Multi-agent task arrangement method based on vector knowledge base and computer device

Abstract

This application discloses a multi-agent task orchestration method and computer device based on a vector knowledge base, relating to the fields of artificial intelligence and information technology. The method includes: a first agent generating semantic data blocks with scene meta-tags from multi-source unstructured data of a scene, encoding them into knowledge vectors, constructing a vector knowledge base with traceability channels, and retrieving matching and related knowledge vectors according to user instructions; a second agent generating prompt words by combining matching knowledge with user needs, driving a large model to decompose sub-tasks with dependencies; a third agent retrieving computing power and tool performance parameters to generate planning prompt words, outputting an execution plan through model inference; and a fourth agent executing the sub-tasks according to the plan scheduling tool. This method can externally store business knowledge, split the inference chain, reduce dependence on large parameter models, and improve orchestration accuracy and iteration efficiency.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and information technology, and in particular to a multi-agent task orchestration method and computer device based on a vector knowledge base. Background Technology

[0002] With the rapid development of deep learning and large language model (LLM) technologies, utilizing multiple agents to collaboratively process complex business tasks has become a current technological hotspot.

[0003] However, when faced with large-scale, multi-source unstructured data from specific vertical industries, such as tobacco, healthcare, or finance, multi-agent task orchestration schemes mostly rely on preset rule bases, structured knowledge bases, or fixed prompt word templates to complete task decomposition and orchestration. This makes it impossible to effectively perform semantic parsing and feature fusion on fragmented, unformatted unstructured data, making it difficult to extract the contextual knowledge contained in the data. At the same time, most existing multi-agent scheduling systems adopt a single, fixed task execution mode. Task invocation, tool matching, and sub-task scheduling lack dynamic adaptation capabilities and cannot dynamically adjust execution strategies according to scenario knowledge and real-time task requirements. When faced with complex and ever-changing scenario-based tasks, problems such as task execution disorder, sub-task connection failure, and low task completion success rate are prone to occur, which seriously restricts the automated orchestration capabilities and practicality of multi-agent systems in complex real-world scenarios. Summary of the Invention

[0004] In view of this, this application provides a multi-agent task orchestration method and computer device based on a vector knowledge base, aiming to solve the technical problems of high randomness and lack of scene constraints in the existing agent orchestration.

[0005] According to a first aspect of this application, a multi-agent task orchestration method based on a vector knowledge base is provided. The method is applied to an orchestration system, which includes: a lightweight large language model and an agent collaboration unit. The agent collaboration unit drives the large language model to complete automated task orchestration and execution scheduling. The agent collaboration unit includes a first agent, a second agent, a third agent, and a fourth agent. The method includes: The first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags, calls a large language model to transform the semantic data blocks into low-dimensional knowledge vectors, and constructs a vector knowledge base that integrates semantic features and scene features; wherein, the vector knowledge base contains a tracing channel from the knowledge vectors to the unstructured data; The first intelligent agent performs a similarity search on the vector knowledge base based on the semantic feature vector of the user instruction to obtain a knowledge vector that matches the user instruction; The second agent generates a first prompt word in response to the knowledge vector and user instruction data transmitted by the first agent, and inputs the first prompt word into the large language model. The large language model is then used to decompose the task to obtain multiple sub-tasks with dependencies. The third agent responds to the multiple sub-tasks transmitted by the second agent, calls the first agent to obtain the available system resources and the performance parameters of the system tools, and inputs the available system resources, the performance parameters, and the second prompt words generated collaboratively by the multiple sub-tasks into the large language model. Through the large language model, planning and reasoning are performed to obtain the execution plan that includes time windows and computing resource allocation schemes. The fourth agent responds to the execution plan and dynamically invokes system tools to execute the multiple sub-tasks based on the execution plan.

[0006] Optionally, the first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags, including: The first intelligent agent performs standardization processing on the multi-source unstructured data to obtain semi-structured data; The first intelligent agent identifies scene metadata in the semi-structured data and binds the scene metadata to the semi-structured data to obtain tagged structured data; The first intelligent agent performs semantic slicing on the structured data to obtain the semantic data block, and establishes a mapping relationship between the semantic data block, the structured data, and the multi-source unstructured data.

