interactive operation by calling the agent execution interface in the SDK module 407, namely the agent.run() interface. Thus, the SDK module 407 can hand over the target task corresponding to the interactive operation to the agent gateway 405 for orchestration. During the process of the agent processing the target task, it calls the inference engine module 410 for inference and calls the page description module 402, namely the A2UI module, to generate a JSON component tree. `inference.run()` and `a2ui.generate()` can also be called independently. For example, terminal platform 406 can respond to text input operations on the terminal and call `inference.run()` in SDK module 407 to complete the response to the input text. SDK module 407 is responsible for monitoring and reporting observable data, and can monitor and report performance indicators, abnormal events, and task execution chains. SDK module 407 can also receive configuration information issued by management platform 408. This configuration information may include configuration information for the edge model, first rule, second rule, scheduling strategy, etc. For example, the page interaction system may also include intelligent agent gateway 405, which contains a rule engine. SDK module 407 can pass the first rule, second rule, etc. to the rule engine and pass the scheduling strategy to the edge-cloud scheduler 401. This scheduling strategy may include edge evaluation strategy and cloud evaluation strategy, etc. SDK module 407 also supports hot updates of configuration information.

[0094] In some embodiments, the system may further include a management platform 408, which may be used to perform at least one of the following: detect whether a tenant or scenario has interface call permissions; if so, generate an interface function opening instruction; if not, generate an interface function closing instruction; respond to a model release instruction, send a new model to a tenant belonging to a grayscale group or an experimental group according to a preset grayscale ratio or experimental group identifier; collect data corresponding to alarm indicators, detect whether the data corresponding to the alarm indicators meets the corresponding alarm conditions; if so, generate an alarm event, wherein the alarm indicators include at least one of response latency, task execution success rate, and resource utilization rate during task execution; respond to a query request based on a target link identifier, determine the target query task corresponding to the target link identifier, and generate a waterfall chart of the target query task based on the task execution log of the target query task; the waterfall chart displays the task processing stages of the target query task on the vertical axis and the time consumed by each task processing stage on the horizontal axis; detect whether the data in the task execution log meets the risk conditions; if so, generate a risk warning event; calculate resource consumption according to resource statistics dimensions, wherein the resource statistics dimensions include at least one of tenant, model, and scenario; and provide debugging functions based on the end-side operating environment.

[0095] The management platform uses a 408 detection task execution log to check if the data meets risk conditions for security auditing. The debugging tool is designed for developers and is used to perform at least one of the following: simulating different terminal devices and operating scenarios to allow developers to test the performance of their agent applications on various platforms within the development environment; time travel debugging to record the application's state change history, allowing developers to revert to any historical state for problem localization; and displaying at least one performance indicator from rendering performance, memory usage, CPU usage, network performance, and inference performance through a performance dashboard to help developers identify performance bottlenecks.

[0096] The management platform 408 can also issue model preheating instructions for models to be preheated. These instructions can include the model's identifier and preheating timing information. The model management module, deployed in the cloud, receives and parses these instructions from the management platform 408. Based on the model identifier and target tenant information in the instructions, it distributes the model preheating task to the SDK module 407 integrated into the corresponding edge device. The SDK module 407, upon receiving the preheating task and the timing of the preheating, loads the model from local storage into memory to complete the model preheating. As a unified control hub, the management platform 408 is responsible for access control, model publishing and preheating, monitoring and alarms, link tracing, risk detection, resource statistics, and edge debugging, enabling visibility, manageability, and controllability of both the edge and cloud sides.

[0097] In some embodiments, the system may further include an agent framework module 409.

[0098] The agent framework module 409 can be used to, in response to an agent invocation instruction for a target agent, determine the second adapter corresponding to the target agent, translate the agent invocation instruction into an invocation instruction of the target agent framework on which the target agent depends through the second adapter, and drive the target agent framework to run the target agent.

