A multi-agent fusion scheduling method, system, device, equipment and medium

Abstract

The present application relates to the technical field of artificial intelligence customer service, and discloses a multi-agent fusion scheduling method, system, device, equipment and medium, the present application obtains scheduling planning information containing each subtask by analyzing and splitting service object request, so that composite appeal can be executed in order according to scheduling planning; by matching each subtask to the corresponding agent and obtaining context information, the collaborative call and information sharing of multi-agent are realized; by performing security check on data access in the execution process, the data leakage risk is eliminated; by fusing the business processing results of each agent to generate a unified reply, the problem of scattered and inconsistent content of multiple replies is avoided, and the efficiency and output quality of multi-agent collaborative scheduling are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence customer service technology, specifically to a multi-agent fusion scheduling method, system, device, equipment, and medium. Background Technology

[0002] Currently, intelligent customer service systems driven by large language models are widely used in industries such as power, e-commerce, and business travel. However, users often have multiple requests in a single consultation. Complex requests require the system to have multiple processing capabilities such as bill inquiry, work order generation, and fault reporting. Existing methods usually train an independent agent for each business scenario, with multiple agents responding to different user needs. However, existing methods have problems in actual operation, such as lack of a unified scheduling mechanism, isolation of the context of each agent, lack of hard isolation of data access, and risk of unauthorized query.

[0003] In existing methods, each agent operates independently. If a message contains multiple requests, the relevant agents can only be called sequentially, with the next agent starting only after the previous one has finished processing. Response time increases linearly with the number of requests, lacking a unified scheduling mechanism. The contexts of each agent are isolated. If a user has already provided their account number in the electricity bill agent, they need to be asked for the account number and address again when transferred to the repair agent, forcing the user to repeat the description. When multiple agents run concurrently, resources are preempted without order. VIP users' urgent requests are queued at the same time as ordinary users' inquiries, making it impossible to guarantee service quality. Furthermore, in existing methods, each agent shares the same context storage space. An agent from company A can read order and user information from company B. When an agent needs to generate database statements using natural language to query the database, it relies solely on prompts to limit the query scope. The generated database statements may omit service provider filtering conditions, leading to a full data table scan in a single query or extraction of confidential fields, causing data leakage. Summary of the Invention

[0004] This invention provides a multi-agent fusion scheduling method, system, device, equipment, and medium to solve the problems of complex requests being processed serially, low response efficiency, and data leakage in multi-service scenarios.

[0005] In a first aspect, the present invention provides a multi-agent fusion scheduling method, comprising: Obtain the service object request, parse and split the service object request to obtain scheduling planning information, which includes each subtask; Based on the scheduling planning information, each subtask is matched to a preset intelligent agent, and the context information corresponding to the service object's request is obtained; An intelligent agent is used to execute corresponding subtasks based on context information, perform security checks on data access during execution, and obtain business processing results based on context information and check results. Based on the business processing results, generate the fusion scheduling results.

[0006] This embodiment provides a multi-agent fusion scheduling method. By parsing and splitting service object requests, it obtains scheduling planning information containing each sub-task, enabling complex requests that originally required sequential processing to be executed in an orderly manner according to the scheduling plan, thus shortening the overall response time. By matching each sub-task to the corresponding agent and obtaining context information, it achieves parallel collaborative invocation and information sharing among multiple agents. By performing security verification on data access during execution, it intercepts unauthorized queries and extraction of confidential fields, eliminating the risk of data leakage. By fusing the business processing results of each agent to generate a unified response, it avoids the problems of fragmented and inconsistent responses from multiple sources, thereby improving the efficiency and output quality of multi-agent collaborative scheduling.

[0007] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, the service object request is parsed and decomposed to obtain scheduling planning information, including: Extract the original request text, service recipient level identifier, and business urgency from the service recipient's request; The original appeal text was semantically parsed and broken down into multiple sub-appeals; Identify the dependencies between the sub-requests, generate a task topology based on the dependencies, and the task topology includes the execution order of each sub-task; The service recipient level score is determined based on the service recipient level identifier, and the urgency score is determined based on the business urgency corresponding to each sub-task. The scheduling priority score of each subtask is obtained by weighting and summing the service object level score and the urgency score. The task topology and scheduling priority scores are combined to form scheduling planning information.

[0008] The beneficial effects of this preferred technical solution are as follows: by semantically parsing the original request text and splitting it into multiple sub-requests, and identifying the dependencies between the sub-requests to generate a task topology structure, the complex requests that originally needed to be processed serially can be distributed and executed in an orderly manner according to the topological order, and the topological order of dependent sub-tasks can be processed in parallel, shortening the overall response time in multi-request scenarios; at the same time, the scheduling priority is obtained by weighted summation of the service object level score and the urgency score, so that different service objects and urgent requests receive different processing resources, thereby improving the service quality assurance capability.

[0009] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, the rules for determining the service object level score and urgency score include: The rules for determining the service recipient level score are as follows: the first type of service recipient corresponds to the first score, the second type of service recipient corresponds to the second score, the third type of service recipient corresponds to the third score, and the fourth type of service recipient corresponds to the fourth score. The rules for determining the urgency score are as follows: the first type of request corresponds to the first score, the second type of request corresponds to the second score, the third type of request corresponds to the third score, the fourth type of request corresponds to the fourth score, the fifth type of request corresponds to the fifth score, and the sixth type of request corresponds to the sixth score.

[0010] The beneficial effects of this preferred technical solution are as follows: by formulating quantitative service object level scores and urgency score determination rules, different scores are assigned to requests of different importance and urgency, providing a clear and executable quantitative basis for scheduling priority calculation, avoiding the subjective arbitrariness of scheduling decisions, and improving scheduling fairness and service quality assurance capabilities.

[0011] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, the preset agents include business agents and natural language query agents; according to the scheduling planning information, each subtask is matched with a preset agent, including: Based on the scheduling priority scores from high to low and combined with the execution order, each subtask is matched to the business intelligence agent. Subtasks without dependencies in the task topology are matched in parallel, while subtasks with prerequisite dependencies wait for the prerequisite subtasks to be executed before matching.

[0012] The beneficial effects of this preferred technical solution are as follows: by matching according to the scheduling priority score from high to low and in combination with the execution order in the task topology, it ensures that high-priority subtasks get processing resources first, avoiding the blocking of high-level service object requests during concurrency; at the same time, subtasks without dependencies are matched in parallel, while subtasks with dependencies wait serially, thereby compressing the overall processing time to the greatest extent and improving the efficiency of concurrent processing while ensuring the correctness of business logic.

[0013] In a preferred embodiment of the multi-agent fusion scheduling method of the present invention, context information is stored in a context memory pool. The context memory pool is divided into fragments and isolated storage for the context information of each service provider according to both service provider identifier and business line identifier. Obtaining the context information corresponding to the service object request includes: A business intelligence agent initiates a context read request carrying an authentication vector, which includes the service provider identifier and business line identifier of the business intelligence agent. The authentication vector is matched with the fragment identifier vector of each context fragment in the context memory pool. If the match is successful, the context information in the corresponding context fragment can be read.

[0014] The beneficial effects of this preferred technical solution are as follows: by storing the context memory pool in a dual-dimensional partitioned manner based on the service provider identifier and the business line identifier, and by adopting an access control mechanism that matches the authentication vector with the partition identifier vector, it ensures that the business intelligence agent can only read the context information within its own service provider and business line scope. This effectively prevents unauthorized access to data across service providers and business lines from the storage architecture level, thus ensuring the secure isolation of customer privacy and business data in multi-service provider scenarios.

[0015] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, the method of matching the authentication vector with the fragment identifier vector of each context fragment in the context memory pool further includes: Calculate the cosine similarity between the authentication vector and the fragment identifier vector; When the cosine similarity is greater than or equal to the preset similarity threshold, the match is considered successful, and the context reading request is allowed. When the cosine similarity is less than the preset similarity threshold, the match is deemed unsuccessful, and the context reading request is blocked.

[0016] The beneficial effects of this preferred technical solution are as follows: by calculating the cosine similarity between the authentication vector and the fragmentation identifier vector, the abstract vector matching relationship is transformed into a quantifiable similarity value, and a judgment is made in combination with a preset similarity threshold, so that the matching process has a clear mathematical basis and a controllable judgment boundary, avoiding the uncertainty of permission judgment caused by fuzzy matching; when the matching fails, the read request is directly intercepted, eliminating unauthorized behavior from the access entry point, and further strengthening the security isolation effect of context information storage in multi-service scenarios.

