Multi-agent based dialogue content determination method and device, computer device, and storage medium

Abstract

The application provides a multi-agent-based dialogue content determination method and device, computer equipment and a storage medium. After determining the current session content, the main agent extracts labels from the current session content to generate a current intent vector. Based on the current intent vector and the ability matrix data corresponding to each sub-agent, a target sub-agent is determined. The target sub-agent determines meta-output data according to the received meta-information. The meta-output data corresponding to the target sub-agent is taken as a node, and the conflict type between different nodes is taken as a connection edge to construct a conflict association graph. According to the conflict type of the connection edge, the severity of the conflict and the arbitration data, the conflict arbitration sub-agent is controlled to perform conflict resolution processing on the conflict association graph to generate a processed association graph. According to the processed association graph and the business priority order, a target component tree is constructed. Based on the target component tree, the reply content corresponding to the current session content is generated and displayed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, computer device, and storage medium for determining dialogue content based on multiple agents. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent dialogue systems have been widely applied in various fields such as customer service, finance, e-commerce, and automobiles. Among related technologies, intelligent dialogue systems typically employ a pipelined architecture combining intent recognition, slot filling, and dialogue management. In practice, the intelligent dialogue system first identifies the user's intent and extracts key slot information through a natural language understanding module. Then, the dialogue management module determines the next action based on a predefined dialogue flow. Finally, the natural language generation module outputs the response content. However, because this intelligent dialogue system relies on intent recognition and dialogue orchestration, it suffers from high complexity, rigid business logic, poor scalability, stiff personalized interaction, and limitations in semantic understanding. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, computer device, and storage medium for determining dialogue content based on multiple agents.

[0004] In a first aspect, embodiments of this application provide a method for determining dialogue content based on multiple agents, including:

[0005] Determine the current session content of the target user; The main agent extracts tags from the current session content to generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, at least one target sub-agent matching the current session is determined. The encapsulated metadata is sent to the target sub-agent, and the target sub-agent is controlled to determine the metadata output data based on the metadata; wherein the metadata includes the user profile data of the target user and the current session context data; Using the meta-output data corresponding to the target sub-agent as nodes, and the conflict types between different nodes as connecting edges, a conflict association graph is constructed; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. Based on the conflict type and severity of the connecting edges, and the acquired adjudication data, the conflict adjudication sub-agent is controlled to perform conflict resolution processing on the conflict association graph, generating a processed association graph, and constructing a target component tree based on the processed association graph and the set business priority order; the adjudication data includes the meta-output data and confidence level associated with each node in the conflict association graph, as well as the user attribute data and historical interaction preference data of the target user; Based on the target component tree, generate and display the reply content corresponding to the current session content.

[0006] In an optional implementation, after generating and displaying the reply content corresponding to the current session content, the method further includes: Based on the response content and the unresolved decisions in the decision chain, generate and display at least one guiding question; The system receives implicit feedback signals from the target user in response to the guiding question, wherein the implicit feedback signals include one or more of the following: user clicks on recommended cards, page dwell time, and modification / deletion of recommended options; the implicit feedback signals are used to adjust the dynamic weight coefficients of the sub-agent, and / or the user profile preference data of the target user.

[0007] In an optional implementation, the method further includes: obtaining the current load state of each sub-agent and the lock state of the resources it depends on; and determining the expected response time of the sub-agent based on the current load state of the sub-agent and the lock state of the resources it depends on. The step of determining at least one target sub-agent for the current session based on the current intent vector and the capability matrix data corresponding to each sub-agent includes: Based on the current intent vector and the capability matrix data corresponding to each sub-agent, calculate the matching score corresponding to the sub-agent; Based on the matching score corresponding to the sub-agent, the expected response time, and the dynamic weight coefficient, at least one target sub-agent is determined for the current session; wherein when there are multiple target sub-agents and multiple target sub-agents call the same dependent resource, a resource reservation and time slice round-robin mechanism is used for parallel scheduling.

[0008] In one optional implementation, determining the target user's current session content includes: Receive the raw session content input by the target user; Obtain the target user's user attribute data and historical dialogue data, and determine the target user's current user status; Based on the user attribute data, the historical dialogue data, and the current user status, the original conversation content is adjusted to generate the current conversation content for the target user.

[0009] In an optional implementation, after receiving the original session content, the method further includes: The original session content is subjected to content risk control detection to obtain the risk control detection results; When the risk control detection result indicates the presence of illegal content, preset acceptance content is generated and displayed.

[0010] In one optional implementation, controlling the target sub-agent to determine meta-output data based on the meta-information includes: The target sub-agent is controlled to analyze the metadata and generate analysis results; When the analysis results indicate that an external interface needs to be called, a call request is initiated, wherein the call request is used to obtain the required interactive data from the target data structure; the interactive data includes one or more of the following: dynamic knowledge base data, static knowledge base data, page links, and user terminal operation cards; Based on the interaction data and the analysis results, meta-output data is generated.

[0011] In one optional implementation, the step of controlling a conflict resolution sub-agent to perform conflict resolution processing on the conflict association graph based on the conflict type and severity of the connecting edges, and the acquired resolution data, to generate a processed association graph, includes: Based on the combination of the conflict type and the severity, a corresponding target resolution strategy is matched from a preset resolution strategy set, wherein the target resolution strategy includes node retention rules and presentation rules; Based on the adjudication data, adjudication calculations are performed on the part of the target resolution strategy involving node selection to determine the resolution results of the nodes at both ends of the connecting edge; According to the node retention rules and the resolution results, delete, retain or merge the nodes at both ends of the connection edge to generate the resolved node data; According to the presentation rules, determine the presentation method of the resolved node data; The resolved node data is written into the conflict association graph to generate the processed association graph.

[0012] In one optional implementation, the step of matching a corresponding target resolution strategy from a preset resolution strategy set based on the combination of the conflict type and the severity includes: If the conflict type is numerical and the severity is high, the first resolution strategy for matching includes: the node retention rule is to merge the meta-output data of the two nodes into a single node, and the value in the meta-output data is taken as the weighted average of the two; the presentation rule is exclusive presentation. If the conflict type is numerical and the severity is medium or low, the matching second resolution strategy includes: the node retention rule is to calculate the deviation rate of the numerical values ​​in the output data of the two node nodes; if the deviation rate falls within a preset tolerance range, the weighted average of the two is taken and merged into a single node; if the deviation rate does not fall within the preset tolerance range, one node is retained and the other node is deleted according to the adjudication data; the presentation rule is exclusive presentation. If the conflict type is conditional, the third resolution strategy includes: the node retention rule is to extract the trigger condition set from the output data of the nodes at both ends, match the trigger condition set with the user attribute data and the current user status, and retain the node corresponding to the condition that is matched, while deleting the node that does not match the condition; if both nodes match the condition or neither matches the condition, then in the case of high severity, the node with the higher business priority is retained, and in the case of medium or low severity, both nodes are retained; the presentation rule is exclusive presentation in the case of high severity, and interactive selection presentation in the case of medium or low severity. If the conflict type is logical and the severity is high, the matching fourth resolution strategy includes: the node retention rule is to extract logical assertions from the meta-output data of the two end nodes respectively, obtain the evidence chain identifier corresponding to each logical assertion, retain the node of the party with a traceable evidence chain and delete the other party's node; the presentation rule is exclusive presentation. If the conflict type is logical and the severity is medium or low, the matching fifth resolution strategy includes: the node retention rule is to retain the node of the party with a traceable chain of evidence and delete the other party's node; the presentation rule is to attach the logical assertion of the deleted node and the corresponding reason for the missing chain of evidence to the meta-output data of the retained node as supplementary explanation, and present it as supplementary explanation.

[0013] In one optional implementation, the step of performing adjudication calculations on the portion of the target resolution strategy involving node selection based on the adjudication data, and determining the resolution results of the nodes at both ends of the connecting edge, includes: When the node retention rule of the target resolution strategy indicates that one party's node should be retained and the other party's node should be deleted based on the adjudication data, the following steps are performed: Obtain the confidence level of the meta-output data corresponding to the nodes at both ends of the connection edge, the business priority of the sub-agent to which it belongs, and the timestamp of the referenced data source; Obtain the user profile preference data of the target user, and calculate the semantic matching degree between each meta output data and the user preference tag; Obtain the target user's selection results for similar conflicts in historical interactions, and calculate the matching frequency of each meta-output data with the user's historical selection preferences; Obtain the current user state of the target user and calculate the applicability matching degree between each meta output data and the current user state; The comprehensive credibility score of each node is calculated by weighting the confidence level of the meta-output data, the semantic matching degree, the matching frequency, the applicability matching degree, the business priority, and the timeliness determined based on the timestamp. The node with the higher overall credibility score is selected as the node to be retained, and the node with the lower overall credibility score is selected as the node to be deleted, which is the resolution result.

[0014] In one optional implementation, the step of performing deletion, retention, or merging operations on the nodes at both ends of the connecting edge according to the node retention rules and the resolution results to generate resolved node data, and writing the resolved node data into the conflict association graph to generate a processed association graph, includes: If the node retention rule indicates that the nodes should be merged into a single node, a new node is generated, the meta-output data of the two nodes are merged and written into the new node, the original nodes at both ends are deleted, and the new node is written into the conflict association graph. If the node retention rule indicates that one side should be retained and the other side should be deleted, then the node to be retained in the resolution result should be retained, and the node to be deleted in the resolution result should be deleted from the conflict association graph. If the node retention rule indicates that both ends of the node should be retained, then the nodes at both ends should remain unchanged, and the connection edge between the two ends of the node should be marked as resolved.

[0015] In one optional implementation, generating and displaying the reply content corresponding to the current session content based on the target component tree includes: Based on the target component tree and the set response strategy, the response content corresponding to the current session content is generated and displayed; wherein the response strategy includes at least one of the following: empathy support strategy, concern elimination strategy, and solution provision strategy.

[0016] Secondly, embodiments of this application also provide a dialogue content determination device based on multiple agents, comprising: The first determination module is used to determine the current session content of the target user; The first generation module is used to extract tags from the current session content using the main agent and generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, it determines at least one target sub-agent that matches the current session. The second determining module is used to send the encapsulated meta-information to the target sub-agent and control the target sub-agent to determine the meta-output data based on the meta-information; wherein the meta-information includes the user profile data of the target user and the current session context data; The construction module is used to construct a conflict association graph by taking the meta-output data corresponding to the target sub-agent as nodes and using the conflict types between different nodes as connecting edges; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. The second generation module is used to control the conflict adjudication sub-agent to perform conflict resolution processing on the conflict association graph based on the conflict type and severity of the connection edge and the acquired adjudication data, generate a processed association graph, and construct a target component tree based on the processed association graph and the set business priority order; the adjudication data includes the meta-output data and confidence level associated with each node in the conflict association graph, as well as the user attribute data and historical interaction preference data of the target user; The third generation module is used to generate and display the reply content corresponding to the current session content based on the target component tree.