[0007] Optionally, the first agent invokes a lightweight large language model to transform the semantic data block into a low-dimensional knowledge vector, constructing a vector knowledge base that integrates semantic features and scene features, including: The first intelligent agent inputs the semantic data block and its scene metadata tag into a lightweight large language model, and vectorizes the semantic data block into a semantic feature vector through the three-layer encoding operator of the large language model. The semantic feature vector of the semantic data block and the scene feature vector corresponding to the scene metadata tag are weighted and fused with a preset fusion weight to generate an enhanced feature vector. The enhanced feature vector is then compressed to obtain a low-dimensional dense feature vector corresponding to the semantic data block. The first intelligent agent constructs the vector knowledge base using multiple low-dimensional dense feature vectors as knowledge vectors.

[0008] Optionally, the second agent generates a first prompt word based on the knowledge vector transmitted by the first agent and the user instruction's requirement data, including: The second agent converts the knowledge vector back into scene constraint rule text; The second intelligent agent concatenates the user instruction's requirement data with the scene constraint rule text to form the first prompt word.

[0009] Optionally, the method further includes: The third agent invokes the first agent to retrieve historical task templates that match the multiple sub-tasks; If the matching degree between the multiple subtasks and the historical task template is greater than the matching degree threshold, the third intelligent agent reuses the execution plan corresponding to the historical task template. If the matching degree between the multiple subtasks and the historical task template is less than or equal to the matching degree threshold, the third agent calls the first agent to obtain the performance parameters of the system's available resources and system tools.

[0010] Optionally, the fourth intelligent agent includes a control and scheduling intelligent agent and at least one subtask execution intelligent agent; the fourth intelligent agent dynamically invokes system tools to execute the multiple subtasks based on the execution plan, including: The control and scheduling agent parses the execution plan and generates scheduling instructions based on the dependencies of the parsed subtasks; The control and scheduling agent assigns different subtasks to the corresponding subtask execution agents based on the scheduling instructions; The subtask execution agent responds to the received subtask by invoking the matching system tool to perform the task operation.

[0011] Optionally, the orchestration system further includes a fifth intelligent agent; the method further includes: The fifth agent monitors the execution status of the system tools. If any subtask execution abnormality is detected, the fifth agent extracts the context stack information of the subtask and transmits the context stack information to the third agent. The third agent, based on the context stack information, calls the large language model again to adjust the plan for the multiple subtasks, and obtains a revised execution plan. The fourth agent responds to the modified execution plan by re-invoking the corresponding system tool to execute any of the sub-tasks until execution is successful or the preset maximum retry threshold is reached.

[0012] Optionally, the method further includes: When the subtask is completed, the fifth agent transmits the task completion feature obtained by feature extraction of the execution result to the first agent. The first agent updates the task completion features to the vector knowledge base.

[0013] Optionally, the orchestration system further includes a sixth intelligent agent; the method further includes: The sixth intelligent agent responds to the user command by inputting the user command into the large language model for intent parsing and obtaining the demand information. After the subtask is completed, the sixth intelligent agent obtains the user feedback content; The first intelligent agent updates the user feedback content transmitted by the sixth intelligent agent to the vector knowledge base.

[0014] According to a second aspect of this application, a readable storage medium is provided on which a program or instructions are stored, which, when executed by a processor, implement the steps of the above-described multi-agent task orchestration method based on a vector knowledge base.

[0015] According to a fourth aspect of this application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described multi-agent task orchestration method based on a vector knowledge base.