[0099] Invocation instructions targeting a target intelligent agent can be task execution instructions, tool invocation instructions, or streaming output instructions. The intelligent agent framework module 409 can integrate intelligent agent frameworks such as OpenClaw and HermasClaw. For example, the second adapter can convert the `tool` involved in the tool invocation instructions targeting the target intelligent agent into a `tool` definition recognizable by the target intelligent agent framework, translating the tool invocation instructions of the target intelligent agent into invocation instructions recognizable by the target intelligent agent framework. The intelligent agent framework module 409 can also translate the raw output data of the target intelligent agent framework into a preset data format.

[0100] In some embodiments, the system may also include an inference engine module 410.

[0101] The inference engine module 410 is used to respond to the model invocation command for the target model, determine the first adapter corresponding to the target model, translate the model invocation command into the invocation command of the target inference engine on which the target model depends through the first adapter, and drive the target inference engine to run the target model.

[0102] The model invocation instruction for the target model can be a model loading instruction, a model inference instruction, or a model resource release instruction. The inference engine module 410 can integrate inference engines such as WebLLM and ONNX.js. The inference engine module 409, in response to the model loading instruction for the target model, creates the WebGPU (Web Graphics Processing Unit) context and Worker thread corresponding to the target inference engine that the target model depends on through the first adapter, and then translates the model invocation instruction into the invocation instruction of the target inference engine that the target model depends on through the first adapter. The inference engine module 410, in response to the model resource release instruction for the target model, destroys the WebGPU context and Worker thread corresponding to the target inference engine that the target model depends on through the first adapter.

[0103] In some embodiments, the inference engine module 410 may include a post-processor. The post-processor can be used to validate the inference results generated by the target model. If validation fails, a validation failure event is reported; if validation succeeds, the inference results are corrected to generate interactive intent information, and an inference completion event is reported. For example, the post-processor can perform schema validation on the inference results. Correcting the inference results can refer to at least one of the following: desensitizing sensitive fields, correcting numerical ranges, and filling in default values. This ensures the security, compliance, and data quality of the interactive intent information. Automatic desensitization of sensitive information effectively avoids the risk of data privacy leakage. Numerical correction of outliers exceeding reasonable business boundaries and filling in default values ​​for missing fields ensure the data integrity of the interactive intent information.

[0104] In some embodiments, the system may include an event bus 411. Verification failure events or inference completion events reported by the post-processor can be reported to the event bus 411. The event bus 411 can then report the events to the management platform 408.

[0105] In some embodiments, the system may include an inference history storage module 412, which can be used to subscribe to inference events on the event bus and store the inference events. The inference events may include at least one of inference completion events and verification failure events. By decoupling the inference of the inference engine module 410 from the event storage of the inference history storage module, the main inference chain is no longer affected by the time spent writing to storage, significantly reducing inference latency and improving system throughput. Failure, expansion, or upgrade maintenance of the inference history storage module will not affect the availability of the inference engine, thereby enhancing the overall resilience of the system.

[0106] In some embodiments, the system may further include a model management module 413. The model management module can be used to: preload a new model in response to detecting a model update event for a new model; switch the service pointer from the old model to the new model through an atomic operation after the new model is loaded; mark the old model as drained, wait for the requests being processed on the old model to finish, and unload the old model. Wherein, after the new model is loaded, its state can be marked as ready. This embodiment of the disclosure isolates the time-consuming operations of model loading and verification from the model service traffic through preloading and atomic pointer switching, compressing the traffic switching action itself to the nanosecond level, completely eliminating jitter or blockage during the switching process. Draining further ensures data consistency, ensuring that the old model is not forcibly reclaimed before all legacy requests are completed, avoiding request interruption or state loss. The hash comparison and digital signature dual verification after model loading eliminates the risk of model file corruption or untrusted sources from the source, making the entire update process highly available, data-integrity, and operational atomicity. In addition, the model management module 413 can also be used to perform model pre-updates and differential updates. For example, performing a model pre-update involves pushing the model file of the pre-updated model to a specified storage directory on the terminal side when the terminal device is detected to be in an idle state. Performing a differential update involves extracting the differential data packets of the new model and the old model in response to a model update event and sending the differential data packets.