[0017] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for determining dialogue content based on multiple agents.

[0018] This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method for determining dialogue content based on multiple agents.

[0019] The method, apparatus, computer device, and storage medium for determining dialogue content based on multiple agents proposed in this application match sub-agents based on the current intent vector and the capability matrix data of each agent by the main agent, and controls the target sub-agent to make subsequent routing decisions. This supports parallel routing or single-point routing of multiple sub-agents under composite intents, and achieves dynamic decoupling of business logic. When a new business domain is added, business responses can be achieved by adding the corresponding sub-agents and configuring the capability matrix, which can effectively reduce the complexity and cost of system expansion.

[0020] Furthermore, this application constructs a conflict association graph, classifies and grades conflicts, and combines this with intelligent resolution by a conflict adjudication sub-agent. This effectively handles various conflicts arising when multiple sub-agents return output results in parallel, avoiding the loss of key information due to simple overwriting and ensuring the consistency and integrity of the final output. Moreover, by assembling the resolved meta-output data into a box computing component tree, a paradigm shift from plain text responses to interactive user interface components is achieved. This allows users to directly perform interactive operations such as clicking, selecting, modifying, and submitting within the dialog box, improving interaction efficiency and user experience. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for determining dialogue content based on multiple agents; Figure 2 This is a flowchart illustrating a multi-agent-based dialogue content determination method, as shown in another exemplary embodiment of this application. Figure 3 This is a schematic diagram of a multi-agent-based dialogue content determination device shown in an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in this application. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0025] To facilitate understanding of this embodiment, a detailed description of the multi-agent-based dialogue content determination method disclosed in this application embodiment will be provided first. The executing entity of the multi-agent-based dialogue content determination method provided in this application embodiment is generally a computer device with a certain computing power. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device can be user equipment (UE), a mobile device, a user terminal, a handheld device, a computing device, an in-vehicle device, an embedded device, etc. In some possible implementations, this multi-agent-based dialogue content determination method can be implemented by a processor calling computer-readable instructions stored in memory.

[0026] See Figure 1 The diagram shows a flowchart of a multi-agent-based dialogue content determination method provided in an embodiment of this application. The method includes steps S101 to S106, wherein: S101. Determine the current session content of the target user; S102. Use the main agent to extract tags from the current session content and generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, determine at least one target sub-agent that matches the current session. S103. Send the encapsulated metadata to the target sub-agent and control the target sub-agent to determine the metadata output data based on the metadata; wherein the metadata includes the user profile data of the target user and the current session context data; S104. Using the meta-output data corresponding to the target sub-agent as nodes, and the conflict types between different nodes as connecting edges, construct a conflict association graph; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. S105. Based on the conflict type and severity of the connection edges, and the acquired adjudication data, control the conflict adjudication sub-agent to perform conflict resolution processing on the conflict association graph, generate a processed association graph, and construct a target component tree based on the processed association graph and the set business priority order; the adjudication data includes the meta-output data associated with each node in the conflict association graph and its confidence level, as well as the user attribute data and historical interaction preference data of the target user; S106. Based on the target component tree, generate and display the reply content corresponding to the current session content.

[0027] This application uses a master agent to match sub-agents based on the current intent vector and the capability matrix data of each agent, and controls the target sub-agent to make subsequent routing decisions. This supports parallel routing or single-point routing of multiple sub-agents under composite intents, and achieves dynamic decoupling of business logic. When a new business domain is added, business responses can be achieved by adding the corresponding sub-agent and configuring the capability matrix, which can effectively reduce the complexity and cost of system expansion.

[0028] This application constructs a conflict association graph and classifies (e.g., numerical, conditional, logical) and hierarchically processes conflicts. Combined with intelligent resolution by a conflict adjudication sub-agent, it effectively handles various conflicts arising when multiple sub-agents return output results in parallel. This avoids the loss of key information due to simple overwriting and ensures the consistency and integrity of the final output. Furthermore, by assembling the resolved meta-output data into a box computing component tree, a paradigm shift from plain text responses to interactive user interface (UI) components is achieved. This allows users to directly perform interactive operations such as clicking, selecting, modifying, and submitting within the dialog box, improving interaction efficiency and user experience.

[0029] Regarding S101, during implementation, the target user can input the original conversation content in the dialog box, and the obtained original conversation content will be used as the current conversation content. Alternatively, to ensure the accuracy of the dialogue process, this application can also rewrite the original conversation content and use the rewritten conversation content as the current conversation content.

[0030] The original conversation content can include one or more formats such as text, speech, images, and video. For non-text inputs, corresponding multimodal recognition modules (such as speech recognition models and image understanding models) can be used to recognize and convert the input content to generate text representations for subsequent processing.

[0031] Optionally, determining the current conversation content of the target user includes: receiving the original conversation content input by the target user; acquiring the user attribute data and historical dialogue data of the target user, and determining the current user state of the target user; adjusting the original conversation content according to the user attribute data, the historical dialogue data and the current user state to generate the current conversation content of the target user.

[0032] After receiving the original conversation content, user attribute data and historical conversation data of the target user can be obtained. User attribute data includes, but is not limited to, static profile information such as age, gender, occupation, region, preferences, and historical transaction records. Historical conversation data can include the target user's historical conversation records and behavior logs.

[0033] Determining the current user state of the target user can be achieved by using neural network models to assess the current user state based on the content of the user's multi-turn dialogue, and / or by judging the current user state based on the user's operational behavior, etc. The user state includes, but is not limited to: emotional state (such as hesitation, eagerness, satisfaction, dissatisfaction, anxiety, doubt, etc.), decision-making stage state (such as cognitive exploration state, solution comparison state, decision-making state, etc.), and interaction preference state (such as preference for text descriptions, video explanations, quick links, etc.).

[0034] For example, based on user attribute data, historical dialogue data, and the current user state, a large language model can be invoked to semantically rewrite the original conversation content. This includes completing omitted subjects and objects, replacing pronouns with specific entities, and explicitly writing hard constraints from the user profile into the rewritten content. This generates complete and unambiguous adjusted conversation content, which is then used as the target user's current conversation content.

[0035] For example, in a used car transaction scenario, if a user inputs via voice, "My budget is limited, so I don't want anything too expensive," the original conversation content is first parsed. For instance, if the parsed content is "What kind of car can I buy with a limited budget that isn't too expensive?", then by combining user attribute data and historical dialogue data, such as user attribute data including a budget range of 80,000-120,000 yuan and a preference for SUVs, and historical dialogue data such as previous discussions about ownership transfer, the original conversation content is rewritten as "The user wants to buy an SUV model that fits their budget profile and offers good value for money."

[0036] This application uses user profile data, historical dialogue data, and current user status to perform semantic rewriting, which can fully understand the user's incomplete expressions such as omissions and references, and combine user preferences to realize personalized services, effectively improving the accuracy of semantic understanding and the level of personalization of interaction.

[0037] Optionally, after receiving the original session content, the method further includes: performing content risk control detection on the original session content to obtain a risk control detection result; and generating and displaying preset acceptance content when the risk control detection result indicates the presence of illegal content.

[0038] In practice, after receiving the original conversation content, risk control detection neural network models and large language models can be used to perform content risk control detection on the original conversation content to obtain the risk control detection result. If the risk control detection result indicates that there is no illegal content, subsequent content rewriting processing will be carried out. Conversely, if the risk control detection result indicates that there is illegal content, no subsequent processing process will be carried out, and a preset response message will be directly generated and displayed, such as "Sorry, I cannot answer this question."

[0039] For S102, the main agent can be used to extract tags from the current session content and generate the current intent vector. For example, the current intent vector can be generated based on the extracted tag content through a text embedding model.

[0040] Sub-agents can be divided according to business direction, such as pushcart sub-agents, financial sub-agents, service rights sub-agents, etc., and / or they can also be divided according to technical implementation, such as knowledge bases, tools, etc.

[0041] Each agent is configured with capability matrix data, which reflects the processing capabilities of that agent. For example, a corresponding responsibility description text can be generated for each agent using natural language. This responsibility description text can be converted into capability matrix data using a text embedding model, and the capability matrix data can be stored in a storage unit accessible to the master agent. This allows the master agent to determine the matching score of each agent based on the current intent vector and the capability matrix data corresponding to each sub-agent, and then determine at least one target sub-agent for the current session based on the matching.

[0042] For example, the main agent can calculate the matching score of each sub-agent based on the current intent vector I and the capability matrix data capability_i corresponding to each sub-agent, such as matching score_i = cosine(I,capability_i) × priority_i, where priority_i is the preset business priority coefficient corresponding to the i-th sub-agent.

[0043] For example, the main agent can extract domain labels such as {pushcart, finance, service rights} and generate an intent vector I based on the domain labels. It can then calculate the matching score between the intent vector I and each sub-agent. For instance, if the matching score of the pushcart sub-agent is 0.92, the finance sub-agent is 0.78, the service rights sub-agent is 0.65, and the transfer sub-agent is 0.2, and the set matching threshold is 0.6, then the pushcart sub-agent, the finance sub-agent, and the service rights sub-agent are identified as target sub-agents for subsequent parallel routing of these three target sub-agents.

[0044] In this application, parallel routing (i.e., parallel processing of multiple target sub-agents) can be used when the intent involves a complex business domain, and single-point routing (i.e., processing of a single target sub-agent) can be used when the intent is a single domain.

[0045] Optionally, the method further includes: obtaining the current load state of each sub-agent and the lock state of the resources it depends on; and determining the expected response time of the sub-agent based on the current load state of the sub-agent and the lock state of the resources it depends on.

[0046] The step of determining at least one target sub-agent for the current session based on the current intent vector and the capability matrix data corresponding to each sub-agent includes: calculating the matching score corresponding to the sub-agent based on the current intent vector and the capability matrix data corresponding to each sub-agent; determining at least one target sub-agent for the current session based on the matching score corresponding to the sub-agent, the expected response time, and the dynamic weight coefficient; wherein when there are multiple target sub-agents and multiple target sub-agents call the same dependent resource, a resource reservation and time slice round-robin mechanism is used for parallel scheduling.

[0047] During implementation, each sub-agent can periodically report its current load status to the monitoring module, such as the current number of concurrent requests, CPU / memory usage, and the average processing time of the last N calls. The master agent can read the current load status during routing decisions to calculate queuing time, for example, queuing time = (current concurrency / maximum concurrency threshold) × baseline time × coefficient. High load is considered a negative factor in the routing benefit function to avoid routing requests to overloaded sub-agents.