[0017] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, an agent is employed to execute corresponding sub-tasks according to context information, perform security verification on data access during execution, and obtain business processing results based on context information and verification results, including: A business intelligence agent executes corresponding subtasks based on context information. When a subtask needs to query the database, the natural language query agent is invoked to convert the natural language query requirement into a raw database query statement. Three layers of security checks are performed on the raw database query statement. After the checks pass, the database query is executed and the query results are obtained. The business agent generates the business processing results based on the context information and the query results. When a subtask does not require querying the database, the business intelligence agent directly generates the business processing result based on the context information.

[0018] The beneficial effects of this preferred technical solution are as follows: by performing three-layer security checks on database query statements, it blocks cross-service queries, extraction of confidential fields, and unauthorized access to data tables. Compared with the traditional method that relies solely on prompt word constraints, it upgrades the security mechanism from soft constraints to hard interception, fundamentally eliminating the risk of data leakage caused by unauthorized queries. It covers both query-type and non-query-type sub-task scenarios, taking into account both functional integrity and data security.

[0019] As a preferred embodiment of the multi-agent fusion scheduling method of the present invention, generating a fusion scheduling result based on the business processing result includes: Summarize the business processing results and eliminate conflicting and duplicate content between different business processing results to obtain a standard business result; Based on the standard business results, generate a complete response text with a consistent tone, and use the complete response text as the fusion scheduling result.

[0020] The beneficial effects of this preferred technical solution are as follows: by summarizing the processing results of each business agent and eliminating conflicting and duplicate content, it integrates them into a standard business result. Then, based on the standard business result, a complete response text with a unified tone is generated, avoiding the problems of fragmented and inconsistent content from multiple responses. This ensures that the service recipient receives a complete, coherent, and accurate response at once, thereby improving the output quality and user experience of multi-agent collaborative services.

[0021] In a second aspect, the present invention provides a multi-agent fusion scheduling system, applied to a multi-agent fusion scheduling method of the first aspect or any corresponding embodiment thereof, the system comprising: The scheduling center is used to obtain service object requests, parse and break down the service object requests to obtain scheduling planning information, which includes each sub-task; and match each sub-task to a preset intelligent agent according to the scheduling planning information. The context memory pool is used to obtain context information corresponding to the service object request; The security verification engine is used to employ intelligent agents to execute corresponding subtasks based on context information, perform security verification on data access during the execution process, and obtain business processing results based on context information and verification results. The result fusion unit is used to generate fusion scheduling results based on the business processing results.

[0022] Thirdly, the present invention provides a multi-agent fusion scheduling device, the device comprising: The task splitting module is used to obtain service object requests, parse and split the service object requests to obtain scheduling planning information, which includes each sub-task. The scheduling and reading module is used to match each subtask to a preset intelligent agent according to the scheduling planning information, and to obtain the context information corresponding to the service object request; The business execution module is used to employ intelligent agents to execute corresponding sub-tasks based on context information, perform security verification on data access during the execution process, and obtain business processing results based on context information and verification results. The results fusion module is used to summarize business processing results and generate fused scheduling results.

[0023] Fourthly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform a multi-agent fusion scheduling method of the first aspect or any corresponding embodiment described above.

[0024] Fifthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute a multi-agent fusion scheduling method according to the first aspect or any corresponding embodiment described above.

[0025] In a sixth aspect, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute a multi-agent fusion scheduling method according to the first aspect or any corresponding embodiment described above. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of a multi-agent fusion scheduling according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first process of a multi-agent fusion scheduling method according to an embodiment of the present invention; Figure 3 This is a second flowchart illustrating a multi-agent fusion scheduling method according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the third process of a multi-agent fusion scheduling method according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a multi-agent fusion scheduling device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0030] As an optional application scenario of this invention, such as Figure 1 As shown, this invention provides a multi-agent fusion scheduling system, which includes: The scheduling center is used to obtain service object requests, parse and break down the service object requests to obtain scheduling planning information, which includes each sub-task; and match each sub-task to a preset intelligent agent according to the scheduling planning information. The context memory pool is used to obtain context information corresponding to the service object request; The security verification engine is used to employ intelligent agents to execute corresponding subtasks based on context information, perform security verification on data access during the execution process, and obtain business processing results based on context information and verification results. The result fusion unit is used to generate fusion scheduling results based on the business processing results.

[0031] The scheduling hub 101 is deployed on a cloud server and includes, but is not limited to, a composite intent parsing module, a DAG task splitter, a weighted priority calculation unit, and a distribution scheduler; the context memory pool 102 is a distributed memory storage cluster that stores the context information of each service provider in a dual-dimensional partitioned and isolated manner according to the service provider identifier and the business line identifier; the security verification engine 103 includes, but is not limited to, a service provider layer verification unit, a confidential field verification unit, and a data table permission verification unit; and the result fusion unit 104 includes a summarization unit and a response generation unit.

[0032] The scheduling hub 101 is communicatively connected to the context memory pool 102, the security verification engine 103, and the result fusion unit 104. Service objects initiate service requests via customer service hotlines, mini-programs, or online APP sessions. These requests are standardized by the access gateway and then sent to the scheduling hub 101. The scheduling hub 101 calls the composite intent parsing module to identify multiple sub-requests in the service object request, uses a DAG task splitter to identify the dependencies between these sub-requests and generate a task topology, and a weighted priority calculation unit to calculate the scheduling priority of each sub-task based on the service object's level score and urgency score. The dispatch scheduler distributes each sub-task to a preset agent according to its scheduling priority from high to low, combined with the execution order in the task topology. The context memory pool 102 receives context read requests carrying authentication vectors from business agents, matches the authentication vectors with the fragment identifier vectors of each context fragment in the context memory pool, and allows reading the context information in the corresponding context fragment after a successful match, returning the information to the business agent. The security verification engine 103... When a business intelligence agent executes a subtask and accesses data, it is invoked to perform three layers of security checks on the database query statement: the service-side layer checks force the concatenation of the current service-side's equality filtering conditions into the query conditions; the confidential field checks block the extraction of confidential fields such as electricity fees and mobile phone numbers; and the data table permission checks check whether the target data table belongs to the preset access range of the current business intelligence agent. Only after all three layers of checks pass can the database query be executed. Each business intelligence agent returns the business processing results to the result aggregator 104. The result aggregator 104 summarizes the return results of all business intelligence agents, eliminates conflicting and duplicate content between different business processing results, integrates the summarized results, generates a complete reply text with a unified tone, and pushes it to the service object's session end as the fusion scheduling result.

[0033] According to an embodiment of the present invention, a multi-agent fusion scheduling method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] It should be noted that embodiments of the present invention may also include a persistent storage layer, serving as the system's persistent data foundation to provide data support for upper-layer scheduling, verification, and execution. This persistent storage layer includes an atomic component library data table, three tables for the intelligent agent (i.e., the intelligent agent process master table, the intelligent agent node table, and the intelligent agent edge table), a service-side sharded context cold and hot database, and a database permission configuration table.

[0035] The atomic component library data table is used to store the metadata of all standardized atomic capability units, including five core atomic capabilities: intent recognition, work order generation, natural language to structured query language query, intelligent outbound calling, and quality inspection scoring. Each atomic component defines a unified input parameter, output parameter, and call interface specification. Business intelligence agents complete complex tasks by combining and calling different atomic components.

[0036] The three tables for intelligent agents are used to persistently store the complete process logic of each business intelligent agent, corresponding to the intelligent agent structure generated by dragging and dropping on the front-end canvas. Among them, the intelligent agent process master table stores the unique identifier, name, business line to which it belongs, three-element version number, and enabled status of the intelligent agent; the intelligent agent node table stores the type, configuration parameters, and canvas coordinate position of all functional nodes of the intelligent agent; and the intelligent agent edge table stores the flow connections and flow rule expressions between nodes.

[0037] The service provider's sharded context hot and cold database adopts a hot and cold separation architecture to store all context data. The hot database stores the service provider's sharded context in the process of a session, supporting real-time read and write. The cold database stores the historical archived context of ended sessions, which is used for subsequent revisits, data statistics and model optimization, and does not participate in real-time scheduling.