[0016] By employing the aforementioned technical solution, on the one hand, the first intelligent agent transforms multi-source unstructured data into semantic data blocks with scene metadata tags and performs dimensionality reduction. This results in a vector knowledge base that not only possesses textual semantic features but also binds environmental features such as topology and rules of the target scene. Furthermore, it records the mapping relationships during the transformation of multi-source unstructured data, forming a traceability channel. Thus, relying on vector similarity retrieval, it can accurately match the knowledge vectors corresponding to user commands, effectively solving the problems of ineffective utilization of unstructured data and the lack of transparency in the vectorization process in traditional solutions. On the other hand, by combining the second and third intelligent agents, only a small amount of accurately matched knowledge fragments and user requirements are fed into the large model. The large input language model completes task decomposition, logical dependency analysis, and execution planning. Therefore, while ensuring that task planning aligns with actual business scenarios, the large model no longer bears the burden of massive original text reading and coarse filtering, significantly reducing its dependence on large parameter models. Finally, the fourth agent completes each sub-task according to the execution plan dynamic scheduling system tool, realizing full-process automation and dynamic scheduling of tasks. This simplifies the integration and scheduling complexity of the multi-agent orchestration system, and significantly reduces the difficulty of each orchestration step, enabling stable output results even for lightweight large models.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating the multi-agent task orchestration method based on a vector knowledge base provided in an embodiment of this application is shown. Figure 2 A schematic diagram of the electronic structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0019] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0020] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0021] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “attached” to another element, it can be directly connected or attached to the other element, or there may be intermediate elements present. Furthermore, the term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0022] Exemplary embodiments according to this application will now be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments may be implemented in many different forms and should not be construed as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided so that the disclosure of this application is thorough and complete, and that the concept of these exemplary embodiments is fully conveyed to those skilled in the art.

[0023] This embodiment provides a multi-agent task orchestration method based on a vector knowledge base, which is applied to an orchestration system.

[0024] This orchestration system comprises two core components: a lightweight large language model (LLM) and an agent collaboration unit. The large language model serves as the system's central reasoning and control core, responsible for understanding natural language instructions and providing generation capabilities. The agent collaboration unit is the driving unit for system-level automated task orchestration and scheduling. In the core embodiment of this application, the agent collaboration unit is finely divided into multiple functionally decoupled first, second, third, and fourth agents.

[0025] like Figure 1 As shown, the multi-agent task orchestration method for vector knowledge bases includes the following steps: Step 101: The first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags, calls a large language model to transform the semantic data blocks into low-dimensional knowledge vectors, and constructs a vector knowledge base that integrates semantic features and scene features.

[0026] The vector knowledge base includes a traceability channel from knowledge vectors to unstructured data. This channel records the original data source, data collection time, scenario-related tags, semantic segmentation nodes, and data version information for each knowledge vector, achieving a one-to-one mapping between knowledge vectors and the original unstructured data. Unstructured data refers to fragmented, unstructured business and operational text data within the target business scenario, such as system logs, equipment manuals, network topology configuration texts, and historical troubleshooting manuals. The data can be in the form of documents, text, tables, or question-and-answer data.

[0027] In this embodiment, a first intelligent agent performs semantic segmentation, scene metadata annotation, and vectorization encoding on multi-source unstructured data to construct a scene-based vector knowledge base with traceability capabilities. This enables the standardization and unified vectorization representation of multi-source heterogeneous unstructured data, integrates semantic and scene features to solve the shortcomings of poor scene adaptability and single features in traditional knowledge storage, and ensures the traceability of knowledge by relying on the traceability channel, reducing the black-box problem.

[0028] Step 102: The first intelligent agent performs a similarity search in the vector knowledge base based on the semantic feature vector of the user instruction to obtain a knowledge vector that matches the user instruction.

[0029] In this embodiment, by vectorizing user natural language commands and utilizing cosine similarity or inner product, precise nearest neighbor retrieval is performed in the contextualized vector knowledge base based on the tracing channel from knowledge vectors to unstructured data. This overcomes the matching limitations of traditional keyword retrieval. Based on the implicit semantics of user commands and scenario requirements, highly suitable professional knowledge vectors are accurately recalled, irrelevant and redundant knowledge is eliminated, and the accuracy of knowledge matching and scenario adaptability are significantly improved, providing accurate knowledge input for subsequent task planning.