[0048] It can also maintain a resource occupancy table, which records information such as the current occupant, estimated release time, and average occupancy duration for each shared dependent resource (such as an external application programming interface (API), dynamic database, etc.). When multiple sub-agents depend on the same external resource, the master agent adds a serialization penalty coefficient to the total time consumed by the above-mentioned sub-agents or adjusts the route combination when calculating the routing benefits to avoid simultaneous invocation of conflicting resources.

[0049] Furthermore, the main agent can determine the expected response time of the sub-agents based on the current load state and the lock state of the resources it depends on. For example, the expected response time can be the sum of the baseline time, queuing time, and resource contention time (if lock conflicts exist).

[0050] After the main agent calculates the matching score based on the current intent vector and the capability matrix data of the sub-agents, it determines at least one target sub-agent for the current session based on the matching score of the sub-agent, the expected response time, and the dynamic weight coefficient. For example, a routing benefit function can be set to calculate the global benefit, and the combination of sub-agents that maximizes the global benefit can be selected as the routing target (i.e., the target sub-agent). The dynamic weight coefficient is stored in a dedicated storage space for each user and can be updated based on the user's historical interaction feedback; that is, the dynamic weight coefficient corresponding to the same sub-agent can be different for different users.

[0051] For example, the calculation process of the routing reward function can be as follows: The main agent enumerates all possible combinations of sub-agents. When there are many sub-agents, N sub-agents can be selected for combination in descending order of matching scores, or sub-agents with matching scores greater than a preset threshold can be selected for combination. After constructing the sub-agent combinations, the total matching degree of each combination (the weighted sum of the matching scores of each sub-agent, where the weight of each agent can be a dynamic weight coefficient) and the total time consumption can be calculated (in parallel calls, the total combined time consumption is the maximum of the expected response times among multiple sub-agents; in serial calls, the total combined time consumption is the sum of the expected response times among multiple sub-agents). Substituting these values ​​into the reward function R = α × total combined matching degree... The total time consumed by the combination is β×, and the combination with the highest global return is selected as the final routing target. Here, α and β are adjustable hyperparameters used to control the weights of matching accuracy and response speed.

[0052] After identifying the target sub-agents, if there are multiple target sub-agents and multiple target sub-agents call the same dependent resource, a resource reservation and time-slice round-robin mechanism can be used for parallel scheduling. For example, a resource reservation scheduler can be set up. Before calling a resource, multiple target sub-agents send reservation requests to the resource reservation scheduler (the reservation request can include the estimated duration). The resource reservation scheduler checks the resource status; if it is idle, it immediately grants the resource lock; if it is busy, it adds the request to the resource's First-In-First-Out (FIFO) waiting queue. Then, the target sub-agent that has acquired the resource lock executes the call process and starts a time-slice timer. After the time slice expires, regardless of whether the target sub-agent's call is completed, the resource reservation scheduler forcibly suspends the current target sub-agent (while saving its context) and moves it to the end of the queue, allowing the next waiting target sub-agent in the waiting queue to execute the resource scheduling process. After the target sub-agent completes the call, it voluntarily releases the resource lock so that the resource reservation scheduler can wake up the next resource caller. This mechanism can fundamentally break the "circular wait" condition caused by deadlock and ensure operational stability in high-concurrency scenarios.

[0053] This application introduces load awareness, resource reservation, and time-slice round-robin scheduling mechanisms to effectively avoid resource contention and deadlock in scenarios with multiple sub-agents calling in parallel, thus ensuring stability and response speed in high-concurrency scenarios.

[0054] Regarding S103, after identifying the target sub-agent, the user profile data and current session context data of the target user can be encapsulated into metadata, which is then sent to each target sub-agent. The target sub-agents are then controlled to determine the metadata output data based on the metadata. The user profile data may include long-term memory data such as user attribute data and historical dialogue data, while the current session context data may include short-term memory data such as the current user state and current session content. This application achieves personalized services by accessing both long-term and short-term memory data of the user.

[0055] Optionally, the step of controlling the target sub-agent to determine the metadata output data based on the metadata includes: The target sub-agent is controlled to analyze the metadata and generate analysis results; When the analysis results indicate that an external interface needs to be called, a call request is initiated, wherein the call request is used to obtain the required interactive data from the target data structure; the interactive data includes one or more of the following: dynamic knowledge base data, static knowledge base data, page links, and user terminal operation cards; Based on the interaction data and the analysis results, meta-output data is generated.

[0056] During implementation, each target sub-agent receives metadata, analyzes it, and generates analysis results. These results may include the question content, required data, user preferences, etc. When the analysis results indicate that an external interface needs to be called, the target sub-agent initiates a call request to obtain the necessary interaction data from the target data structure. This interaction data includes, but is not limited to: dynamic knowledge base data (such as real-time inventory, real-time interest rates, etc.), static knowledge base data (such as product parameters, frequently asked questions (FAQ), etc.), page links (such as official website details pages), and user-side operation cards (such as buttons, sliders, forms, etc.).

[0057] The target sub-agent generates structured meta-output data based on interaction data and analysis results. For example, meta-output data may include, but is not limited to: data subjects (such as a list of recommended car models or details of financial plans), descriptions of interactive elements (such as button labels and corresponding APIs), confidence scores (the target sub-agent's confidence score in its output results), and timestamps of the referenced data sources.

[0058] For S104, the main agent uses the meta-output data returned by each target sub-agent as nodes, and constructs a conflict association graph with the conflict types between different nodes as connecting edges. The conflict types include: numerical conflicts, which involve specific numbers, such as the cart-pushing sub-agent recommending a car model priced at 120,000 yuan, while the financial sub-agent calculates the user's maximum credit limit to be 100,000 yuan; conditional conflicts, where the preconditions for one meta-output data to be true are negated by another, such as the cart-pushing sub-agent recommending a discounted car model requiring "local residency," but the user profile shows the user is not a local resident; and logical conflicts, where two meta-output data are logically mutually exclusive, such as the cart-pushing sub-agent recommending a "full payment purchase plan," while the financial sub-agent recommends "installment purchase plan B," creating a logical choice between the two. This conflict association graph clearly displays the conflict relationships between different meta-output data.

[0059] For S105, conflict resolution is performed based on the constructed conflict association graph. Specifically, based on the conflict type and severity (e.g., high / medium / low) of each connection edge, and combined with the acquired adjudication data, the conflict adjudication sub-agent is controlled to process the conflict association graph for conflict resolution. The adjudication data includes: the meta-output data and confidence scores of each node association, the target user's user attribute data, historical interaction preference data, and the current user state. The historical interaction preference data includes one or more of the following: historical dialogue data, the target user's selection results for similar conflicts in historical interactions, and the target user's user profile preference weights.

[0060] For example, nodes can be categorized into core nodes, reconcilable nodes, and mutually exclusive nodes based on the type and severity of the conflict. Core nodes are output answers with high value and best match the user's needs. Reconcilable nodes can be those where there is a conflict, but the conflict can be resolved by adjusting parameters. For example, a financial plan may have a high down payment ratio, but the monthly payment can be reduced by adjusting the loan term to match the user's budget. Mutually exclusive nodes are those where there is an irreconcilable logical conflict, and one must be chosen. For example, only one of paying in full or purchasing in installments can be recommended.

[0061] Furthermore, the conflict adjudication sub-agent can combine adjudication data to retain core nodes in the conflict association graph, selectively retain one of mutually exclusive nodes, and adjust the parameters of reconcilable nodes. Through node adjustment, dynamic pruning, and reattaching operations on the conflict association graph, conflict resolution is achieved, generating a processed association graph. For example, if historical interaction preference data or historical interaction data determines that a user frequently views or understands installment payment plans, then the installment plan is identified as the recommended plan, and the full payment plan is deleted. Alternatively, selection can be based on the confidence level of the meta-output data, i.e., selecting meta-output data with higher confidence levels for retention. This allows for the construction of a target component tree based on the processed association graph and the set business priority order.

[0062] Optionally, the step of controlling the conflict adjudication sub-agent to perform conflict resolution processing on the conflict association graph based on the conflict type and severity of the connecting edges and the acquired adjudication data, and generating a processed association graph, includes: Step a1: Based on the combination of the conflict type and the severity, match the corresponding target resolution strategy from the preset resolution strategy set. The target resolution strategy includes node retention rules and presentation rules. Step a2: Based on the adjudication data, perform adjudication calculations on the part of the target resolution strategy involving node selection to determine the resolution results of the nodes at both ends of the connecting edge; Step a3: According to the node retention rules and the resolution results, perform deletion, retention or merging operations on the nodes at both ends of the connection edge to generate the resolved node data; Step a4: Determine the presentation method of the resolved node data according to the presentation rules; Step a5: Write the resolved node data into the conflict association graph to generate the processed association graph.

[0063] In step a1, multiple preset resolution strategies can be set and stored in a preset resolution strategy set. Then, based on the combination of conflict type and severity, the corresponding target resolution strategy is matched from the preset resolution strategy set. The target resolution strategy includes node retention rules and presentation rules.

[0064] Optionally, matching a corresponding target resolution strategy from a preset resolution strategy set based on the combination of the conflict type and the severity includes: Case 1: If the conflict type is numerical and the severity is high, the first resolution strategy includes: the node retention rule is to merge the meta-output data of the two nodes into a single node, and the value in the meta-output data is taken as the weighted average of the two; the presentation rule is exclusive presentation.

[0065] Scenario 2: If the conflict type is numerical and the severity is medium or low, the matching second resolution strategy includes: the node retention rule is to calculate the deviation rate of the numerical values ​​in the output data of the two node nodes; if the deviation rate falls within a preset tolerance range, the weighted average of the two is taken and merged into a single node; if the deviation rate does not fall within the preset tolerance range, one node is retained and the other node is deleted according to the adjudication data; the presentation rule is exclusive presentation.

[0066] Scenario 3: If the conflict type is conditional, the third resolution strategy includes: the node retention rule is to extract the trigger condition set from the output data of both nodes, match the trigger condition set with the user attribute data and the current user status, and retain the node corresponding to the condition that is matched, while deleting the node that does not match the condition; if both nodes match the condition or neither matches the condition, then in the case of high severity, the node with the higher business priority is retained, and in the case of medium or low severity, both nodes are retained; the presentation rule is exclusive presentation in the case of high severity, and interactive selection presentation in the case of medium or low severity.