[0038] The database permission configuration table is the core configuration basis for the security verification of natural language to structured query language queries. It stores the data access permission matrix for each business intelligence agent, including the range of business data tables that the business intelligence agent is allowed to access, the whitelist of fields that can be queried, and the boundaries of the service providers that can be accessed. The data table permission verification and the confidential field verification in the three-layer verification are both performed based on the configuration rules of the database permission configuration table.

[0039] This embodiment provides a multi-agent fusion scheduling method, which can be used in the aforementioned multi-agent fusion scheduling system. Figure 2 This is a flowchart of a multi-agent fusion scheduling method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain the service object request, and parse and split the service object request to obtain scheduling planning information, which includes each subtask.

[0040] Among them, a service recipient request refers to an original inquiry message initiated by a service recipient through channels such as customer service hotlines, mini-programs, or online conversations on an app. This service recipient request carries information such as the original request text, the service recipient's level identifier, and the urgency of the business. The original request text is the specific problem described by the service recipient in natural language, such as "Check this month's electricity bill and report my electricity meter for repair." The service recipient's level identifier is used to distinguish the importance of different service recipients, such as ordinary service recipients, VIP (Very Important Person) service recipients, and strategic service recipients. The business urgency is used to indicate the urgency of the current request.

[0041] Parsing and decomposition refers to multi-level processing of service object requests. This involves extracting the original request text, service object level identifier, and business urgency from the request; performing semantic parsing on the original request text to break it down into multiple independent sub-requests; identifying the dependencies between these sub-requests; generating a task topology based on these dependencies, which includes the execution order of each sub-task. For example, querying an electricity bill and reporting meter repairs are independent and can be processed in parallel, while querying a complaint processing progress depends on the completion of the complaint submission; and calculating the scheduling priority score for each sub-task based on the service object level identifier and business urgency. Finally, the task topology and the scheduling priority scores of each sub-task are combined to form scheduling planning information.

[0042] As an example, a VIP service recipient sends a service request: "Check my electricity balance and report the street light in front of my door for repair." Based on the service recipient's request, two sub-requests are parsed out: electricity balance inquiry and street light repair. It is identified that the two are independent and correspond to the electricity bill agent and the repair agent, respectively. At the same time, a higher scheduling priority is assigned according to the VIP level, and scheduling planning information containing these two sub-tasks is generated.

[0043] Step S202: Based on the scheduling planning information, each subtask is matched to a preset intelligent agent, and the context information corresponding to the service object request is obtained.

[0044] The pre-defined intelligent agents refer to multiple pre-deployed intelligent agents with specific business capabilities, including but not limited to business intelligent agents such as electricity bill intelligent agents, repair intelligent agents, industrial product after-sales intelligent agents, Tesco return and exchange intelligent agents, business travel follow-up intelligent agents, and natural language query intelligent agents. Each business intelligent agent is responsible for handling specific types of sub-tasks, such as the electricity bill intelligent agent handling electricity balance inquiries and bill detail inquiries; the natural language query intelligent agent is responsible for converting natural language query requirements into database query statements. The sub-tasks split in step S201 are assigned to the corresponding intelligent agents according to the scheduling priority from high to low, and combined with the execution order in the task topology. Sub-tasks without dependencies are matched and distributed in parallel, while sub-tasks with prerequisite dependencies wait for the prerequisite sub-tasks to be executed before matching and distribution.

[0045] Context information refers to historical interaction records and structured information related to service object requests, such as the service object's identity information, historical work order records, and account number information confirmed in the last dialogue. The process of obtaining context information is as follows: the agent initiates a context read request carrying an authentication vector, which contains the agent's service provider identifier and business line identifier; the authentication vector is matched with the fragment identifier vector of each context fragment in the context memory pool, and if the match is successful, the context information in the corresponding fragment is allowed to be read; the context memory pool stores the context information of each service provider in isolation according to the dual dimensions of service provider identifier and business line identifier, ensuring that the data between different service providers is isolated from each other.

[0046] As an example, after receiving the subtask of querying the electricity balance, the electricity bill agent initiates a context read request and obtains the account number information provided by the service object in the previous conversation through authentication vector matching. This allows the service object to directly perform subsequent queries without having to enter the account number again.

[0047] Step S203: An intelligent agent is used to execute the corresponding sub-task based on the context information, perform security verification on data access during the execution process, and obtain the business processing result based on the context information and the verification result.

[0048] Executing the corresponding sub-task refers to the intelligent agent handling the assigned sub-tasks according to its own responsibilities, such as the electricity bill intelligent agent handling electricity bill inquiries and the repair intelligent agent handling fault repairs. Security verification refers to performing three layers of security verification on database access during the execution of sub-tasks. When a sub-task needs to query the database, a natural language query intelligent agent is invoked to convert the natural language query request into a raw database query statement. This raw database query statement then undergoes three layers of verification sequentially: the first layer is the service provider layer verification, checking whether the query conditions contain the current service provider's equality filtering conditions; if not, the query statement is discarded. The second layer is the confidential field verification, checking whether the query fields contain confidential fields such as electricity bill amount and mobile phone number; if so, the query statement is discarded. The third layer is the data table permission verification, checking whether the target data table belongs to the preset access range of the current business intelligent agent; if not, the query statement is discarded. Database queries can only be executed after all three layers of verification pass, preventing unauthorized data access from the bottom up.

[0049] The verification result includes two cases: verification passed and verification failed. If the verification passes, a database query is executed to obtain the query result. The agent generates the business processing result based on the context information and the query result. If the verification fails, the query statement is discarded. When the subtask does not need to query the database, such as a pure consultation request, the agent directly generates the business processing result based on the context information without triggering the security verification process.

[0050] As an example, when the electricity billing agent executes the subtask of querying the electricity bill balance, it needs to access the electricity billing business database. The agent calls the natural language query agent to convert the natural language requirement of querying the electricity bill balance of the current service object into a database query statement. The database query statement is validated in three layers to confirm that the query conditions include the current service provider's equality filtering conditions, no confidential fields are extracted, and the target data table is within the allowed access range of the electricity billing agent. After the validation is passed, the query is executed, and the electricity billing agent generates a business processing result of the current electricity bill balance being 120 yuan based on the query results.

[0051] Step S204: Generate fusion scheduling results based on business processing results.

[0052] Among them, the business processing result refers to the processing result returned by each business intelligence agent after completing its own sub-task. Since each sub-task is processed by different intelligence agents, the returned results may be fragmented, repetitive, or inconsistent in terms of content. The fusion scheduling result refers to the unified response generated after aggregating and merging the business processing results of multiple business intelligence agents. The process of generating the fusion scheduling result is as follows: aggregating the business processing results returned by all business intelligence agents; eliminating conflicting and repetitive content between different business processing results to obtain a standard business result; generating a complete response text with a unified tone based on the standard business result; and pushing this complete response text as the fusion scheduling result to the service object.

[0053] As an example, the electricity bill agent returns: "Your current electricity balance is 120 yuan." The repair agent returns: "Streetlight malfunction has been registered, work order number GZ20240801001. The repairman is expected to arrive in 30 minutes." The results fusion module combines these two results, eliminating possible discrepancies, and generates a unified and complete response text: "Hello, your current electricity balance is 120 yuan; streetlight repair has been registered, work order number GZ20240801001. The repairman is expected to arrive in 30 minutes. Please keep your phone accessible." The service recipient receives a complete response in one go, without needing to interact with multiple agents separately.

[0054] This embodiment provides a multi-agent fusion scheduling method. It obtains scheduling planning information containing sub-tasks by parsing and splitting service object requests, identifies dependencies using task topology, and performs parallel matching and distribution, effectively shortening the overall response time in multi-request scenarios. By segmenting and isolating the context memory pool according to both service provider and business line identifiers and using authentication vector matching, it prevents unauthorized access to cross-service provider data at the storage level. By performing security checks on database query statements, it eliminates the risk of unauthorized access and sensitive information leakage at the database query level. Through a cross-agent context synchronization mechanism, it enables the sharing and reuse of collected information among multiple agents, eliminating duplicate queries. By fusing the business processing results of each agent to generate a unified and complete response, it improves the output quality of multi-agent collaborative services and solves the problems of missing data isolation, high risk of unauthorized queries, and fragmented content in multi-path responses.

[0055] This embodiment provides a multi-agent fusion scheduling method, which can be used in the aforementioned multi-agent fusion scheduling system. Figure 3 This is a flowchart of a multi-agent fusion scheduling method according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain the service object request, and parse and split the service object request to obtain scheduling planning information, which includes each subtask.