[0030] It is understandable that the first agent responds to the user's command and uses a large model to vectorize the user's command to obtain the semantic feature vector of the user's command. Alternatively, the sixth agent, which serves as the interaction interface, can be triggered to perform intent parsing and then vectorize the user's command after receiving the user's request information to obtain the semantic feature vector.

[0031] Step 103: The second agent generates a first prompt word in response to the knowledge vector and user instruction data transmitted by the first agent, and inputs the first prompt word into a large language model. The large language model is used to decompose the task and obtain multiple sub-tasks with dependencies.

[0032] In this embodiment, a second intelligent agent integrates user needs and retrieved knowledge to generate standardized first prompt words, which assists the large language model in carrying out hierarchical reasoning. This avoids reasoning biases caused by fuzzy natural language input and knowledge gaps, accurately decomposes complex overall tasks, sorts out the logical dependencies of each sub-task, and outputs structured sub-tasks, thus achieving accurate transformation from needs to tasks.

[0033] Step 104: In response to the multiple sub-tasks transmitted by the second agent, the third agent calls the first agent to obtain the available system resources and the performance parameters of the system tools. It then inputs the available system resources, performance parameters, and the second prompt words generated collaboratively by the multiple sub-tasks into the large language model. Through planning and reasoning using the large language model, it obtains an execution plan that includes time windows and computing resource allocation schemes.

[0034] The available system resources include the current CPU idle rate, GPU memory availability, and remaining network bandwidth of each node. Performance parameters include average execution latency and success rate. The available system resources and current performance parameters of the system tools can be dynamically acquired by the first intelligent agent through a lightweight probe executed periodically in the background.

[0035] In this embodiment, the second agent can obtain the available system resources and performance parameters of system tools through the first agent, and then combine the parsed task, available system resources, and performance parameters to generate a second prompt word. The second prompt word is input into a large model, and the large language model performs comprehensive planning and reasoning with long contexts to generate a customized execution plan containing precise time windows and computing resource allocation schemes. This avoids the logical confusion and excessive input overhead caused by the large model carrying multiple types of mixed information, effectively avoids computing power conflicts and time limits exceeding limits during multi-task concurrency, and reduces the dependence on the autonomous reasoning capabilities of the large model during the planning stage through pre-input physical constraints, ensuring the planning feasibility and execution stability of the lightweight model in complex scheduling scenarios.

[0036] Step 105: The fourth agent dynamically calls system tools to execute multiple sub-tasks based on the execution plan.

[0037] The multi-agent task orchestration method based on a vector knowledge base provided in this application has two main aspects. First, the first agent transforms multi-source unstructured data into semantic data blocks with scene metadata tags and performs dimensionality reduction. This results in a vector knowledge base that not only possesses textual semantic features but also binds environmental features such as topology and rules of the target scene. Furthermore, it records the mapping relationships during the transformation of multi-source unstructured data, forming a traceability channel. Thus, relying on vector similarity retrieval, it can accurately match the knowledge vectors corresponding to user commands, effectively solving the problems of ineffective utilization of unstructured data and the lack of transparency in the vectorization process in traditional solutions. Second, by combining the second and third agents, only a small amount of accurately matched knowledge fragments and user requirements are fed into a large model. The large input language model completes task decomposition, logical dependency analysis, and execution planning. Therefore, while ensuring that task planning conforms to actual business scenarios, the large model no longer bears the burden of massive original text reading and coarse filtering, greatly reducing the dependence on large parameter models. Finally, the fourth agent completes each sub-task according to the execution plan dynamic scheduling system tool, realizing full-process automation and dynamic scheduling of tasks. This simplifies the integration and scheduling complexity of the multi-agent orchestration system, and significantly reduces the difficulty of each orchestration step, enabling stable output results even for lightweight large models.

[0038] In one embodiment, in step 101, the first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags. Specifically, this includes: the first intelligent agent standardizing the multi-source unstructured data to obtain semi-structured data; the first intelligent agent identifying scene metadata in the semi-structured data and binding the scene metadata with the semi-structured data to obtain tagged structured data; and the first intelligent agent performing semantic slicing on the structured data to obtain semantic data blocks and establishing a mapping relationship between semantic data blocks, structured data, and multi-source unstructured data.