[0067] It should be noted that when neither of the two nodes in a conditional conflict meets the triggering condition, it means that the meta-answers (i.e., meta-output data) output by both sub-agents do not match the current user state and user attribute data. In this case, relying solely on business priority to select nodes may not meet the user's actual needs. To alleviate the above problem, this application can also utilize a set linkage processing mechanism for handling.

[0068] Specifically, when it is detected that neither of the two endpoints has met the triggering condition, the conflict resolution sub-agent obtains the current load status and the length of the pending task queue of each target sub-agent, and calculates the expected response time of each target sub-agent.

[0069] If the load status of each target sub-agent is low and the expected response time is lower than the preset time threshold, then the negotiation mode is triggered. The target sub-agents that have conflicted trace back up along their respective target hierarchy, exchange counter-proposals, and attempt to reach a new consensus solution.

[0070] If the target sub-agent is under medium or high load, or if the expected response time exceeds the preset time threshold, the negotiation mode will be abandoned, and the one with higher business priority will be reserved as the node to avoid increasing the system burden and affecting the overall response efficiency during the negotiation process.

[0071] Once the negotiation mode is initiated, if the number of negotiation rounds has not exceeded the preset round threshold but the counter-proposals from both parties are gradually approaching each other, the upper limit of the preset round threshold can be increased to give both parties more rounds of exchange to reach a consensus. If the number of negotiation rounds exceeds the preset round threshold and a consensus is still not reached, the negotiation mode will be exited and the conflict resolution sub-agent will execute the resolution according to the business priority.

[0072] Through the aforementioned linkage mechanism, when the matching results for conditional conflicts are unsatisfactory (no match), the most suitable resolution path can be flexibly selected based on the current resource status. When the load is light, negotiation is prioritized to obtain a better solution; when the load is heavy, direct adjudication is performed to ensure response speed, thus achieving a dynamic balance between resolution quality and system performance.

[0073] Case 4: If the conflict type is logical and the severity is high, the matching fourth resolution strategy includes: the node retention rule is to extract the logical assertions from the meta-output data of the two nodes respectively, obtain the evidence chain identifiers corresponding to each logical assertion, retain the node of the party with a traceable evidence chain and delete the other node; the presentation rule is exclusive presentation.

[0074] Case 5: If the conflict type is logical and the severity is medium or low, the fifth resolution strategy includes: the node retention rule is to retain the node of the party with a traceable chain of evidence and delete the other party's node; the presentation rule is to attach the logical assertion of the deleted node and the corresponding reason for the missing chain of evidence to the meta-output data of the retained node as supplementary explanations.

[0075] In this application, the preset conflict resolution strategy set is not a static rule table. During actual operation, the matching priority of each conflict resolution strategy can be dynamically adjusted based on the collected feedback results. For example, after the conflict resolution strategy completes the conflict resolution process and generates a response, implicit feedback signals are obtained to monitor the target user's acceptance of the resolution result, thereby adjusting the priority of the conflict resolution strategy.

[0076] In step a2, after determining the target resolution strategy, if the target resolution strategy involves node selection, the resolution results of the two endpoints can be determined based on the adjudication data. For example, the second resolution strategy mentioned above involves node selection for deletion and retention.

[0077] In optional step a2, based on the adjudication data, adjudication calculations are performed on the part of the target resolution strategy involving node selection to determine the resolution results of the nodes at both ends of the connecting edge. Specifically, this includes: when the node retention rule of the target resolution strategy indicates that one node should be retained and the other node should be deleted based on the adjudication data, the following steps are performed: Step a21: Obtain the confidence level of the meta-output data corresponding to the nodes at both ends of the connection edge, the business priority of the sub-agent to which it belongs, and the timestamp of the referenced data source; Step a22: Obtain the user profile preference data of the target user, and calculate the semantic matching degree between each meta output data and the user preference tag; Step a23: Obtain the target user's selection results for similar conflicts in the historical interactions, and calculate the matching frequency of each meta-output data with the user's historical selection preferences; Step a24: Obtain the current user state of the target user and calculate the applicability matching degree between each element output data and the current user state; Step a25: Calculate the comprehensive credibility score of each node by weighting the confidence level of the meta-output data, the semantic matching degree, the matching frequency, the applicability matching degree, the business priority, and the timeliness determined based on the timestamp. Step a26: The node with the higher overall credibility score is selected as the node to be retained, and the node with the lower overall credibility score is selected as the node to be deleted, which is the resolution result.

[0078] During implementation, when the node retention rule of the target resolution strategy (such as the second resolution strategy) indicates that one side should be retained and the other side should be deleted based on the adjudication data, the retained node and the deleted node in the two-end nodes can be determined according to the above process.

[0079] First, we can obtain the confidence level of the meta-output data corresponding to the nodes at both ends of the connection edge, the business priority of the corresponding sub-agent, and the timestamp of the referenced data source (i.e., the existence of the referenced data source). We can also obtain the user profile preference data of the target user, such as extracting user profile preference data from user attribute data, extracting user preference tags from the user profile preference data, and calculating the semantic matching degree between each meta-output data and the user preference tags.

[0080] Furthermore, based on historical interaction preference data in the adjudication data, the target user's selection results for similar conflicts in historical interactions are determined, and based on these selection results, the matching frequency of each meta-output data in the user's historical selection preferences is statistically analyzed.

[0081] Obtain the current user state of the target user, and calculate the applicability matching degree between each meta output data and the current user state. For example, the state vector matched by each meta output data can be determined, and the applicability matching degree between the state vector and the current user state can be determined.

[0082] Then, based on the confidence level, semantic matching degree, matching frequency, applicability matching degree, business priority, and timeliness determined by timestamp of the meta output data, the comprehensive credibility score of each node is calculated by weighting. Among them, confidence level, semantic matching degree, matching frequency, applicability matching degree, business priority, and timeliness correspond to preset weight coefficients, which can be set according to business needs.

[0083] Finally, the node with the higher overall credibility score was selected as the node to be retained, and the node with the lower overall credibility score was selected as the node to be deleted, which was used as the resolution result.

[0084] In steps a3 to a5, based on the node retention rules and resolution results, operations such as deletion, retention, data adjustment, or merging can be performed on the nodes at both ends of the connecting edge to generate resolved node data. The presentation method of the resolved node data is determined according to the presentation rules. The resolved node data is then written into the conflict association graph to generate the processed association graph.

[0085] Optionally, according to the node retention rules and the resolution results, delete, retain, or merge the nodes at both ends of the connection edge to generate resolved node data, and write the resolved node data into the conflict association graph to generate a processed association graph, including: If the node retention rule indicates that the nodes should be merged into a single node, a new node is generated, the meta-output data of the two ends are merged and written into the new node, the original nodes at both ends are deleted, and the new node is written into the conflict association graph; thus realizing the merging operation of the nodes at both ends of the connection edge.

[0086] If the node retention rule indicates that one side should be retained and the other side should be deleted, then the node to be retained in the resolution result should be retained, and the node to be deleted in the resolution result should be deleted from the conflict association graph; thus, the deletion operation of the nodes at both ends of the connecting edge is realized.

[0087] If the node retention rule indicates that both ends of the node should be retained, then the nodes at both ends remain unchanged, and the connecting edge between the two ends of the node is marked as resolved, thus realizing the retention operation of the nodes at both ends of the connecting edge.

[0088] This embodiment uses a car sales dialogue scenario as an example to illustrate the construction and resolution process of the conflict association graph. Assume the target user inputs "My budget is limited, don't choose anything too expensive." After semantic understanding and rewriting, target sub-agents are determined, such as: a car sales sub-agent, a financial sub-agent, and a service rights sub-agent. Based on the received metadata, the three target sub-agents determine their metadata output data. For example, the car sales sub-agent determines the metadata output data as follows: recommended car model A (priced at 120,000 yuan), car model B (priced at 98,000 yuan), and car model C (priced at 115,000 yuan), with a confidence level of 0.92. The financial sub-agent determines the metadata output data as follows: the user's maximum credit limit is calculated to be 100,000 yuan, and a 20% down payment plan is recommended, with a confidence level of 0.85. The service rights sub-agent determines the metadata output data as follows: recommended car model A enjoys "3 years of free maintenance" benefits, with a confidence level of 0.78.

[0089] Based on the constructed conflict correlation diagram, the following conflicts were detected: Numerical conflict (severity: high), namely, the recommended vehicle model A (120,000 yuan) by the push-cart sub-agent contradicts the maximum credit limit (100,000 yuan) calculated by the financial sub-agent. Conditional conflict (severity: medium), namely, the "3-year free maintenance" benefit of the service rights sub-agent only applies to vehicle model A, but vehicle model A has been marked as not recommended because it exceeds the budget.

[0090] The conflict adjudication sub-agent matches a resolution strategy based on the conflict type and severity. For high-severity numerical conflicts, the first resolution strategy is applied, which merges the meta-output data of both nodes into a single node, and takes the weighted average of the values ​​in the meta-output data; the presentation rule is exclusive presentation. Specifically, based on the adjudication data, the adjudication calculates the confidence scores (0.92 for the cart, 0.85 for finance), user profile preferences (e.g., budget of 80,000-120,000), and historical interaction preferences (e.g., the user's past tendency towards conservative spending) of both nodes, and calculates a weighted overall credibility score. Although the confidence score of the finance sub-agent's meta-output data is slightly lower, its semantic match with the user profile of "limited budget" is higher, and it is consistent with the user's historical spending preferences, resulting in a higher overall score. Therefore, the adjudication merges the two meta-output data into a single node, retains the finance sub-agent's limit (100,000 yuan), and deletes the cart sub-agent's vehicle type A recommendation.

[0091] For severe conflicts in conditional scenarios, a third resolution strategy is applied (both ends are retained for low to medium severity, with interactive selection). The system retains the vehicle model B and model C recommendations from the pushcart sub-agent, while also retaining the rights information of the service rights sub-agent, but presenting the rights information as "optional," allowing users to decide whether to adjust their budget for the sake of the rights.

[0092] After the resolution is complete, the node data in the processed association graph is assembled into a component tree: the top displays the credit limit calculation card (exclusively presented), the vehicle type B and vehicle type C recommendation cards (with the "View Benefits" entry), and the bottom displays the benefit details (collapsed presentation, which users can expand by clicking).

[0093] This application can achieve more accurate conflict resolution through the above process, thereby improving the accuracy of the response content and ensuring user experience.

[0094] Optionally, when the conflict type is a conditional conflict or a logical conflict, and the severity is medium or low, before the conflict resolution sub-agent performs conflict resolution processing, this application may also set a negotiation mode processing flow. The negotiation mode refers to a distributed negotiation mechanism in which each target sub-agent that generates a conflict attempts to reach a consensus autonomously through multiple rounds of information exchange and scheme adjustment before the conflict resolution sub-agent intervenes.