[0056] Specifically, step S301 above includes: Step S3011: Extract the original request text, service object level identifier, and business urgency from the service object request; perform semantic parsing on the original request text and split it into multiple sub-requests; identify the dependencies between the sub-requests and generate a task topology structure based on the dependencies, the task topology structure containing the execution order of each sub-task.

[0057] Specifically, a service recipient initiates a service request through an online session on the APP. The request content is: check the current electricity balance, report the street light in front of the door for repair, and inquire about the status of industrial product purchase orders. After receiving the service recipient's request, the dispatch center extracts the original request text, service recipient level identifier, and business urgency information from the service recipient's request. In this embodiment, the scheduling center inputs the original request text into a pre-trained large language model. This large language model is a deep neural network model based on the Transformer architecture, pre-trained on a large-scale text corpus, and possesses natural language understanding, intent recognition, and task decomposition capabilities. Semantic parsing is performed on the input to the large language model, outputting three independent sub-requests: querying electricity balance, reporting street light repair, and querying the status of industrial product purchase orders. The scheduling center identifies the dependencies between these sub-requests: the three sub-requests are respectively directed to the electricity bill business database, the repair order system, and the industrial product supply chain database, and there are no data or temporal dependencies between them; each can be executed independently without waiting for other sub-requests to complete. Based on these dependencies, the scheduling center generates a task topology, marking all three sub-tasks as having no dependencies, and the execution order is that all three are executed in parallel.

[0058] Step S3012: Determine the service object level score based on the service object level identifier, and determine the urgency score based on the business urgency corresponding to each sub-task; perform a weighted summation of the service object level score and the urgency score to obtain the scheduling priority score of each sub-task; combine the task topology and scheduling priority score into scheduling planning information.

[0059] Specifically, the dispatch center determines the service object level score based on the service object level identifier and the business urgency corresponding to each sub-task, and determines the urgency score of each sub-task according to the service object level score determination rules. The dispatch center substitutes the service recipient's level score and the urgency score of each subtask into the weighted summation formula: (1); in, As the scheduling priority score, For the service recipient level score, The score represents the urgency level. and The preset weighting coefficient is used to adjust the relative importance of service object level and urgency in the scheduling priority calculation. The preset weighting coefficient is pre-configured in the scheduling priority according to the service object level and urgency, and the scheduling priority score of each subtask is calculated. The scheduling priority score of each subtask is calculated, such as setting... It is 1.2. The priority is 1.0. For street light repair, the priority is P=1.2×3+1.0×3=6.6. For electricity balance inquiry, the priority is P=1.2×3+1.0×1=4.6. For industrial product purchase order status inquiry, the priority is P=1.2×3+1.0×1=4.6. Among the three sub-tasks, the sub-task corresponding to fault-related requests has the highest scheduling priority due to its higher urgency score. The two sub-tasks corresponding to ordinary query requests have the same scheduling priority. The scheduling center combines the task topology and the scheduling priority scores of each sub-task into scheduling planning information.

[0060] In some optional implementations, step S3012 above includes: Step a1, the rules for determining the service recipient level score are as follows: the first type of service recipient corresponds to the first score, the second type of service recipient corresponds to the second score, the third type of service recipient corresponds to the third score, and the fourth type of service recipient corresponds to the fourth score.

[0061] Specifically, in this embodiment, the service object level score is used to quantify the importance of different service objects. The higher the score, the higher the scheduling priority. The score is determined according to the service object level identifier and the following judgment rules: the first type of service object is a regular service object with a first score of 0 points, and no exclusive service rights; the second type of service object is a regular VIP service object with a second score of 1 point, and basic priority rights; the third type of service object is a high-level VIP service object with a third score of 2 points, and high priority handling rights; the fourth type of service object is a strategic service object with a fourth score of 3 points, and the highest priority exclusive service channel rights.

[0062] Step a2, the rules for determining the urgency score are as follows: the first type of request corresponds to the first score, the second type of request corresponds to the second score, the third type of request corresponds to the third score, the fourth type of request corresponds to the fourth score, the fifth type of request corresponds to the fifth score, and the sixth type of request corresponds to the sixth score.

[0063] Specifically, the first, second, third, fourth, fifth, and sixth categories of requests correspond to pure consultation requests, general inquiry requests, general business processing requests, fault or service anomaly requests, complaint requests, and group complaints or high-risk compliance requests, respectively. The fifth category can be scored as 4 points, and the sixth category can be scored as 5 points.

[0064] The urgency score for each sub-request is determined according to the following rules: Pure inquiry requests receive 0 points (no time limit, e.g., what are the electricity pricing rules?); general inquiry requests receive 1 point (low time limit, e.g., checking electricity balance); general business processing requests receive 2 points (medium time limit, e.g., changing payment methods); fault or service anomaly requests receive 3 points (high time limit, e.g., street light repair); complaint requests receive 4 points (high time limit, e.g., disagreement with electricity bills); major faults, group complaints, or high-risk compliance requests receive 5 points (highest urgency level, requiring immediate action, e.g., large-scale power outages). For example, in this embodiment, street light repair falls under the fault or service anomaly category, receiving a urgency score of 3. Checking electricity balance and checking industrial product purchase order status are both general inquiry requests, each receiving a urgency score of 1 point.

[0065] Step S302: Based on the scheduling planning information, each subtask is matched to a preset intelligent agent, and the context information corresponding to the service object request is obtained.

[0066] Specifically, step S302 above includes: Step S3021: The preset intelligent agents include business intelligent agents and natural language query intelligent agents; according to the scheduling priority score from high to low and combined with the execution order, each subtask is matched to the business intelligent agent; among them, subtasks without dependencies in the task topology are matched in parallel, and subtasks with prerequisite dependencies wait for the prerequisite subtasks to be executed before matching.

[0067] Specifically, multiple business intelligence agents are pre-deployed, including but not limited to an electricity bill intelligence agent for handling electricity bill inquiries, bill details inquiries, payment record inquiries, overdue payment status inquiries, payment method guidance, electricity price policy standards, and electricity invoice issuance; a repair intelligence agent for handling fault reporting and repair progress inquiries; an industrial product intelligence agent for handling order status inquiries, logistics tracking inquiries, return and exchange application submissions, and after-sales review progress inquiries; and a natural language query intelligence agent is also pre-deployed to convert natural language query requirements into database query statements for each business intelligence agent to call when it needs to query the database.

[0068] Based on the scheduling priority scores from high to low, and combined with the execution order in the task topology, each subtask is matched to the corresponding business agent. For example, if the scheduling priority scores of the three subtasks are 6.6 for the street light repair request, 4.6 for the electricity balance query, and 4.6 for the industrial product purchase order status query, the street light repair request will be matched to the repair agent first. The electricity balance query and industrial product purchase order status query subtasks have equal scheduling priority and no dependency, and will be matched to the electricity bill agent and the industrial product agent respectively, and the three will be executed in parallel. If there are pre-dependent subtasks, such as the complaint submission subtask which must be executed before the complaint processing progress query subtask, the complaint submission subtask will wait for the agent corresponding to the complaint submission subtask to complete and return the result before the complaint processing progress query subtask is matched to the corresponding business agent.

[0069] In step S3022, the context information is stored in the context memory pool. The context memory pool is divided into two dimensions, namely, the service provider identifier and the business line identifier, to store the context information of each service provider in isolation. A business agent initiates a context read request carrying an authentication vector, which includes the service provider identifier and the business line identifier of the business agent.

[0070] Specifically, the context information of each service provider is uniformly stored in the context memory pool. The context memory pool adopts a storage architecture with dual-dimensional sharding and isolation based on service provider identifier and business line identifier. That is, each context shard is divided using the service provider identifier and business line identifier as a composite key. Data from different service providers is stored in different shards, and the data between shards is physically isolated and cannot be read across shards. After matching the electricity balance query subtask, the electricity billing agent obtains the context information of the current service object to determine the account number. The electricity billing agent initiates a context read request carrying an authentication vector, which contains the service provider identifier (e.g., service provider A) and the business line identifier (e.g., electricity billing business line). After receiving the context read request, the context memory pool locates the corresponding context fragment based on the service provider identifier and business line identifier in the authentication vector. In addition, if the strategic service object with the service provider identifier of service provider A has previously provided the account number information 1001 through other channels such as customer service hotline, this account number information is already stored in the service provider A-electricity billing business line fragment. After the electricity billing agent matches this fragment through the authentication vector, it can read the account number information as 1001 without having to ask the service object for the account number again. If the service provider identifier in the electricity billing agent's authentication vector is service provider B, but the requested context fragment is service provider A, then the authentication vector does not match the service provider identifier, and the context read request is rejected.