[0107] In some embodiments, the model management module 413 can be used to mark an old model as being in a drained state, and then uninstall the old model if the processing time of the old model for requests meets a preset timeout condition. This can prevent the uninstallation process of the old model from being blocked due to prolonged blocking by individual abnormal requests.

[0108] In some embodiments, the model management module 413 can be used to switch the service pointer from the new model back to the old model via an atomic operation in response to a rollback command. This embodiment of the disclosure switches the service pointer from the new model back to the old model via an atomic operation, reducing the complex rollback process to nanosecond-level pointer assignment and completely avoiding the time-consuming process of reloading the model.

[0109] In some embodiments, the model management module 413 can be used to: in response to the completion of downloading the main model resource package obtained from the primary content delivery network (CDN), perform a first integrity check on the main model resource package; if the first integrity check passes, load the main model resource package into memory; if the first integrity check fails, obtain a backup model resource package from the backup content delivery network and report the model resource check failure event. The first integrity check on the main model resource package can be performed by performing digital signature verification and SHA-256 hash comparison. Digital signature verification can refer to decrypting and verifying the digital signature in the manifest file corresponding to the main model resource package using a preset public key to verify the authenticity and legality of the manifest file's source, preventing malicious tampering or forgery. SHA-256 hash comparison can refer to comparing the hash value of the main model resource package with the value recorded in the manifest file corresponding to the main model resource package to verify whether the data is corrupted or tampered with.

[0110] In some embodiments, the system may further include a component management module 414, which can be used to: perform a second integrity check on the main component resource package in response to the completion of downloading the main component resource package obtained from the main content delivery network; load the main component resource package into memory if the second integrity check passes; and obtain a backup component resource package of the main component resource package from a backup content delivery network and report the component resource check failure event if the second integrity check fails. The second integrity check on the component resource package may involve digital signature verification and SHA-256 hash comparison of the main component resource package. Digital signature verification of the main component resource package may refer to decrypting and verifying the digital signature in the manifest file corresponding to the main component resource package using a preset public key to verify the authenticity and legality of the manifest file and prevent it from being maliciously tampered with or forged. SHA-256 hash comparison of the main component resource package may refer to comparing the hash value of the main component resource package with the value recorded in the manifest file corresponding to the main component resource package to verify whether the data is corrupted or tampered with. In addition, the component management module 414 is also used to parse the component tree and extract the component type corresponding to each node; for the target node in the component tree, it checks whether the component type corresponding to the target node has been loaded; if it has been loaded, it verifies the version information of the component, and if the version is not the latest, it performs an update; if it has not been loaded, it dynamically loads the corresponding component code according to the component type. The component management module 414 is also used for version management, recording the version information of each component on the terminal device, performing component updates when a new version is released in the cloud, and performing version rollback when a component anomaly is detected.

[0111] In some embodiments, the system may further include a platform abstraction module 415, which is configured to perform at least one of the following: in response to a call request to the StorageFactory interface, execute a storage instance corresponding to the runtime environment; in response to a call request to the WorkerFactory interface, execute a background task instance corresponding to the runtime environment; in response to a call request to the EventBusFactory interface, execute an event communication instance corresponding to the runtime environment. The runtime environment refers to the runtime environment of the terminal platform 406, such as a PC Web application, a PC desktop application, or a mobile client (e.g., an App). For example, in response to a call request to the StorageFactory interface, execute an IndexedDB instance corresponding to the PC Web application; in response to a call request to the StorageFactory interface, execute a local file system instance corresponding to the PC desktop application; and in response to a call request to the StorageFactory interface, execute a native storage instance corresponding to the mobile client. In response to a call request to the background task factory interface, the Web Worker instance corresponding to the computer webpage is executed; in response to a call request to the background task factory interface, the child process instance corresponding to the computer desktop program is executed; in response to a call request to the background task factory interface, the Web Worker instance or native process instance corresponding to the mobile client is executed. This disclosed embodiment enables upper-layer applications to call a unified interface without being aware of environmental differences, significantly reducing the complexity and maintenance costs of cross-platform development; centralized management of instance creation and switching through the factory pattern facilitates on-demand expansion to new platforms, replacement of specific implementations, or environmental isolation, improving the system's scalability and testability.