[0095] This application sets up a negotiation mode for conditional or logical conflicts with a medium or low severity. This is because numerical conflicts typically involve explicit numerical comparisons (such as price and credit limit), and their resolution paths are relatively predictable, eliminating the need for a negotiation mechanism. Conditional and logical conflicts, however, involve complex condition matching and logical assertions, and their resolution directly impacts the correctness of subsequent business logic and user experience, exhibiting multiple solutions. Therefore, negotiation is suitable for obtaining a better solution. Furthermore, when the severity is medium or low, the time cost of negotiation is acceptable; if the severity is high, rapid adjudication is necessary to avoid response delays, making multiple rounds of negotiation unsuitable.

[0096] In this application, the negotiation mode is the preferred path for prior adjudication, preceding the adjudication execution of the conflict adjudication sub-agent. When negotiation is successful, the negotiation result replaces the adjudication result as the resolution result. When negotiation fails (timeout or inability to reach a consensus), the negotiation mode is exited, and the process jumps back to the adjudication execution flow of the conflict adjudication sub-agent. This design ensures that the resolution result can still be output normally when negotiation fails, guaranteeing the robustness of the session system.

[0097] Specifically, the negotiation mode process may include: triggering each conflicting target sub-agent to enter the negotiation mode; obtaining the target hierarchy structure of each target sub-agent, which includes upper-level target nodes and lower-level implementation nodes, with the upper-level target node being an abstraction of the lower-level implementation node; tracing back along the target hierarchy structure of each target sub-agent to determine the upper-level target node corresponding to its current conflict node; generating a counter-proposal based on the upper-level target node and exchanging the counter-proposal among the target sub-agents; when a counter-proposal is accepted by all conflicting target sub-agents, using the meta-output data corresponding to the counter-proposal as the resolution result; and exiting the negotiation mode when the number of negotiation rounds exceeds a preset round threshold and consensus is not reached, with the conflict adjudication sub-agent executing the adjudication.

[0098] After triggering the negotiation mode, the target hierarchy structure of each target sub-agent is first obtained. The target hierarchy structure is a directed acyclic graph structure pre-constructed for each sub-agent to represent its decision logic hierarchy. It includes upper-level target nodes and lower-level implementation nodes. The upper-level target nodes are abstractions of the lower-level implementation nodes, and the lower-level implementation nodes are concrete implementations of the upper-level target nodes.

[0099] For example, the target hierarchy of a push-cart sub-agent can be represented as follows: upper-level target node – meeting user travel needs; middle-level target node – recommending suitable car models and providing car purchase plans; lower-level implementation nodes – selecting car models, comparing parameters, calculating prices, and matching financial solutions, etc. As another example, the target hierarchy of a financial sub-agent can be represented as follows: upper-level target node – meeting user funding needs; middle-level target node – providing financial solutions and assessing credit limits; lower-level implementation nodes – querying user credit reports, calculating down payment ratios, estimating monthly payments, and generating loan plans, etc.

[0100] In negotiation mode, the complete target hierarchy structure of each sub-agent can be obtained by reading the target hierarchy structure data pre-stored in the configuration information of each sub-agent. The target hierarchy structure data can be stored in JSON format or graph database format, and it includes fields such as the unique identifier of each node in the target hierarchy structure, node name, node type (upper-level target node / lower-level implementation node / intermediate-level node), parent node identifier, and child node identifier.

[0101] After obtaining the target hierarchy structure of each target sub-agent, each target sub-agent backtracks upwards along its respective target hierarchy structure to determine the upper-level target node corresponding to its current conflict node. The specific process of backtracking upwards can include: assuming the current conflict node is conflict node one, starting from conflict node one, traversing upwards layer by layer along the parent node pointers in the target hierarchy structure, passing through intermediate layer nodes until reaching the preset abstraction level, obtaining the identifiers and semantic descriptions of each layer's nodes, and finally determining the upper-level target node two corresponding to the current conflict node. When multiple candidate upper-level target nodes are encountered during the backtracking process, the upper-level target node closest to the current conflict node (i.e., with the smallest hierarchy difference) can be used as the target node.

[0102] For example, suppose the pushcart sub-agent and the financial sub-agent have a conditional conflict due to "the down payment ratio of the recommended car model". The conflict node of the pushcart sub-agent is matching a financial solution, and the conflict node of the financial sub-agent is calculating the down payment ratio. In this case, the conflict nodes of the two sub-agents can be backtracked separately: For the pushcart sub-agent, by backtracking "matching a financial solution → recommending a suitable car model → meeting the user's travel needs", the upper-level target node is determined to be meeting the user's travel needs; for the financial sub-agent, by backtracking "calculating the down payment ratio → providing a financial solution → meeting the user's funding needs", the upper-level target node is determined to be meeting the user's funding needs.

[0103] By tracing back upwards as described above, conflicts between sub-agents at the specific implementation level can be elevated to the semantic space at the target level, allowing conflicts that might otherwise be irreconcilable at the implementation level to find a common ground for negotiation at a higher level of abstraction.

[0104] After determining the target node, a counter-proposal can be generated based on the upper-level target node, and these counter-proposals can be exchanged among the target sub-agents. The generation of counter-proposals can include: each target sub-agent using the upper-level target node as a constraint, expanding downwards within its target hierarchy starting from that upper-level target node, and generating alternative implementation schemes that satisfy the upper-level target node; these alternative implementation schemes are output in structured data form as the counter-proposal for that sub-agent.

[0105] For example, suppose the push-cart sub-agent determines that the upper-level target node is to meet the user's travel needs, and its generated counter-proposal is: "Recommend alternative model D that meets the user's budget range (80,000-120,000), with a down payment ratio of 20%"; the financial sub-agent determines that the upper-level target node is to meet the user's funding needs, and its generated counter-proposal is: "Adjust the loan term to 48 months and reduce the down payment ratio to 15%". Both parties send their respective counter-proposals to each other.

[0106] The exchange of counter-proposals can be achieved through a message bus or point-to-point communication channel, employing an asynchronous, non-blocking communication mode to avoid any sub-agent becoming blocked while waiting for a response from the other. Upon receiving a counter-proposal, each sub-agent analyzes and evaluates it to determine whether it falls within its acceptable adjustment range.

[0107] For example, suppose sub-agent A receives a counter-proposal P_B from sub-agent B. Sub-agent A matches the implementation scheme in P_B with its own target hierarchy, determining whether P_B is compatible with the upper-level target nodes in its own target hierarchy. If P_B ​​is within its adjustable range, sub-agent A sends an acceptance response; otherwise, based on the target constraints in P_B, sub-agent A regenerates a new counter-proposal P_A' as a response to P_B and sends P_A' to sub-agent B, entering the next round of negotiation.

[0108] When a counter-proposal is accepted by all conflicting target sub-agents, the corresponding meta-output data is used as the resolution result. That is, in the current negotiation round, if a counter-proposal proposed by any sub-agent receives acceptance responses from all conflicting parties, it indicates that the parties have reached a consensus based on the counter-proposal, and the negotiation is successful. The structured data (i.e., alternative implementation schemes) corresponding to the accepted counter-proposal is transformed into meta-output data, serving as the final resolution result of this conflict. This data then directly enters the component tree assembly process without further adjudication by the conflict adjudication sub-agent.

[0109] Conversely, if no consensus is reached after more than a preset threshold number of negotiation rounds, the negotiation mode is exited, and the conflict adjudication sub-agent executes the adjudication. The preset threshold number of rounds can be flexibly configured, for example, from 2 to 5 rounds. After exiting the negotiation mode, the meta-output data of both (or more) conflicting parties, along with the exchanged historical counter-proposal records, are submitted to the conflict adjudication sub-agent. The conflict adjudication sub-agent then executes the final adjudication according to the adjudication calculation process described above (i.e., weighted calculation based on comprehensive credibility scores).

[0110] Optionally, when the number of negotiation rounds does not exceed the preset round threshold but the counter-proposals of both parties gradually approach each other (i.e., the difference between counter-proposals in adjacent rounds continues to decrease), the upper limit of the preset round threshold is allowed to be increased (e.g., 1-2 rounds are added on the basis of the original threshold), giving the two parties more rounds to exchange ideas to reach a consensus and further improving the negotiation success rate.

[0111] This application improves the accuracy of conflict resolution by setting up a negotiation-based processing flow. Specifically, sub-agents trace back along the target hierarchy to higher levels of abstraction to generate counter-proposals, seeking consensus at a more macro-level and avoiding zero-sum games at the implementation level. This results in resolutions that better meet the user's overall needs. Simultaneously, it reduces the uncertainty of adjudication. In negotiation mode, the resolution result is reached autonomously by the target sub-agents involved in the conflict, rather than being unilaterally decided by the conflict adjudication sub-agent. Because each target sub-agent has a deeper understanding of the business logic and data within its respective domain, its autonomously reached consensus is usually more in line with business reality than externally adjudicated solutions, reducing the deviation between the adjudication result and actual business logic.

[0112] Furthermore, it ensures system robustness. If no consensus is reached after a negotiation round times out, it automatically exits the negotiation mode and degrades to the adjudication mode, ensuring that a resolution result is output under any circumstances, preventing system crashes or response timeouts due to negotiation process anomalies. It also retains scalability; the negotiation round threshold is configurable, and the upper limit can be dynamically increased as both sides' counter-proposals approach each other. This allows the negotiation mechanism to flexibly adjust according to the complexity and real-time requirements of different scenarios, balancing resolution quality and response efficiency.

[0113] The following example illustrates this scenario. For instance, the scenario is set as follows: A target user, "Zhang San," inquires about car purchase options in a car sales dialogue system. The car-pushing agent recommends a car model X that fits his budget, but the financial agent calculates a down payment ratio (30%) that exceeds Zhang San's short-term payment capacity (Zhang San currently has limited available cash). This creates a conditional conflict regarding the "down payment ratio," and the severity of the conflict is assessed as "medium." The system triggers a negotiation mode.

[0114] In the first round of negotiation (backtracking and generating counter-proposals), the pusher agent backtracks upwards along the target hierarchy: the current conflict node matches a financial solution → the mid-level node recommends a suitable car model → the upper-level target node meets the user's travel needs. Based on this upper-level target node, the pusher agent generates a counter-proposal: "Recommend car model Y with a lower price, whose down payment ratio can be reduced to 20%." The finance agent backtracks upwards along the target hierarchy: the current conflict node calculates the down payment ratio → the mid-level node provides a financial solution → the upper-level target node meets the user's funding needs. Based on this upper-level target node, the finance agent generates a counter-proposal: "Keep car model X unchanged, but extend the loan term from 24 months to 48 months, reducing the down payment ratio to 20%."