[0071] Step S3023: Calculate the cosine similarity between the authentication vector and the fragment identifier vector; when the cosine similarity is greater than or equal to the preset similarity threshold, the match is deemed successful and the context reading request is allowed; when the cosine similarity is less than the preset similarity threshold, the match is deemed unsuccessful and the context reading request is blocked.

[0072] Specifically, the matching between the authentication vector and the fragment identifier vector is achieved by calculating cosine similarity. Each context fragment has a unique fragment identifier vector, and the authentication vector of the business agent includes the service provider identifier and business line identifier of that business agent. During matching, the cosine similarity between the authentication vector and the fragment identifier vector of the target fragment is calculated using the following formula: (2); in, The cosine similarity value ranges from -1 to 1. The closer the cosine similarity value is to 1, the more likely the two vectors are in the same direction. This is the authentication vector; This is a fragment identifier vector; It is the dot product of two vectors; and These are the magnitudes of the authentication vector and the fragmentation identifier vector, respectively.

[0073] If a three-dimensional rectangular space is used to represent the combination of service provider identifiers and business line identifiers, where the first dimension represents the service provider identifier (e.g., service provider A corresponds to the value 1, service provider B corresponds to the value 2), the second dimension represents the e-commerce business line (represented by the value 1), and the third dimension represents the industrial products business line (represented by the value 1), then the e-commerce business line identifier vector A can be represented as Service Provider A - E-commerce Business Line, in vector form as follows: The target shard is Service Provider A - Electricity Sales Business Line, and the shard identifier vector is... Both are completely identical. The preset similarity threshold is adjusted according to actual security needs. For example, in this embodiment, the preset similarity threshold is set to 0.8. If 1.0 > 0.8, the match is deemed successful, and the business intelligence agent is allowed to read the context information in the segment of the service provider A-electric business line. If the authentication vector A of the industrial product intelligence is A-industrial product business line of the service provider, the vector form is: Try reading the context information in the shard of service provider A-electronic business line, then the shard identifier vector of the target shard. The calculated cosine similarity is 0.5, which is less than the preset similarity threshold of 0.8. Therefore, the match is deemed unsuccessful, the request to read the context is intercepted, an unauthorized access log is recorded, and an empty context is returned.

[0074] Step S303: An intelligent agent is used to execute the corresponding sub-task based on the context information, perform security verification on data access during the execution process, and obtain the business processing result based on the context information and the verification result.

[0075] Specifically, step S303 above includes: Step S3031: The business intelligence agent executes the corresponding sub-task based on the context information.

[0076] Specifically, after completing the context information reading, the corresponding business logic is executed according to the sub-tasks they are matched with. For example, the three sub-tasks are matched to the corresponding business intelligent agents: the street light repair sub-task is matched to the repair intelligent agent, the electricity balance query sub-task is matched to the electricity intelligent agent, and the industrial product purchase order status query sub-task is matched to the industrial product intelligent agent.

[0077] Since step S301 identifies no dependency between the three subtasks, the three business intelligence agents execute their respective subtasks in parallel. The repair intelligence agent directly generates a repair work order based on the service object address read from the context information, without needing to ask the service object for address information again. The electricity bill intelligence agent prepares to perform an electricity balance query based on the account number information 1001 read from the context information. The industrial products intelligence agent prepares to perform a purchase order status query based on the historical order records of the service object read from the context information.

[0078] Step S3032: When a subtask needs to query the database, the natural language query agent is invoked to convert the natural language query requirement into a raw database query statement. Three-layer security verification is performed on the raw database query statement. After the verification is passed, the database query is executed to obtain the query result. The business agent generates the business processing result based on the context information and the query result.

[0079] Specifically, when using the electricity bill intelligent agent to execute the electricity bill balance query subtask, it is necessary to obtain the real-time electricity bill balance data of the current service object from the electricity bill business database. This electricity bill balance query subtask requires querying the database. The natural language query intelligent agent is called to convert the natural language query requirement of querying the current electricity bill balance of the service object with the service provider identified as service provider A and the account number information as 1001 into the original database query statement. The original database query statement is then subjected to three layers of security checks in sequence. The first layer is the service provider layer validation: it checks whether the query conditions of the original database query statement contain the current service provider's equality filter conditions. If the query conditions are checked and it is confirmed that there is an equality condition where the service provider identifier is the same as the current service provider identifier, the service provider layer validation passes; if the original database query statement does not contain the service provider identifier filter conditions, such as only stating that the customer number is equal to a certain number, the validation fails, the original database query statement is discarded and a regeneration is triggered.

[0080] The second layer is the confidential field verification: It iterates through the query field list and checks whether it contains any field from the preset confidential field set. The preset confidential field set is pre-configured in the database permission configuration table according to data security compliance requirements. The preset confidential field set includes fields such as electricity bill details, customer mobile phone number, and enterprise qualification document fields. The query field of the current original database query statement is the electricity bill balance, which is not a confidential field, so the confidential field verification passes. If the query field contains the customer mobile phone number, the verification is deemed to have failed, and the original database query statement is directly discarded.

[0081] The third layer is data table permission verification: it checks whether the target data table being queried is within the preset access range of the electricity billing agent; the preset access range of the electricity billing agent is pre-configured in the database permission configuration table, including the electricity account table, electricity bill table, payment record table, etc. If the target data table being queried is the electricity account table, which is within the access range of the electricity billing agent, the data table permission verification passes; if the target data table is the industrial product order table, it is not within the access range of the electricity billing agent, the verification fails, and the original database query statement is directly discarded.

[0082] After all three layers of verification pass, the original database query statement is executed to perform a database query, and the query result shows that the electricity balance is 120 yuan. The electricity bill intelligent agent generates the business processing result based on the context information such as the service object name Mr. Zhang, the account number information 1001 and the query result of 120 yuan: Hello Mr. Zhang, your current electricity balance is 120 yuan. When the industrial product intelligent agent executes the sub-task of querying the status of industrial product purchase orders, this sub-task also needs to query the database, such as querying the order status data in the industrial product supply chain system. Similarly, after calling the natural language query intelligent agent to perform three-layer security verification, the query results are obtained and the business processing results are generated: Hello Mr. Zhang, your industrial product purchase order, order number PO202407001, is currently in the status of "shipped" and is expected to arrive in 3 days.

[0083] Step S3033: When the subtask does not need to query the database, the business intelligence agent directly generates the business processing result based on the context information.

[0084] Specifically, when the intelligent repair agent executes the street light repair sub-task, this sub-task falls under the category of business processing requests and does not require querying the database to obtain additional data. The intelligent repair agent directly generates a repair work order based on the context information read in step S302, such as the service recipient's address and the service recipient's name being Mr. Zhang, and returns the business processing result: "Hello Mr. Zhang, a repair work order for the street light at your door has been generated. The work order number is GZ20240801001. The repairman is expected to arrive in 30 minutes. Please keep your phone accessible."

[0085] Step S304: Generate fusion scheduling results based on business processing results.

[0086] Specifically, step S304 above includes: Step S3041: Summarize the business processing results and eliminate conflicting and duplicate content between different business processing results to obtain standard business results; based on the standard business results, generate a complete response text with a consistent tone, and use the complete response text as the fusion scheduling result.

[0087] Specifically, after each business intelligence agent completes its parallel execution, the business processing results returned by all business intelligence agents are aggregated. For example, the three business intelligence agents may return the following business processing results: Electricity bill intelligent agent returned: Hello Mr. Zhang, your current electricity bill balance is 120 yuan.

[0088] The repair agent replied: "Hello Mr. Zhang, a repair order for the street light in front of your door has been generated. The order number is GZ20240801001. The repairman is expected to arrive in 30 minutes. Please keep your phone accessible."

[0089] Industrial Product Intelligent Agent Response: Hello Mr. Zhang, your industrial product purchase order, order number PO202407001, is currently in the "shipped" status and is expected to arrive in 3 days.