[0112] In some embodiments, the system further includes a hardware adaptation module 416, which can be used to perform at least one of the following: detecting at least one hardware performance parameter among the terminal device's computing power, memory capacity, and storage space; detecting at least one network status information among the terminal device's network type and network quality; detecting the inference backend supported by the terminal device and determining the available inference engine; and obtaining at least one platform identification information among the terminal device's operating system type, device model, and application environment.

[0113] In some embodiments, the system may further include a resource scheduler 417, which can be used to determine the priorities of multiple tasks to be executed and schedule the computing resources occupied by the tasks in descending order of priority. It can also be used to configure resource quotas for tenants or tasks and limit the maximum resource usage of the corresponding tenants or tasks according to the resource quotas. The resource quotas include at least one of memory quotas, CPU utilization quotas, and GPU memory quotas. Furthermore, it can be used to create and manage a Worker pool, which contains one or more Workers, each Worker corresponding to an execution thread or process for executing the scheduled tasks.

[0114] In some embodiments, the system may further include a security module 418, which can be used to desensitize the input information corresponding to the target task. For example, if an object enters "I want to check the recent orders" on the platform's smart conversation page and clicks the send control, the corresponding target task is the task of "checking the recent orders," and the input information of the target task is "I want to check the recent orders." Sensitive information includes at least one of user privacy data, identity authentication information, and confidential data. Desensitization processing includes at least one of replacement, obfuscation, encryption, and truncation. It can also be used to isolate the data storage, computing resources, and access permissions of different tenants based on tenant identifiers to prevent unauthorized access across tenants.

[0115] In some embodiments, the system may further include an offline synchronization manager 419, which monitors the network status of the terminal device and determines whether the current network meets the data transmission conditions based on the network status. The offline synchronization manager can also be used to persistently store the target task in a local task queue when the target task cannot be executed offline due to its reliance on a cloud-based AI service cluster, and retrieve the target task according to the queue order when the network status meets the data transmission conditions.

[0116] In some embodiments, the system may further include an agent capability management module 420, which manages the registration, review, and publication of all available tools, skills, and knowledge bases.

[0117] The system's end-to-end cloud scheduler 401, page description module 402, page generation module 403, component protocol module 404, agent gateway 405, terminal platform 406, SDK module 407, agent framework module 409, inference engine module 410, event bus 411, inference history storage module 412, platform abstraction module 415, hardware adaptation module 416, resource scheduler 417, security module 418, and offline synchronization manager 419 can all be deployed on the terminal, while the management platform 408, model management module 413, component management module 414, agent capability management module 420, event bus 411, and inference history storage module 412 can all be deployed in the cloud.

[0118] like Figure 6 A data flow diagram of page interaction is shown.

[0119] Terminal platform 406 calls the configuration request interface in SDK module 407 to send a configuration retrieval request to management platform 408. This configuration retrieval request may carry the tenant's tenant identifier. SDK module 407 can receive configuration information from management platform 408 in response to this configuration retrieval request. This configuration information may include configuration information for the terminal-side model, first rule, second rule, and scheduling strategy. The scheduling strategy may include terminal-side evaluation strategy and cloud-based evaluation strategy.

[0120] The terminal platform can respond to the terminal's interactive operations by calling the agent execution interface provided by the SDK module 407, namely the agent.run() interface, to drive the target agent. The SDK module 407 can hand over the target task corresponding to the interactive operation to the agent gateway 405 of the target agent for orchestration. The agent gateway 405 performs rule matching between the target task and multiple first rules through the rule engine. If the target task successfully matches multiple first rules, the agent gateway 405 returns the second response data corresponding to the matched first rule to the SDK module 407. The page generation module 403 receives the second response data returned by the SDK module 407, renders it based on the second response data to generate a second page, and displays the second page on the terminal platform. The second response data returned by the SDK module 407 is also the second response data returned by the target agent.