[0115] Both parties exchange counter-proposals via a message bus. In the second round of negotiation (evaluation and response), after receiving the counter-proposal from the financial agent, the push-car agent assesses and determines that model X better meets Zhang San's car needs, and extending the loan term is within Zhang San's long-term repayment capacity. Therefore, the push-car agent sends an acceptance response. After receiving the counter-proposal from the push-car agent, the financial agent assesses and determines that model Y is also within Zhang San's budget, but model X has a higher overall score, and the financial agent's solution (extending the loan term) already meets the down payment requirement. Therefore, the financial agent also sends an acceptance response.

[0116] Consensus reached: Both parties accepted the financial sub-agent's counter-proposal (model X + 48-month loan term + 20% down payment), and this counter-proposal was determined as the resolution of this conflict. The system then proceeds to the component tree assembly process, presenting this resolution as the final recommended solution to user Zhang San.

[0117] In the above scenario, if after 5 rounds of negotiation (assuming a preset threshold of 5 rounds), the push-cart agent still insists on recommending car model Y, and the financial agent still insists on maintaining car model X but extending the loan term, and the two parties cannot reach a consensus, then the system exits the negotiation mode and submits the meta-output data of both parties to the conflict resolution agent. The conflict resolution agent performs a weighted calculation based on decision data such as confidence level, user profile preference weight, and historical interaction preferences to determine the retained recommendation scheme and ultimately generate the decision result.

[0118] After obtaining the processed association graph, the decomposed node data is assembled into a box computing component tree (i.e., the target component tree) based on the processed association graph and the set business priority order (e.g., cart priority is greater than financial priority, and financial priority is greater than equity priority).

[0119] The box-based component tree can be an abstract syntax tree used to describe conversational UIs. It uses a nested JSON structure to specify which UI elements are displayed within the dialog box, how they are arranged, and the corresponding interactive behaviors of each element. Node types in the component tree include, but are not limited to: card containers, carousels, tables, button groups, forms, and sliders.

[0120] When constructing the target component tree, you can select an appropriate layout template (such as a comparison table template, a recommendation details + calculator template, a tab pagination template, etc.) based on the current intent and the current user state. Based on the processed association graph, the resolved node data is filled into the corresponding component slots according to the template structure to generate a complete component tree JSON structure.

[0121] For S106, based on the constructed target component tree, generate and display the response content corresponding to the current session content. The output content format can include various media formats, as well as operation links, tools, and cards. The output method for the response content can be: server-sent event (SSE) streaming output, static output, etc.

[0122] This application, based on box computing logic, integrates and restates the meta-output data from multiple target sub-agents to generate and display response content, maximizing the completion of all interactions within the dialog box, effectively and efficiently answering user questions, reducing user effort, and improving the overall interactive experience.

[0123] Each agent in this application can call external APIs in real time to obtain dynamic data. Combined with SSE streaming output and user interaction tools, the meta-answers (i.e., meta-output data) returned by multiple sub-agents are assembled into a box computing component tree, enabling all interactions to be completed within the dialog box and improving the user experience. This application constructs a scalable, personalized, and real-time intelligent dialog computing system, solving the technical problems of traditional frameworks in business expansion, data fusion, and contextual coherence.

[0124] During implementation, the response content can be tested for content risk control first, and the response content can be displayed after the test is passed.

[0125] Optionally, generating and displaying the response content corresponding to the current session content based on the target component tree includes: generating and displaying the response content corresponding to the current session content based on the target component tree and the set response strategy; wherein the response strategy includes at least one of the following: empathy support strategy, concern elimination strategy, and solution provision strategy.

[0126] The empathy-based approach involves acknowledging and responding to the user's emotional state before addressing their concerns, thus reducing their defensiveness. For example, if user anxiety is detected, the strategy might begin by saying, "I understand your struggle when choosing a car; it's normal to want to control your budget but not settle," before presenting recommended content. The concern-relief strategy involves proactively providing reassuring information or evidence to alleviate user anxieties about risks such as overpaying, being ripped off, or the car not being suitable. For example, when showcasing financing options, it might include a "2-year 0% interest" policy and a "7-day no-questions-asked cancellation" guarantee. The solution-providing strategy involves simplifying complex issues into "one-click operations" or "ready-made options," reducing the effort required from the user. For example, it might directly display a decision card with the recommendation "Comprehensive Recommendation: Car B," instead of requiring the user to compare numerous parameters themselves.

[0127] During implementation, the component tree is pushed to the front end piece by piece through SSE streaming output. The front-end component tree rendering engine performs a deep traversal of the JSON structure, matches the corresponding UI component from the local component library based on the field type, injects the props data, and then mounts and renders it to the document object model (DOM) node of the dialog box, so that the content seen by the user is presented step by step without any blank screen waiting, resulting in a smooth experience.

[0128] Optionally, after generating and displaying the response content corresponding to the current session content, the method further includes: generating and displaying at least one guiding question based on the response content and the unfinished decisions in the decision chain; receiving implicit feedback signals from the target user regarding the guiding questions, wherein the implicit feedback signals include one or more of the following: user clicks on recommended cards, page dwell time, modification / deletion of recommended options; the implicit feedback signals are used to adjust the dynamic weight coefficients of the sub-agent, and / or, the user profile preference data of the target user.

[0129] After generating and displaying the response content, at least one guiding question is dynamically generated and displayed based on the response content and the unresolved decisions in the decision chain. This guiding question is not generated based on a fixed template, but rather dynamically generated based on the user's long-term memory predictions of the unresolved decision chain.

[0130] The incomplete decision chain refers to the sequence of key decision steps that a user has not yet completed in order to achieve their ultimate goal. For example, a general decision chain directed graph can be pre-constructed for each business domain (such as the car purchase decision chain: budget confirmation → model selection → parameter comparison → financial calculation → test drive appointment → placing a deposit). Based on the content of the current dialogue round, historical operation records, and known conditions in the user profile, the user's current position on the decision chain can be determined. The first incomplete node along the directed graph with all its prerequisites completed is taken as the predicted incomplete decision, and then the larger model transforms it into a natural language generation guiding question.

[0131] In implementation, each sub-agent can have its own defined behavioral observation metrics (e.g., the cart sub-agent observes the click-through rate and dwell time of vehicle model cards, the financial sub-agent observes the click-through rate of the credit limit test button and the form completion rate, etc.). These metrics are then weighted and fused to calculate the actual satisfaction index for each sub-agent. This satisfaction index is used as a reward signal for reinforcement learning, updating the dynamic weight coefficients in the main agent's route matching score. The updated dynamic weight coefficients are persistently stored independently by user identifier. When the same user initiates a subsequent session, the main agent loads that user's dynamic weight coefficients and recalculates the matching score, achieving personalized adaptive evolution of the routing strategy.

[0132] For example, after generating the response content, data such as the user's long and short-term memory can be combined to fully understand the user's questions and underlying concerns, and to guide the conversation and make recommendations, such as generating guiding questions like "Are you still worried about issues related to ownership transfer and after-sales service?"

[0133] This application also provides an adaptive adjustment mechanism for conflict resolution strategies based on implicit feedback. This mechanism automatically identifies the effectiveness of conflict resolution strategies by monitoring users' subsequent behavior regarding the conflict resolution results, and negatively marks and downgrades the priority of ineffective or inefficient strategies, thereby continuously optimizing the matching effect of conflict resolution strategies without manual intervention.

[0134] This mechanism has a close synergy with the conflict resolution method described above (i.e., the mapping rules for determining resolution strategies based on conflict type and severity). The aforementioned conflict resolution method provides the initial conflict resolution strategy matching and execution logic, while this mechanism dynamically adjusts the matching priority of conflict resolution strategies based on user feedback after execution. Together, they form a complete closed loop of strategy execution, effect monitoring, and strategy optimization. As the system runs longer, the matching of conflict resolution strategies will increasingly approximate the true preference distribution of the target user group, achieving adaptive evolution that becomes more accurate with use.

[0135] By introducing implicit feedback signal acquisition and dynamic weight adjustment mechanisms, user preferences can be learned from user clicks, dwell times, modifications, and other behaviors, enabling personalized adaptive evolution of routing strategies and achieving a technical effect of becoming more accurate with use.

[0136] Optionally, the method further includes: when the implicit feedback signal indicates that the conflict resolution result is rejected or modified by the user, adding a negative mark to the resolution strategy corresponding to the combination of conflict type and severity that triggered this conflict resolution process; when the negative marks of the same resolution strategy accumulate to a preset number, lowering the matching priority of the resolution strategy.

[0137] In this application, after the conflict resolution sub-agent completes the conflict resolution process, the resolution result is presented to the target user in the form of a component tree. During the interaction between the target user and the component tree, the target user's interaction behavior is continuously monitored by receiving implicit feedback signals to determine the target user's acceptance level of the resolution result.

[0138] Implicit feedback signals include, but are not limited to, one or more of the following: user clicks on recommended cards, page dwell time, and modification / deletion of recommended options. These implicit feedback signals are used to determine whether the conflict resolution result has been rejected or modified by the user. Furthermore, when implicit feedback signals indicate that the conflict resolution result has been rejected or modified by the user (e.g., the user manually changes the system-recommended loan term or down payment ratio), a negative marker is added to the resolution strategy corresponding to the combination of conflict type and severity that triggered this conflict resolution process. When the cumulative number of negative markers for the same resolution strategy reaches a preset number, the matching priority of that resolution strategy is lowered, so that subsequent similar conflicts will no longer preferentially use that resolution strategy.

[0139] For example, an implicit feedback signal can be determined to indicate that the conflict resolution result has been rejected or modified by the user when one of the following behaviors exists: First, the user modified the recommended options, such as adjusting the loan term or down payment ratio based on the system-recommended financing plan, or modifying the configuration parameters of the system-recommended car model. This modification indicates that the system's recommendations did not fully match the user's actual needs, and the user needs to manually correct these errors to compensate for the system's decision-making bias.

[0140] Second, the user deleted a recommended option, such as removing a car model or financing plan recommended by the system. This deletion indicates that the user has a clear negative attitude towards the recommended option.

[0141] Third, the user triggered follow-up questions on the page displaying the resolution results, and the questions concerned the solutions that were ruled out. For example, if a user asks, "Why wasn't model A recommended?", it indicates that the user has doubts about the reasonableness of the decision and needs the system to provide further explanation or reassessment.

[0142] Fourth, if a user stays on the resolution results display page for less than the preset first time threshold (e.g., 3 seconds) and does not click on any recommended cards, it indicates that the user lacks interest in the resolution results and quickly leaves.