[0090] Summarize these three business processing results and check for conflicting and duplicate content. The detection rules include, but are not limited to, duplicate content detection and conflicting content detection. For example, if all three business processing results begin with "Hello Mr. Zhang," it is a duplicate greeting. Retain one consistent greeting and eliminate the rest of the duplicates. Also check for factual conflicts between the business processing results. For example, if one agent returns that the current balance is sufficient while another agent returns that the current balance is insufficient, determine which one is correct based on the actual data source. The three business processing results involve electricity balance, repair work order, and order status, respectively. Since they are different business dimensions and have no conflicting content, their respective result information is directly retained.

[0091] After eliminating duplicate content, the three business processing results are merged into a standard business result: Your current electricity balance is 120 yuan; a repair order for the street light at your door has been generated, with the order number GZ20240801001, and the repairman is expected to arrive in 30 minutes; your industrial product purchase order, with the order number PO202407001, is currently in the "shipped" status and is expected to arrive in 3 days. Based on the standard business results, a complete response text with a consistent tone is generated. In this embodiment, a large language model is used to refine the standard business results, merging scattered statements into a logically coherent and consistent response text. For example, the large language model reorganizes the three results in logical order: first the electricity balance, then the repair order, and finally the order status, all using the second-person pronoun "you," generating a complete response text: Hello, according to our inquiry, your current electricity balance is 120 yuan; a repair order for the street light at your door has been generated, order number GZ20240801001, and the repairman is expected to arrive in 30 minutes, please keep your phone accessible; in addition, your industrial product purchase order, order number PO202407001, is currently in the "shipped" status and is expected to arrive in 3 days.

[0092] The complete response text is pushed to the service recipient's session as the fusion scheduling result. The service recipient receives the complete response to all three sub-tasks at once, eliminating the need to communicate with multiple agents separately, avoiding content fragmentation and repetitive queries, and improving the service experience. If an agent fails to execute the three sub-tasks, such as the industrial product agent, and does not return a result due to an anomaly, the result fusion module will only include two results in the complete response text: electricity balance and repair order, and add the prompt message: Industrial product order status query is temporarily unavailable. Please try again later or contact customer service.

[0093] This embodiment provides a multi-agent fusion scheduling method that semantically parses service object requests into multiple sub-requests, identifies dependencies between sub-requests to generate a task topology, enabling dependent sub-tasks to execute in parallel and dependent sub-tasks to wait sequentially, effectively shortening the overall response time of complex requests; it obtains scheduling priorities so that high-level service objects and urgent requests receive processing resources first; it uses a context memory pool to partition and isolate storage based on both service provider and business line identifiers, combined with cosine similarity matching of authentication vectors, to prevent unauthorized access to cross-service provider data at the storage architecture level; it performs three-layer security checks on database query statements to eliminate the risk of unauthorized access and leakage of confidential fields at the database query level; and it generates a unified and complete response by summarizing the processing results of each business and eliminating conflicting and duplicate content, avoiding the problems of fragmented and inconsistent responses from multiple sources, and solving the problems of low response efficiency, lack of data isolation, high risk of unauthorized queries, and fragmented responses from multiple sources.

[0094] As a specific application embodiment of the present invention, the multi-agent fusion scheduling method provided by the present invention will be further described in detail, such as... Figure 4 As shown, it includes the following steps: 1. Service object request access and multiple request parsing: The service recipient initiates a service request through an online session on the APP. The request content includes checking the current electricity balance, reporting the street light in front of the door for repair, and inquiring about the status of industrial product purchase orders. After receiving the service recipient's request, the dispatch center extracts the original request text, service recipient level identifier, and business urgency information from the service recipient's request. Among them, the service recipient level identifier is strategic service recipient, and the business urgency is marked as fault type.

[0095] The scheduling center inputs the original request text into a large language model for semantic parsing. The large language model outputs three independent sub-requests: querying electricity balance, reporting street light repairs, and querying the status of industrial product purchase orders. It also identifies the dependencies between these sub-requests: the three sub-requests target the electricity billing database, the repair order system, and the industrial product supply chain database, respectively. There are no data or temporal dependencies between them, and each can be executed independently without waiting for the completion of other sub-requests. Based on these dependencies, the scheduling center generates a task topology, marking all three sub-tasks as dependency-free, and executing them in parallel.

[0096] 2. Scheduling priority calculation and scheduling planning information generation: The dispatch center determines the service recipient level score for strategic service recipients based on the service recipient level identifier. According to the service recipient level score determination rules, such as 0 points for ordinary service recipients, 1 point for ordinary VIP service recipients, 2 points for advanced VIP service recipients, and 3 points for strategic service recipients, the service recipient level score is determined to be V=3.

[0097] The dispatch center determines the urgency score for each sub-task based on its corresponding business urgency level, according to the urgency score determination rules. For example, a pure inquiry request corresponds to 0 points, a general query request corresponds to 1 point, a fault or service anomaly request corresponds to 3 points, a complaint request corresponds to 4 points, and a major fault, group complaint, or high-risk compliance request corresponds to 5 points. Inquiring about electricity balance and checking the status of industrial product purchase orders are considered general queries, with an urgency score of U=1; street light repair requests are considered fault or service anomaly requests, with an urgency score of U=3.

[0098] The dispatch center substitutes the service recipient's level score and the urgency score of each subtask into the weighted summation formula: (1); in, As the scheduling priority score, For the service recipient level score, The score represents the urgency level. and The scheduling priority score of each subtask is calculated using preset weighting coefficients. The scheduling priority score of each subtask is calculated, such as setting... It is 1.2. The priority is 1.0. For street light repair, the priority is P=1.2×3+1.0×3=6.6. For electricity balance inquiry, the priority is P=1.2×3+1.0×1=4.6. For industrial product purchase order status inquiry, the priority is P=1.2×3+1.0×1=4.6. Among the three sub-tasks, the sub-task corresponding to fault-related requests has the highest scheduling priority due to its higher urgency score. The two sub-tasks corresponding to ordinary query requests have the same scheduling priority. The scheduling center combines the task topology and the scheduling priority scores of each sub-task into scheduling planning information.

[0099] 3. Agent matching and context authentication reading: The scheduling center matches the street light repair subtask to the repair agent, the electricity balance query subtask to the electricity bill agent, and the industrial product purchase order status query subtask to the industrial product agent, according to the scheduling priority from high to low. If there is no dependency, a parallel matching and distribution strategy can be adopted, and task execution instructions are issued to the three business agents at the same time.

[0100] Before executing business logic, each business intelligence agent obtains the context information of the current service object. The context information is stored in a context memory pool, which is divided into two dimensions: service provider identifier and business line identifier to store the context information of each service provider in isolation.

[0101] The electricity billing agent initiates a context read request carrying an authentication vector, which is the service provider A-electricity billing business line. The context memory pool calculates the cosine similarity between this authentication vector and the shard identifier vector of the target shard, which is the service provider A-electricity billing business line, to achieve matching between the authentication vector and the shard identifier vector. The calculation formula is as follows: (2); in, The cosine similarity value ranges from -1 to 1. The closer the cosine similarity value is to 1, the more likely the two vectors are in the same direction. This is the authentication vector; This is a fragment identifier vector; It is the dot product of two vectors; and These are the magnitudes of the authentication vector and the fragmentation identifier vector, respectively.

[0102] If a three-dimensional rectangular space is used to represent the combination of service provider identifiers and business line identifiers, where the first dimension represents the service provider identifier (e.g., service provider A corresponds to the value 1, service provider B corresponds to the value 2), the second dimension represents the e-commerce business line (represented by the value 1), and the third dimension represents the industrial products business line (represented by the value 1), then the e-commerce business line identifier vector A can be represented as Service Provider A - E-commerce Business Line, in vector form as follows: The target shard is Service Provider A - Electricity Sales Business Line, and the shard identifier vector is... The two are completely identical. A preset similarity threshold, such as 0.8, is used. If 1.0 > 0.8, the match is considered successful, and the business intelligence agent is allowed to read the context information in the service provider A - the e-commerce business line segment to obtain the account number information as 1001. The repair intelligence agent's authentication vector is service provider A - the repair business line. After a successful match, it reads the service provider A - the repair business line segment to obtain the service object address. The industrial product intelligence agent's authentication vector is service provider A - the industrial product business line. After a successful match, it reads the service provider A - the industrial product business line segment to obtain historical order records, which include the order number PO202407001. The three business intelligence agents obtain the context information they need without having to repeatedly ask the service object for the account number, address, or order number.