[0121] If the target task fails to match multiple first rules, the intelligent agent gateway 405 queries the memory for the session context associated with the target task based on the current terminal window, and sends the target task and session context to the edge-cloud scheduler 401. The edge-cloud scheduler 401 can determine at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task corresponding to the interaction operation, the session context of the target task, and the running status; and generate interaction intent information based on the target task and session context using at least one target model.

[0122] If at least one target model call fails, the edge-cloud scheduler 401 sends a downgrade instruction for the target task to the agent gateway 405. Based on this downgrade instruction, the agent gateway 405 uses a rule engine to perform rule matching between the target task and multiple second rules. If a match is successful, the third response data corresponding to the matched second rule is returned to the SDK module 407. The page generation module 403 receives the third response data returned by the SDK module 407, renders it based on the third response data to generate a third page, and displays the third page on the terminal platform. If the target task does not match any of the second rules, the agent gateway 405 sends an inference instruction for the target task to the edge-cloud scheduler 401. If at least one target model is an edge-side model, the edge-cloud scheduler 401 calls a model in the cloud AI service cluster to generate interactive intent information based on the target task; if at least one target model is a model in the cloud AI service cluster, the edge-cloud scheduler 401 calls an edge-side model to generate interactive intent information based on the target task.

[0123] The edge-cloud scheduler 401 can return the interaction intent information to the intelligent agent gateway 405. The intelligent agent gateway 405 sends the interaction intent information to the page description module 402. The page description module 402 can determine the component tree of the page to be generated based on the interaction intent information and the component protocol of the component protocol module 404, and return the component tree to the intelligent agent gateway 405. The intelligent agent gateway 405 returns the component tree to the SDK module 407.

[0124] The page generation module 403 can receive the component tree returned by the SDK module 407, which is also the component tree returned by the target intelligent agent. The page generation module 403 sends the component tree to the scene adapter and calls the scene adapter. The scene adapter calls the scene interface that matches the scene of the component tree, processes the data returned by the scene interface to obtain the first response data, and fills the first response data into the position of the corresponding data binding placeholder to generate a complete component tree. The page generation module 403 can receive the complete component tree returned by the scene adapter, render the complete component tree, and obtain the first page.

[0125] Computer equipment can be implemented as a terminal. Figure 7 This is a block diagram illustrating a terminal 700 according to an exemplary embodiment. Figure 7 A structural block diagram of a terminal 700 provided in an exemplary embodiment of this disclosure is shown. The terminal 700 may be a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. The terminal 700 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names.

[0126] Typically, terminal 700 includes a processor 701 and a memory 702.

[0127] Processor 701 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 701 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 701 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 701 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, processor 701 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0128] The memory 702 may include one or more computer-readable storage media, which may be non-transitory. The memory 702 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 702 are used to store at least one computer program, which is executed by the processor 701 to implement the page interaction method provided in the method embodiments of this application.

[0129] In some embodiments, the terminal 700 may also optionally include a peripheral device interface 703 and at least one peripheral device. The processor 701, memory 702, and peripheral device interface 703 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 703 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 704, a display screen 705, a camera assembly 706, an audio circuit 707, and a power supply 708.

[0130] Peripheral device interface 703 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 701 and memory 702. In some embodiments, processor 701, memory 702 and peripheral device interface 703 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 701, memory 702 and peripheral device interface 703 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0131] The radio frequency (RF) circuit 704 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 704 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 704 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. In some embodiments, the RF circuit 704 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 704 can communicate with other terminals through at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 704 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0132] Display screen 705 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 705 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 701 for processing. In this case, display screen 705 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 705, disposed on the front panel of terminal 700; in other embodiments, there may be at least two display screens, disposed on different surfaces of terminal 700 or in a folded design; in other embodiments, display screen 705 may be a flexible display screen, disposed on a curved or folded surface of terminal 700. Furthermore, display screen 705 may be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. Display screen 705 may be made of materials such as LCD (Liquid Crystal Display) or OLED (Organic Light-Emitting Diode).