[0143] Fifth, the user manually selected the option corresponding to the eliminated solution. For example, in an interactive selection presentation, if the user chooses an option other than the system's recommended one, it indicates that the system's decision does not reflect the user's true preferences.

[0144] When an implicit feedback signal indicates that the conflict resolution result has been rejected or modified by the user, a negative flag is added to the resolution strategy corresponding to the combination of conflict type and severity that triggered this conflict resolution process. For example, the negative flag may include the cumulative count and the latest cumulative timestamp corresponding to the resolution strategy.

[0145] During implementation, each time a user denies or corrects a resolution result, the cumulative count of the corresponding negative flag is incremented by one based on the conflict type, severity, and resolution strategy identifier used in the current conflict resolution process, and the latest cumulative timestamp is updated.

[0146] When multiple correction operations are repeatedly performed on the same resolution result in the same session for the same user, only one negative mark can be added to the resolution policy. This avoids multiple superimposed negative marks caused by the user's operation behavior in a single conflict resolution, thereby preventing the excessive impact of occasional extreme behavior of a single user on the policy matching priority.

[0147] When the cumulative negative flags for the same resolution strategy reach a preset number, the matching priority of that resolution strategy is lowered. For example, the preset number of times can be set based on the sensitivity requirements of the business scenario. For business scenarios with high sensitivity to user experience, the preset number of times can be set to a lower value (e.g., 3 times) to allow the system to adjust the resolution strategy more quickly; for business scenarios with high requirements for system stability, the preset number of times can be set to a higher value (e.g., 10 times) to avoid frequent strategy changes due to a small amount of noise feedback.

[0148] Matching priority refers to the order in which resolution strategies are matched within a predefined set of resolution strategies. Matching priority determines the probability or order in which each resolution strategy is matched when the system encounters the same combination of conflict type and severity again. When the matching priority of a resolution strategy is lowered, its matching order within that combination shifts to the next lower priority; the system will prioritize matching other resolution strategies whose priority has not been lowered or which have the fewest negative flags.

[0149] The technical effects of the aforementioned negative marking and matching priority downgrading mechanism are reflected in the following aspects: It enables adaptive evolution of the conflict resolution strategy. Traditional conflict resolution methods rely on manually pre-set static rules. When the rules are unreasonable or do not meet user expectations, manual intervention is required for adjustment, resulting in high maintenance costs and slow response times. This application, through the accumulation of negative markings and the priority downgrading mechanism, enables the system to automatically adjust the matching preferences of the conflict resolution strategy based on real user feedback, continuously optimizing without manual intervention and significantly reducing system operation and maintenance costs. Furthermore, the initial matching priority of the conflict resolution strategy is preset based on business experience, which may deviate from the actual preferences of the target user group. Through this mechanism, the system can gradually converge to the true distribution of user preferences during operation, effectively solving the problem of inaccurate strategy matching during the cold start phase.

[0150] It also enables personalized adaptation. Because negative tags are stored separately for each user, the system can automatically perceive the differences in preferences among different user groups through the cumulative distribution of negative tags. For example, younger users may prefer flexible installment plans, while middle-aged users may prefer lower down payment plans. The system maintains separate negative tags for different user groups, resulting in a differentiated distribution of the matching priority of mitigation strategies across different user groups, thus achieving personalized mitigation strategy adaptation. This also ensures the system's robustness. The preset number of negative tags prevents excessive interference from occasional negative behaviors of individual users on the system's strategy. Negative tags need to accumulate to a preset number before triggering a priority downgrade, avoiding frequent policy fluctuations caused by single abnormal behaviors and ensuring the stability of the system's decisions.

[0151] This mechanism has a close synergy with the aforementioned resolution strategy decision table. Specifically, the preset resolution strategy set described above provides a combination of conflict types and severity levels, and an initial mapping relationship between them and the resolution strategies, serving as the basis for resolution execution. This mechanism, based on the negative flag accumulation-driven matching priority downgrading, is a dynamic optimization mechanism for the aforementioned mapping relationship. Its output (the adjusted matching priority) serves as feedback input to the aforementioned strategy matching process, enabling the resolution strategy set to continuously optimize with the accumulation of actual operational data, rather than being a static and unchanging rule table.

[0152] The above-mentioned collaborative relationship can be summarized as follows: first, an initial solution is provided; then, user feedback signals are collected; and finally, priorities are adjusted based on the feedback. The three functions complement each other and are linked by data to form a complete closed-loop control loop of execution, monitoring, feedback, and adjustment.

[0153] The following example illustrates this with specific application scenarios. For instance, in a used car trading platform, for the combination of conflict type and severity of "conditional conflict + medium severity", the system's preset resolution strategy is: both ends of the node are retained, and an interactive selection presentation method is used, allowing the user to choose for themselves.

[0154] During the first three conversations on the platform, the system encountered a conditional conflict (moderate severity) of "limited budget vs. high configuration requirements" three times consecutively, and adopted the interactive selection presentation strategy described above in each instance. In all three uses, the user's implicit feedback signals were: page dwell time less than 3 seconds, no clicks on any option, and ultimately, typing "Is there anything cheaper?" on the platform, indicating user dissatisfaction with the outcome of the interactive selection presentation strategy.

[0155] Negative marker accumulation: In the three uses mentioned above, each time the user's implicit feedback signal indicates that the conflict resolution result was rejected by the user (such as the page quickly exiting or entering negative information), the system adds a negative marker to the third resolution strategy corresponding to "conditional conflict + medium severity". The negative markers accumulate to a preset number (e.g., 3 times).

[0156] Priority Downgrade: The system detects that the negative flag of the third resolution strategy under the combination of "conditional conflict + medium severity" has accumulated to the preset number of times, and performs a matching priority downgrade operation.

[0157] When a user uses the platform for the fourth time, if the platform encounters a conditional conflict with the same combination of conflict type and severity, the system reads the updated matching priority data when matching resolution strategies. The third resolution strategy is no longer in the highest priority position. The system detects that the second resolution strategy (i.e., business priority adjudication, exclusive presentation) has the fewest negative markers under this combination and has the highest matching priority. Therefore, the system automatically adopts the second resolution strategy to perform conflict resolution. Thus, the system replaces the strategy of "letting the user choose" with a strategy of "quickly providing a definite answer".

[0158] If, during the fourth use, the user's implicit feedback signals manifest as clicking on a system-exclusive recommended card and staying on the page for a duration far exceeding a preset threshold, it indicates that the new resolution strategy has produced resolution results that better meet the user's expectations. Through the above process, this application can automatically complete the optimization and switching of the resolution strategy based on implicit user feedback, without manual intervention.

[0159] The conflict resolution scheme in this application forms a complete closed loop of resolution execution, effect monitoring, and strategy evolution. First, a decision table for conflict resolution is defined, and mapping rules for matching resolution strategies are determined based on the conflict type and severity. After resolution, user behavior is monitored, and implicit feedback signals are collected as a source of feedback on the resolution effect. Then, based on the collected implicit feedback signals, the defined resolution strategies are dynamically adjusted. For example, when the negative feedback of a certain resolution strategy accumulates to a preset number of times under a specific combination of conflict type and severity, the matching priority of that strategy is lowered, so that subsequent similar conflicts will be preferentially matched with other resolution strategies.

[0160] The aforementioned closed loop requires no human intervention, enabling continuous optimization of conflict resolution strategies. Through this collaborative design, the system can continuously accumulate user preference data during operation, making the resolution strategies increasingly aligned with the actual needs of the target users, ultimately improving the quality of dialogue content and user satisfaction.

[0161] See Figure 2 As shown, combined with Figure 2 The method proposed in this application is illustrated by example.

[0162] First, at the user input layer, the system can receive multimodal raw conversation content (such as text, voice, images, videos, etc.) input by the target user. After passing the content risk control check, the system uses the acquired user attribute data, historical dialogue data, and the determined current user state to semantically rewrite the raw conversation content and generate complete and unambiguous current conversation content.

[0163] Next, at the Agent routing layer, the Agent extracts domain tags from the rewritten session content, generates the current intent vector, and simultaneously reads the capability matrix data, current load status, and lock status of dependent resources of each sub-agent. Based on this, the Agent calculates the matching score and expected response time of each sub-agent, selects the combination of sub-agents that maximizes the global benefit as the routing target according to the set routing benefit function, obtains multiple target sub-agents, and distributes the encapsulated metadata to each target sub-agent in parallel through the message bus.

[0164] Then, in the parallel processing layer of the sub-agents, each target sub-agent independently performs analysis, calling external APIs as needed to obtain dynamic knowledge base data, static knowledge base data, page links or user operation cards and other interactive data, generating structured meta-output data, along with confidence scores, timestamps of the referenced data sources, and evidence chain identifiers.

[0165] Subsequently, in the conflict resolution layer, the main agent uses all meta-output data as nodes and connects different nodes by conflict types, constructing a conflict association graph. The conflict adjudication sub-agent then processes the conflict association graph based on the adjudication data, such as performing deletion, retention, data adjustment, or merging operations on nodes, generating a processed association graph.

[0166] Subsequently, in the output and interaction layer, based on the processed association graph and business priority order, the resolved node data is assembled into a box computing component tree (i.e., the target component tree), and then the response content is generated and displayed based on the target component tree.

[0167] Finally, in the closed-loop optimization layer, based on the response content and unresolved decisions in the decision chain, guiding questions are dynamically generated and displayed to the user, receiving implicit feedback signals from the user regarding these guiding questions. Based on these implicit feedback signals, the agent's dynamic weight coefficients and user profile preference data can be adjusted, and the adjusted dynamic weight coefficients and user profile preference data are stored in the user's dedicated weight library for loading and use in subsequent sessions.

[0168] Corresponding to the aforementioned embodiments of the multi-agent-based dialogue content determination method, this application also provides embodiments of a multi-agent-based dialogue content determination apparatus. Figure 3A schematic diagram of the multi-agent dialogue content determination device provided in this application specifically includes: The first determining module 301 is used to determine the current session content of the target user; The first generation module 302 is used to extract tags from the current session content using the main agent and generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, it determines at least one target sub-agent that matches the current session. The second determining module 303 is used to send the encapsulated meta-information to the target sub-agent and control the target sub-agent to determine the meta-output data based on the meta-information; wherein the meta-information includes the user profile data of the target user and the current session context data; The construction module 304 is used to construct a conflict association graph by taking the meta-output data corresponding to the target sub-agent as nodes and using the conflict types between different nodes as connecting edges; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. The second generation module 305 is used to control the conflict adjudication sub-agent to perform conflict resolution processing on the conflict association graph according to the conflict type and severity of the connection edge and the acquired adjudication data, generate a processed association graph, and construct a target component tree according to the processed association graph and the set business priority order; the adjudication data includes the meta-output data and confidence level associated with each node in the conflict association graph, as well as the user attribute data and historical interaction preference data of the target user; The third generation module 306 is used to generate and display the reply content corresponding to the current session content based on the target component tree.