[0103] 4. Parallel execution of subtasks and three-layer security checks: Three business intelligence agents execute their respective subtasks in parallel. When the electricity billing intelligence agent performs the electricity balance query subtask, it needs to retrieve real-time electricity balance data from the electricity billing business database and call the natural language query intelligence agent. The natural language query intelligence agent converts the natural language query requirement—querying the current electricity balance of a service object identified as service provider A with account number 1001—into a raw database query statement. The security verification engine performs three layers of security verification on this raw database query statement: The first layer is the service provider layer verification: it checks whether the query conditions of the original database query statement contain the current service provider's equality filter condition. If it confirms that there is an equality condition with the same service provider identifier as service provider A, the service provider layer verification passes; if the original database query statement does not contain the service provider identifier filter condition, such as only writing customer number equal to 1001, the verification is judged to fail, the original database query statement is directly discarded and regeneration is triggered.

[0104] The second layer is the confidential field validation: it iterates through the list of query fields and checks whether any field from the preset confidential field set is included. The preset confidential field set includes fields such as electricity bill details, customer mobile phone number, and enterprise qualification document fields. The query field of the current original database query is the electricity balance, which is not a confidential field, so the confidential field validation passes; if the query field contains the customer mobile phone number field, the validation fails, and the original database query is discarded.

[0105] The third layer is data table permission verification: it checks whether the target data table being queried falls within the preset accessible data table range of the electricity billing agent. The preset accessible range of the electricity billing agent is pre-configured in the database permission configuration table, including the electricity account table, electricity bill table, payment record table, etc. The target data table being queried is the electricity account table, which falls within this accessible range, and the data table permission verification passes. If the target data table is the industrial product order table, it does not fall within the accessible range of the electricity billing agent, the verification fails, and the original database query statement is directly discarded.

[0106] After all three layers of verification pass, the original database query statement is executed to perform a database query, and the query result is that the electricity balance is 120 yuan. The electricity bill agent generates the business processing result based on the context information and the query result: Hello Mr. Zhang, your current electricity balance is 120 yuan.

[0107] Similarly, when the industrial product intelligent agent executes the subtask of querying the status of industrial product purchase orders, it needs to query the industrial product supply chain database to obtain order status data, obtain the query results, and generate business processing results: Hello Mr. Zhang, your industrial product purchase order, order number PO202407001, is currently in the status of "shipped" and is expected to arrive in 3 days.

[0108] When the intelligent agent executes the street light repair sub-task, the street light repair sub-task belongs to the business processing category and does not need to query the database to obtain additional data; it directly generates a repair work order based on the context information and returns the business processing result: Hello Mr. Zhang, a repair work order for the street light at your door has been generated, the work order number is GZ20240801001, the repairman is expected to arrive in 30 minutes, please keep your phone accessible.

[0109] 5. Controlled context synchronization across agents: After each business agent completes its execution, it writes the newly collected structured business information back to the context memory pool. For example, the electricity bill agent updates the account number information 1001 confirmed in this query to the service provider A-electricity bill business line segment and updates the timestamp; the repair agent writes the generated work order number GZ20240801001 to the service provider A-repair business line segment; and the industrial products agent updates the current order status to the service provider A-industrial products business line segment. This updated context information can be directly read and reused by other business agents under the same service provider when executing tasks in the future, without having to collect information repeatedly from the service object. Data synchronization between context segments across service providers is strictly prohibited to ensure clear data boundaries.

[0110] 6. Results fusion and unified response generation: After each business agent completes its parallel execution, the result fusion module summarizes the business processing results returned by all business agents, including: the electricity bill agent returning "Hello Mr. Zhang, your current electricity balance is 120 yuan"; the repair agent returning "Hello Mr. Zhang, a repair order for the street light at your door has been generated, with the order number GZ20240801001, and the repairman is expected to arrive in 30 minutes, please keep your phone accessible"; and the industrial products agent returning "Hello Mr. Zhang, your industrial product purchase order, order number PO202407001, is currently in the 'shipped' status and is expected to arrive in 3 days."

[0111] The result fusion processor detected conflicting and duplicate content in the three business processing results: all three business processing results begin with "Hello Mr. Zhang," which is a duplicate greeting. One consistent greeting is retained, and the remaining duplicates are eliminated. The three results involve electricity balance, repair work order, and order status, respectively. Since they are different business dimensions and have no conflicting content, their respective result information is directly retained.

[0112] After eliminating duplicate content, the three business processing results are merged into a standard business result: Your current electricity balance is 120 yuan; a repair order for the street light at your door has been generated, with the order number GZ20240801001, and the repairman is expected to arrive in 30 minutes; your industrial product purchase order, with the order number PO202407001, is currently in the "shipped" status and is expected to arrive in 3 days.

[0113] The large language model is used to refine the standard business result, merging the scattered statements into a logically coherent and consistent complete response text. For example, the large language model reorganizes the three results in logical order: first the electricity balance, then the repair order, and finally the order status, all using the second-person pronoun "you" to generate a complete response text: Hello, according to our inquiry, your current electricity balance is 120 yuan; a repair order for the street light at your door has been generated, order number GZ20240801001, and the repairman is expected to arrive in 30 minutes, please keep your phone accessible; in addition, your industrial product purchase order, order number PO202407001, is currently in the "shipped" status and is expected to arrive in 3 days.

[0114] The complete response text is then pushed to the service recipient's app session as the fusion scheduling result. The service recipient receives complete responses to all three sub-tasks at once, eliminating the need to interact with multiple agents separately, avoiding fragmented content and repetitive queries, and improving the service experience.

[0115] This invention provides a multi-agent fusion scheduling method that uses a scheduling hub to semantically parse and break down complex requests into multiple sub-requests, identifies dependencies between sub-requests to generate a task topology, and enables dependent sub-tasks to execute in parallel while dependent sub-tasks wait sequentially, effectively shortening the overall response time in multi-request scenarios. By prioritizing scheduling, high-level service objects and urgent requests are given priority in obtaining processing resources, preventing important requests from being blocked during concurrency. Through cosine similarity matching, cross-service data unauthorized access is prevented at the storage architecture level, meeting data security compliance requirements in multi-service scenarios. Furthermore, by analyzing data... The database query statement performs three layers of security checks to eliminate the risks of unauthorized access and leakage of confidential fields at the database query level; through a cross-agent controlled context synchronization mechanism, the collected structured information can be shared and reused among multiple agents, eliminating duplicate queries; by summarizing the results of each business process and eliminating conflicting and duplicate content, a complete response with a consistent tone is generated, avoiding the problems of fragmented and inconsistent responses from multiple sources, improving the output quality of multi-agent collaborative services, realizing efficient collaborative scheduling of multi-agent customer service systems, and solving the problems of low response efficiency, lack of data isolation, high risk of unauthorized queries, and fragmented responses from multiple sources.

[0116] This embodiment also provides a multi-agent fusion scheduling device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0117] This embodiment provides a multi-agent fusion scheduling device, such as Figure 5 As shown, it includes: The task splitting module 501 is used to obtain service object requests, parse and split the service object requests to obtain scheduling planning information, which includes each sub-task. The scheduling and reading module 502 is used to match each subtask to a preset intelligent agent according to the scheduling planning information, and to obtain the context information corresponding to the service object request; The business execution module 503 is used to employ an intelligent agent to execute corresponding sub-tasks based on context information, perform security verification on data access during the execution process, and obtain business processing results based on context information and verification results. The result fusion module 504 is used to summarize the business processing results and generate fused scheduling results.

[0118] In some alternative implementations, the task splitting module 501 includes: The information extraction unit is used to extract the original request text, service object level identifier, and business urgency from the service object's request.

[0119] The semantic parsing unit is used to perform semantic parsing on the original request text, split it into multiple sub-requests, identify the dependencies between the sub-requests, and generate a task topology structure based on the dependencies. The task topology structure contains the execution order of each sub-task.

[0120] The priority calculation unit is used to determine the service object level score based on the service object level identifier, determine the urgency score based on the business urgency corresponding to each sub-task, and perform a weighted sum of the service object level score and the urgency score to obtain the scheduling priority score of each sub-task; and combine the task topology and scheduling priority score into scheduling planning information.

[0121] In some optional implementations, the priority calculation unit includes: The rating score determination sub-unit is used to determine the rating score of the service object based on the service object rating identifier. The rules for determining the rating score of the service object are as follows: the first type of service object corresponds to the first score, the second type of service object corresponds to the second score, the third type of service object corresponds to the third score, and the fourth type of service object corresponds to the fourth score.