[0133] The camera assembly 706 is used to acquire images or videos. In some embodiments, the camera assembly 706 includes a front-facing camera and a rear-facing camera. Typically, the front-facing camera is located on the front panel of the terminal, and the rear-facing camera is located on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 706 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash is a combination of a warm-light flash and a cool-light flash, which can be used for light compensation at different color temperatures.

[0134] The audio circuit 707 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 701 for processing, or input to the radio frequency circuit 704 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each located at a different part of the terminal 700. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert the electrical signals from the processor 701 or the radio frequency circuit 704 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 707 may also include a headphone jack.

[0135] Power supply 708 is used to power the various components in terminal 700. Power supply 708 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 708 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged via a wired line, while a wireless rechargeable battery is a battery that is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0136] Those skilled in the art will understand that Figure 7 The structure shown does not constitute a limitation on terminal 700, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0137] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory including instructions, which can be executed by a processor of a computer device, such as a terminal, to complete the page interaction method described above. Optionally, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0138] This disclosure also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the above-described page interaction method.

[0139] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0140] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A page interaction method, characterized in that, The method includes: In response to interactive operations on the terminal, the target agent determines at least one target model from at least one side of the terminal model and the cloud AI service cluster based on the target task and running status corresponding to the interactive operation. The running status indicates at least one of the terminal running status, the terminal model running status, and the network status. Interactive intent information is generated based on the target task using at least one target model; Based on the interaction intent information, a component tree of the page to be generated is determined, wherein the components in the component tree are used to indicate the page structure of the page to be generated; Based on the component tree and the first response data of the target task, the first page is displayed.

2. The page interaction method according to claim 1, characterized in that, The method further includes: In response to interactive operations on the terminal, the target intelligent agent performs rule matching between the target task and multiple first rules based on a rule engine; If a match is successful, retrieve the second response data based on the first rule of the match and display the second page; If the matching fails, the following step is performed: the target agent determines at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running status corresponding to the interaction operation.

3. The page interaction method according to claim 1, characterized in that, The step of determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running status corresponding to the interaction operation by the target intelligent agent includes: The evaluation score of the edge model is determined by the target agent based on the task complexity, running status and edge evaluation strategy of the target task corresponding to the interaction operation. Based on the task complexity, running status, and cloud evaluation strategy of the target task corresponding to the interactive operation, the evaluation score of the cloud AI service cluster is determined. If the evaluation score of the edge model is higher than the evaluation score of the cloud AI service cluster, the edge model is determined as the target model; if the evaluation score of the cloud AI service cluster is higher than the evaluation score of the edge model, the target model is determined from the cloud AI service cluster.

4. The page interaction method according to claim 1, characterized in that, The step of determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running state corresponding to the interaction operation by the target intelligent agent includes: Based on the target task corresponding to the interaction operation, the session context of the target task, and the running state, the target agent determines at least one target model from at least one side of the edge model and the cloud AI service cluster. The step of generating interactive intent information based on the target task using the at least one target model includes: Based on the target task and the session context, the at least one target model generates interaction intent information.

5. The page interaction method according to claim 1, characterized in that, The step of determining at least one target model from at least one side of the edge model and the cloud AI service cluster based on the target task and running state corresponding to the interaction operation by the target intelligent agent includes: Based on the target task and its running state, the target agent determines multiple sub-tasks and the target model corresponding to each sub-task is determined from at least one side, the edge model and the cloud AI service cluster. Each sub-task corresponds to an inference stage of the target task. The step of generating interactive intent information based on the target task using the at least one target model includes: Each target model performs a corresponding inference phase for its respective sub-task to generate interactive intent information.

6. The page interaction method according to claim 1, characterized in that, The method further includes: In the event that at least one target model call fails, the target agent performs rule matching on the target task and multiple second rules based on a rule engine. If a match is successful, retrieve the third response data based on the second rule of the match and display the third page; If the target task does not match any of the second rules and the at least one target model is an edge model, the target agent calls the model in the cloud AI service cluster to generate interactive intent information based on the target task. If the target task does not match any of the second rules and the at least one target model is a model in a cloud AI service cluster, the target agent calls the edge model to generate interactive intent information based on the target task.