[0169] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0170] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0171] This application also provides a computer-readable storage medium storing a computer program that can be used to execute the multi-agent-based dialogue content determination method described in the above embodiments.

[0172] This application also provides a computer device, see [link to relevant documentation] Figure 4 The diagram shown illustrates the structure of the computer device provided in this application. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the multi-agent-based dialogue content determination method described in the above embodiments. Of course, besides software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM, optical memory, etc.

[0174] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0177] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0178] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, and optical disc read-only memory (CD). ROM, digital multifunction optical disc (DVD) or other optical storage, magnetic cassette tape, magnetic disk storage or other magnetic storage devices or any other non-transfer media, may be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transient media, such as modulated data signals and carrier waves.

[0179] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0180] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0181] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0182] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for determining dialogue content based on multi-agent communication, characterized in that, The method includes: Determine the current session content of the target user; The main agent extracts tags from the current session content to generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, at least one target sub-agent matching the current session is determined. The encapsulated metadata is sent to the target sub-agent, and the target sub-agent is controlled to determine the metadata output data based on the metadata; wherein the metadata includes the user profile data of the target user and the current session context data; Using the meta-output data corresponding to the target sub-agent as nodes, and the conflict types between different nodes as connecting edges, a conflict association graph is constructed; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. Based on the conflict type and severity of the connecting edges, and the acquired adjudication data, the conflict adjudication sub-agent is controlled to perform conflict resolution processing on the conflict association graph, generating a processed association graph, and constructing a target component tree based on the processed association graph and the set business priority order; the adjudication data includes the meta-output data and confidence level associated with each node in the conflict association graph, as well as the user attribute data and historical interaction preference data of the target user; Based on the target component tree, generate and display the reply content corresponding to the current session content.

2. The method according to claim 1, characterized in that, After generating and displaying the reply content corresponding to the current session content, the method further includes: Based on the response content and the unresolved decisions in the decision chain, generate and display at least one guiding question; The system receives implicit feedback signals from the target user in response to the guiding question, wherein the implicit feedback signals include one or more of the following: user clicks on recommended cards, page dwell time, and modification / deletion of recommended options; the implicit feedback signals are used to adjust the dynamic weight coefficients of the sub-agent, and / or the user profile preference data of the target user.

3. The method according to claim 1, characterized in that, The method further includes: Obtain the current load status of each sub-agent and the lock status of the resources it depends on; Based on the current load state of the sub-agent and the lock state of the resources it depends on, determine the expected response time of the sub-agent; The step of determining at least one target sub-agent for the current session based on the current intent vector and the capability matrix data corresponding to each sub-agent includes: Based on the current intent vector and the capability matrix data corresponding to each sub-agent, calculate the matching score corresponding to the sub-agent; Based on the matching score corresponding to the sub-agent, the expected response time, and the dynamic weight coefficient, at least one target sub-agent is determined for the current session; wherein when there are multiple target sub-agents and multiple target sub-agents call the same dependent resource, a resource reservation and time slice round-robin mechanism is used for parallel scheduling.

4. The method according to claim 1, characterized in that, Determining the current session content of the target user includes: Receive the raw session content input by the target user; Obtain the target user's user attribute data and historical dialogue data, and determine the target user's current user status; Based on the user attribute data, the historical dialogue data, and the current user status, the original conversation content is adjusted to generate the current conversation content for the target user.

5. The method according to claim 4, characterized in that, After receiving the original session content, the method further includes: The original session content is subjected to content risk control detection to obtain the risk control detection results; When the risk control detection result indicates the presence of illegal content, preset acceptance content is generated and displayed.

6. The method according to claim 1, characterized in that, The control of the target sub-agent to determine the metadata output data based on the metadata includes: The target sub-agent is controlled to analyze the metadata and generate analysis results; When the analysis results indicate that an external interface needs to be called, a call request is initiated, wherein the call request is used to obtain the required interactive data from the target data structure; the interactive data includes one or more of the following: dynamic knowledge base data, static knowledge base data, page links, and user terminal operation cards; Based on the interaction data and the analysis results, meta-output data is generated.

7. The method according to claim 1, characterized in that, The step of controlling a conflict resolution sub-agent to perform conflict resolution processing on the conflict association graph based on the conflict type and severity of the connecting edges, and the acquired resolution data, to generate a processed association graph, includes: Based on the combination of the conflict type and the severity, a corresponding target resolution strategy is matched from a preset resolution strategy set, wherein the target resolution strategy includes node retention rules and presentation rules; Based on the adjudication data, adjudication calculations are performed on the part of the target resolution strategy involving node selection to determine the resolution results of the nodes at both ends of the connecting edge; According to the node retention rules and the resolution results, delete, retain or merge the nodes at both ends of the connection edge to generate the resolved node data; According to the presentation rules, determine the presentation method of the resolved node data; The resolved node data is written into the conflict association graph to generate the processed association graph.

8. The method according to claim 7, characterized in that, The step of matching a corresponding target resolution strategy from a preset resolution strategy set based on the combination of the conflict type and the severity includes: If the conflict type is numerical and the severity is high, the first resolution strategy for matching includes: the node retention rule is to merge the meta-output data of the two nodes into a single node, and the value in the meta-output data is taken as the weighted average of the two; the presentation rule is exclusive presentation. If the conflict type is numerical and the severity is medium or low, the matching second resolution strategy includes: the node retention rule is to calculate the deviation rate of the numerical values ​​in the output data of the two node nodes; if the deviation rate falls within a preset tolerance range, the weighted average of the two is taken and merged into a single node; if the deviation rate does not fall within the preset tolerance range, one node is retained and the other node is deleted according to the adjudication data; the presentation rule is exclusive presentation. If the conflict type is conditional, the third resolution strategy includes: the node retention rule is to extract the trigger condition set from the output data of the nodes at both ends, match the trigger condition set with the user attribute data and the current user status, and retain the node corresponding to the condition that is matched, while deleting the node that does not match the condition; if both nodes match the condition or neither matches the condition, then in the case of high severity, the node with the higher business priority is retained, and in the case of medium or low severity, both nodes are retained; the presentation rule is exclusive presentation in the case of high severity, and interactive selection presentation in the case of medium or low severity. If the conflict type is logical and the severity is high, the matching fourth resolution strategy includes: the node retention rule is to extract logical assertions from the meta-output data of the two end nodes respectively, obtain the evidence chain identifier corresponding to each logical assertion, retain the node of the party with a traceable evidence chain and delete the other party's node; the presentation rule is exclusive presentation. If the conflict type is logical and the severity is medium or low, the matching fifth resolution strategy includes: the node retention rule is to retain the node of the party with a traceable chain of evidence and delete the other party's node; the presentation rule is to attach the logical assertion of the deleted node and the corresponding reason for the missing chain of evidence to the meta-output data of the retained node as supplementary explanation, and present it as supplementary explanation.

9. The method according to claim 7, characterized in that, The step of performing adjudication calculations on the node selection portion of the target resolution strategy based on the adjudication data, and determining the resolution results of the nodes at both ends of the connecting edge, includes: When the node retention rule of the target resolution strategy indicates that one party's node should be retained and the other party's node should be deleted based on the adjudication data, the following steps are performed: Obtain the confidence level of the meta-output data corresponding to the nodes at both ends of the connection edge, the business priority of the sub-agent to which it belongs, and the timestamp of the referenced data source; Obtain the user profile preference data of the target user, and calculate the semantic matching degree between each meta output data and the user preference tag; Obtain the target user's selection results for similar conflicts in historical interactions, and calculate the matching frequency of each meta-output data with the user's historical selection preferences; Obtain the current user state of the target user and calculate the applicability matching degree between each meta output data and the current user state; The comprehensive credibility score of each node is calculated by weighting the confidence level of the meta-output data, the semantic matching degree, the matching frequency, the applicability matching degree, the business priority, and the timeliness determined based on the timestamp. The node with the higher overall credibility score is selected as the node to be retained, and the node with the lower overall credibility score is selected as the node to be deleted, which is the resolution result.

10. The method according to claim 7, characterized in that, The process involves performing deletion, retention, or merging operations on the nodes at both ends of the connecting edge according to the node retention rules and the resolution results, generating resolved node data, and writing the resolved node data into the conflict association graph to generate a processed association graph, including: If the node retention rule indicates that the nodes should be merged into a single node, a new node is generated, the meta-output data of the two nodes are merged and written into the new node, the original nodes at both ends are deleted, and the new node is written into the conflict association graph. If the node retention rule indicates that one side should be retained and the other side should be deleted, then the node to be retained in the resolution result should be retained, and the node to be deleted in the resolution result should be deleted from the conflict association graph. If the node retention rule indicates that both ends of the node should be retained, then the nodes at both ends should remain unchanged, and the connection edge between the two ends of the node should be marked as resolved.

11. The method according to claim 1, characterized in that, The step of generating and displaying the reply content corresponding to the current session content based on the target component tree includes: Based on the target component tree and the set response strategy, the response content corresponding to the current session content is generated and displayed; wherein the response strategy includes at least one of the following: empathy support strategy, concern elimination strategy, and solution provision strategy.

12. A device for determining dialogue content based on multi-agent communication, characterized in that, The device includes: The first determination module is used to determine the current session content of the target user; The first generation module is used to extract tags from the current session content using the main agent and generate a current intent vector; based on the current intent vector and the capability matrix data corresponding to each sub-agent, it determines at least one target sub-agent that matches the current session. The second determining module is used to send the encapsulated meta-information to the target sub-agent and control the target sub-agent to determine the meta-output data based on the meta-information; wherein the meta-information includes the user profile data of the target user and the current session context data; The construction module is used to construct a conflict association graph by taking the meta-output data corresponding to the target sub-agent as nodes and using the conflict types between different nodes as connecting edges; wherein the conflict types include: numerical conflict, conditional conflict, and logical conflict. The second generation module is used to control the conflict adjudication sub-agent to perform conflict resolution processing on the conflict association graph based on the conflict type and severity of the connection edge and the acquired adjudication data, generate a processed association graph, and construct a target component tree based on the processed association graph and the set business priority order; the adjudication data includes the meta-output data and confidence level associated with each node in the conflict association graph, as well as the user attribute data and historical interaction preference data of the target user; The third generation module is used to generate and display the reply content corresponding to the current session content based on the target component tree.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-11.

14. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor performs the steps of the method according to any one of claims 1-11.

Images