[0122] The urgency score determination sub-unit is used to determine the urgency score based on the business urgency of each sub-task. The urgency score determination rule is as follows: the first type of request corresponds to the first score, the second type of request corresponds to the second score, the third type of request corresponds to the third score, the fourth type of request corresponds to the fourth score, the fifth type of request corresponds to the fifth score, and the sixth type of request corresponds to the sixth score.

[0123] In some alternative implementations, the scheduling read module 502 includes: The task scheduling unit is used to match each subtask to a preset business intelligence agent according to the scheduling priority score from high to low and in combination with the execution order. Among them, subtasks without dependencies in the task topology are matched in parallel, while subtasks with prerequisite dependencies wait for the prerequisite subtasks to be executed before matching.

[0124] The context reading unit is used to obtain the context information corresponding to the service object request.

[0125] In some optional implementations, the context reading unit includes: The authentication request subunit is used to initiate a context read request carrying an authentication vector using a business intelligence agent. The authentication vector includes the service provider identifier and business line identifier of the business intelligence agent.

[0126] The vector matching subunit is used to match the authentication vector with the fragment identifier vector of each context fragment in the context memory pool, and calculate the cosine similarity between the authentication vector and the fragment identifier vector. When the cosine similarity is greater than or equal to the preset similarity threshold, the match is deemed successful and the context read request is allowed. When the cosine similarity is less than the preset similarity threshold, the match is deemed unsuccessful and the context read request is blocked.

[0127] In some alternative implementations, the service execution module 503 includes: The task execution unit is used to execute corresponding subtasks by a business intelligence agent based on context information.

[0128] The security verification unit is used to call the natural language query agent to convert the natural language query requirements into raw database query statements when a subtask needs to query the database. It then performs three layers of security verification on the raw database query statements. After the verification passes, it executes the database query and obtains the query results.

[0129] The result generation unit is used to generate business processing results based on context information and query results when a subtask needs to query the database; and to directly generate business processing results based on context information when a subtask does not need to query the database.

[0130] In some alternative implementations, the result fusion module 504 includes: The summary unit is used to summarize the results of business processing and eliminate conflicting and duplicate content between different business processing results to obtain a standard business result.

[0131] The response generation unit is used to generate a complete response text with a consistent tone based on standard business results, and to use the complete response text as the fusion scheduling result.

[0132] The multi-agent fusion scheduling device provided in this embodiment of the invention can execute a multi-agent fusion scheduling method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method. Further functional descriptions of the various modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0133] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0134] The following is a detailed reference. Figure 6This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0135] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0136] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in a multi-agent fusion scheduling method according to embodiments of the present invention.

[0137] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0138] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, it implements the multi-agent fusion scheduling method shown in the above embodiments.

[0139] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0140] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-agent fusion scheduling method, characterized in that, The method includes: Obtain service object requests, and parse and split the service object requests to obtain scheduling planning information, which includes each subtask; Based on the scheduling planning information, each subtask is matched to a preset intelligent agent, and context information corresponding to the service object request is obtained; The intelligent agent executes corresponding subtasks based on the context information, performs security verification on data access during the execution process, and obtains the business processing result based on the context information and the verification result. Based on the business processing results, a fusion scheduling result is generated.

2. The method according to claim 1, characterized in that, The step of parsing and splitting the service object request to obtain scheduling planning information includes: Extract the original request text, service object level identifier, and business urgency from the service object request; The original appeal text is semantically parsed and broken down into multiple sub-appeals; Identify the dependencies between the sub-requests, and generate a task topology based on the dependencies, wherein the task topology includes the execution order of the sub-tasks; The service object level score is determined based on the service object level identifier, and the urgency score is determined based on the business urgency corresponding to each sub-task. The scheduling priority score of each subtask is obtained by weighted summing of the service object level score and the urgency score. The task topology and the scheduling priority score are combined to form scheduling planning information.

3. The method according to claim 2, characterized in that, The rules for determining the service recipient level score and the urgency score include: The rules for determining the service recipient level score are as follows: the first type of service recipient corresponds to the first score, the second type of service recipient corresponds to the second score, the third type of service recipient corresponds to the third score, and the fourth type of service recipient corresponds to the fourth score. The rules for determining the urgency score are as follows: the first type of request corresponds to the first score, the second type of request corresponds to the second score, the third type of request corresponds to the third score, the fourth type of request corresponds to the fourth score, the fifth type of request corresponds to the fifth score, and the sixth type of request corresponds to the sixth score.

4. The method according to claim 2, characterized in that, The preset intelligent agents include business intelligent agents and natural language query intelligent agents; matching each subtask with a preset intelligent agent according to the scheduling planning information includes: According to the scheduling priority score from high to low and combined with the execution order, each subtask is matched to the business intelligence agent; wherein, subtasks without dependencies in the task topology are matched in parallel, and subtasks with prerequisite dependencies wait for the prerequisite subtasks to be executed before matching.

5. The method according to claim 4, characterized in that, The context information is stored in a context memory pool, which is divided into two dimensions: service provider identifier and business line identifier to store the context information of each service provider in an isolated manner. The step of obtaining the context information corresponding to the service object request includes: The business intelligence agent initiates a context read request carrying an authentication vector, wherein the authentication vector includes the service provider identifier and business line identifier of the business intelligence agent; The authentication vector is matched with the fragment identifier vector of each context fragment in the context memory pool. If the match is successful, the context information in the corresponding context fragment can be read.

6. The method according to claim 5, characterized in that, The step of matching the authentication vector with the fragment identifier vector of each context fragment in the context memory pool further includes: Calculate the cosine similarity between the authentication vector and the fragment identifier vector; When the cosine similarity is greater than or equal to a preset similarity threshold, the match is deemed successful, and the context reading request is allowed. When the cosine similarity is less than the preset similarity threshold, the match is deemed unsuccessful, and the context reading request is blocked.

7. The method according to claim 4, characterized in that, The process involves employing the intelligent agent to execute corresponding subtasks based on the context information, performing security checks on data access during execution, and obtaining business processing results based on the context information and the check results, including: The business intelligence agent executes the corresponding sub-task based on the context information; When the subtask needs to query the database, the natural language query agent is invoked to convert the natural language query requirement into a raw database query statement. Three-layer security verification is performed on the raw database query statement. After the verification is passed, the database query is executed to obtain the query result. The business agent generates a business processing result based on the context information and the query result. When the subtask does not require querying the database, the business intelligence agent directly generates the business processing result based on the context information.

8. The method according to claim 1, characterized in that, The step of generating a fusion scheduling result based on the business processing result includes: The business processing results are summarized, and conflicting and duplicate content between different business processing results is eliminated to obtain a standard business result; Based on the standard business results, a complete response text with a consistent tone is generated, and the complete response text is used as the fusion scheduling result.

9. A multi-agent fusion scheduling system, characterized in that, The system is used to execute a multi-agent fusion scheduling method according to any one of claims 1 to 8, the system comprising: The scheduling center is used to obtain service object requests, parse and decompose the service object requests to obtain scheduling planning information, which includes each sub-task. And according to the scheduling planning information, each subtask is matched to a preset intelligent agent; A context memory pool is used to obtain context information corresponding to the service object request; A security verification engine is used to employ the intelligent agent to execute corresponding subtasks based on the context information, perform security verification on data access during the execution process, and obtain business processing results based on the context information and verification results. The result fusion unit is used to generate fusion scheduling results based on the business processing results.

10. A multi-agent fusion scheduling device, characterized in that, The device includes: The task splitting module is used to obtain service object requests, parse and split the service object requests to obtain scheduling planning information, which includes each sub-task. The scheduling and reading module is used to match each subtask to a preset intelligent agent according to the scheduling planning information, and to obtain the context information corresponding to the service object request; The business execution module is used to employ the intelligent agent to execute corresponding sub-tasks according to the context information, perform security verification on data access during the execution process, and obtain business processing results based on the context information and verification results. The result fusion module is used to summarize the business processing results and generate fusion scheduling results.

11. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the steps of the multi-agent fusion scheduling method according to any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the steps of the multi-agent fusion scheduling method according to any one of claims 1 to 8.

13. A computer program product, characterized in that, Includes computer instructions for causing a computer to perform the steps of a multi-agent fusion scheduling method according to any one of claims 1 to 8.

Images