7. A page interaction system, characterized in that, This includes an edge-cloud scheduler, a page description module, and a page generation module; The edge-cloud scheduler is used to respond to a terminal triggering a target task, and based on the target task and the running status, determine at least one target model from at least one side of the edge-side model and the cloud AI service cluster, and generate interactive intent information based on the target task through the at least one target model. The running status indicates at least one of the terminal running status, the edge-side model running status and the network status. The page description module is used to determine the component tree of the page to be generated based on the interaction intent information, wherein the components in the component tree are used to indicate the page structure of the page to be generated; The page generation module is used to display a first page based on the component tree and the first response data of the target task.

8. The page interaction system according to claim 7, characterized in that, The system also includes an intelligent agent gateway, which is used to respond to a terminal triggering a target task and perform rule matching between the target task and multiple first rules through a rule engine; The page generation module is used to obtain second response data based on the matching rules and display a second page when the target task successfully matches the plurality of first rules. The edge-cloud scheduler is used to determine at least one target model from at least one side of the edge-side model and the cloud AI service cluster, based on the target task corresponding to the interaction operation and the running status, when the target task fails to match the plurality of first rules.

9. The page interaction system according to claim 7, characterized in that, The edge-cloud scheduler is used for: The evaluation score of the terminal model is determined based on the task complexity, running status, and terminal evaluation strategy of the target task corresponding to the interactive operation. Based on the task complexity, running status, and cloud evaluation strategy of the target task corresponding to the interactive operation, the evaluation score of the cloud AI service cluster is determined. If the evaluation score of the edge model is higher than the evaluation score of the cloud AI service cluster, the edge model is determined as the target model; if the evaluation score of the cloud AI service cluster is higher than the evaluation score of the edge model, the target model is determined from the cloud AI service cluster.

10. The page interaction system according to claim 7, characterized in that, The edge-cloud scheduler is used for: Based on the target task corresponding to the interactive operation, the session context of the target task, and the running state, at least one target model is determined from at least one side of the terminal model and the cloud AI service cluster. Based on the target task and the session context, the at least one target model generates interaction intent information.

11. The page interaction system according to claim 7, characterized in that, The edge-cloud scheduler is used for: Based on the target task and the running status, multiple sub-tasks are determined, and the target model corresponding to each sub-task is determined from at least one side, the edge model and the cloud AI service cluster. Each sub-task corresponds to an inference stage of the target task. Each target model performs a corresponding inference phase for its respective sub-task to generate interactive intent information.

12. The page interaction system according to claim 7, characterized in that, The intelligent agent gateway is used to perform rule matching between the target task and multiple second rules through a rule engine when the at least one target model call fails. The page generation module is used to obtain third response data based on the matched second rules and display a third page when the target task and multiple second rules are successfully matched. The edge-cloud scheduler is used to, when the target task does not match any of the second rules and the at least one target model is an edge model, invoke a model in the cloud AI service cluster to generate interactive intent information based on the target task; and when the target task does not match any of the second rules and the at least one target model is a model in the cloud AI service cluster, invoke the edge model to generate interactive intent information based on the target task.

13. The page interaction system according to claim 7, characterized in that, The system also includes an inference engine module; The inference engine module is configured to respond to a model invocation command for a target model, determine a first adapter corresponding to the target model, translate the model invocation command into an invocation command for the target inference engine on which the target model depends through the first adapter, and drive the target inference engine to run the target model.

14. The page interaction system according to claim 13, characterized in that, The inference engine module includes a post-processor; The post-processor is used to verify the inference results generated by the target model. If the verification fails, a verification failure event is reported. If the verification is successful, the reasoning result is corrected to generate interactive intent information, and the reasoning completion event is reported.

15. The page interaction system according to claim 7, characterized in that, The system also includes an intelligent agent framework module; The agent framework module is configured to respond to an agent invocation command for a target agent, determine a second adapter corresponding to the target agent, translate the agent invocation command into an invocation command of the target agent framework on which the target agent depends through the second adapter, and drive the target agent framework to run the target agent.

16. The page interaction system according to claim 7, characterized in that, The system also includes an intelligent agent gateway. The intelligent agent gateway is used to respond to a terminal triggering a target task by querying the session context associated with the target task in memory based on the current terminal window.

17. The page interaction system according to claim 16, characterized in that, The intelligent agent gateway is also used for: The task execution log of the target task is written to the memory to update the session context, and the session context is stored in memory in independent partitions according to the corresponding session identifier.

18. The page interaction system according to claim 7, characterized in that, The system also includes a model management module, which is used for: In response to a model update event that detects a new model, the new model is preloaded; After the new model is loaded, the service pointer is switched from the old model to the new model through an atomic operation; The old model is marked as drained, and the system waits for the requests being processed on the old model to be completed before unloading the old model.

19. The page interaction system according to claim 18, characterized in that, The model management module is used to mark the old model as being in a drained state, and then, if the processing time of the old model for the request meets a preset timeout condition, unload the old model.

20. The page interaction system according to claim 18, characterized in that, The model management module is used for: In response to a rollback instruction, the service pointer is switched back from the new model to the old model via an atomic operation.

21. The page interaction system according to claim 18, characterized in that, The model management module is used for: In response to the completion of downloading the main model resource package obtained from the main content delivery network, a first integrity check is performed on the main model resource package; if the first integrity check passes, the main model resource package is loaded into memory; if the first integrity check fails, a backup model resource package of the main model resource package is obtained from the backup content delivery network, and the model resource check failure event is reported.

22. The page interaction system according to claim 7, characterized in that, The system also includes a component management module, which is used to perform a second integrity check on the main component resource package in response to the completion of downloading the main component resource package obtained from the main content delivery network; if the second integrity check passes, load the main component resource package into memory; if the second integrity check fails, obtain a backup component resource package of the main component resource package from the backup content delivery network, and report the component resource check failure event.

23. The page interaction system according to claim 7, characterized in that, The system also includes a platform abstraction module, which is configured to perform at least one of the following: In response to a call request to the storage factory interface, execute the storage instance corresponding to the runtime environment; In response to a call request to the background task factory interface, execute the background task instance corresponding to the runtime environment; In response to a call request to the event bus factory interface, execute the event communication instance corresponding to the runtime environment.

24. The page interaction system according to claim 7, characterized in that, The system also includes a management platform, which is used to perform at least one of the following: Check whether the tenant or scenario has the permission to call the interface. If yes, generate an instruction to open the interface function; otherwise, generate an instruction to close the interface function. In response to the model release command, the new model is sent to the tenant that belongs to the grayscale group or the experimental group according to the preset grayscale ratio or experimental group identifier. Collect data corresponding to alarm indicators, and check whether the data corresponding to the alarm indicators meets the corresponding alarm conditions. If so, generate an alarm event. The alarm indicators include at least one of the following: response latency during task execution, task execution success rate, and resource utilization rate. In response to a query request based on a target link identifier, the target query task corresponding to the target link identifier is determined, and a waterfall chart of the target query task is generated based on the task execution log of the target query task. The waterfall chart displays the task processing stages of the target query task on the vertical axis and the time consumed by each task processing stage on the horizontal axis. Check whether the data in the task execution log meets the risk conditions; if so, generate a risk warning event. Resource consumption is calculated according to resource statistics dimensions, which include at least one of tenant, model, and scenario. Provides debugging functionality based on the edge runtime environment.

25. A computer device, characterized in that, The computer device includes: One or more processors; Memory used to store the executable program code of the processor; The processor is configured to execute the program code to implement the page interaction method as described in any one of claims 1 to 6.

26. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of a computer device, the computer device is enabled to perform the page interaction method as described in any one of claims 1 to 6.

27. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the page interaction method according to any one of claims 1 to 6.

Images