[0039] Semi-structured data refers to data with clearly defined segmentation markers and key fields. Standardization processing includes, but is not limited to, removing noise, garbled characters, and redundant formatting from the original multi-source unstructured data. Scenario metadata includes, but is not limited to, data source device model, application component name, timestamp range, topology level, and business security level.

[0040] In this embodiment, the first intelligent agent explicitly binds these scene metadata as attribute tags to the semi-structured data. This assigns clear scene attributes and business dimension identifiers to the data, resolving the ambiguity of multi-source data attribution. It also provides scene selection criteria for subsequent slicing, vectorization, and retrieval, ensuring knowledge adaptability. Furthermore, the slicing granularity is dynamically adjusted based on semantic integrity, business logic boundaries, and metadata tag dimensions. Large-granularity semi-structured data is broken down into multiple small-granularity data blocks that can independently express complete semantics. Simultaneously, a full-link mapping relationship is established between semantic data blocks and the original structured and unstructured data, providing complete link support for subsequent vector tracing. This avoids semantic breaks or excessively large granularities caused by fixed slicing rules, solving the semantic breakage problem caused by fixed slicing. It makes the granularity of semantic data blocks more uniform and traceable, ensuring that each data processing step is verifiable, and helps improve the accuracy of subsequent vectorization.

[0041] Understandably, text can be sliced ​​based on natural paragraphs, syntactic tree boundaries, or semantic integrity. Alternatively, it can be sliced ​​in combination with business dimensions of scenario metadata, such as device model, topology level, and business security level, to prioritize the semantic integrity of core business dimensions and adapt to the retrieval granularity requirements of different scenarios.

[0042] For example, multi-source unstructured data sets It covers various data sources, including industry documents, text, tables, and Q&A data. By integrating target scenario tool parsing logic with general text cleaning rules, it transforms raw unstructured data into well-structured semi-structured data. To eliminate data noise and format differences, the formula is: ; In the formula, A finely tuned OCR format conversion operator for the target scenario, used for text conversion of PDF and scanned documents; Text cleaning operators adapted to target scenarios include word segmentation, stop word removal, and industry terminology normalization functions for standardized processing.

[0043] (1) After completing the structured annotation, extract the core metadata and use the splicing operator to bind scene metadata tags to the semi-structured data to generate tagged structured data. The formula is: ; In the formula, For splicing operators; Metadata extraction operators; This is a metadata tag set for the target scenario, which can include core tags such as industry standard number, indicator parameters, clause level, and data type.

[0044] Based on the principles of scene metadata and semantic independence, large-granular structured data is decomposed into small-granular data blocks adapted for retrieval, while retaining key metadata information to make it suitable for vectorization and retrieval tasks, generating the final data block set. The formula is: ; A mapping relationship is established simultaneously between semantic data blocks, structured data, and multi-source unstructured data. This mapping relationship can be represented as: ; ; In the formula, m is the total number of data blocks.

[0045] Taking tobacco product instructions as an example, some semantic data blocks are shown in Table 1.

[0046] Table 1

[0047] In one embodiment, in step 101, the first agent calls a lightweight large language model to convert semantic data blocks into low-dimensional knowledge vectors and constructs a vector knowledge base that integrates semantic features and scene features. Specifically, the first agent inputs the semantic data block and its scene metadata tags into the lightweight large language model, vectorizes the semantic data block into semantic feature vectors through the three-layer encoding operator of the large language model, and performs weighted fusion of the semantic feature vector of the semantic data block and the scene feature vector corresponding to the scene metadata tags with preset fusion weights to generate enhanced feature vectors, and performs dimensionality reduction and compression on the enhanced feature vectors to obtain low-dimensional dense feature vectors corresponding to the semantic data block; the first agent constructs a vector knowledge base using multiple low-dimensional dense feature vectors as knowledge vectors.

[0048] In this embodiment, the first agent inputs semantic data blocks and their associated scene metadata tags into a large language model. Deep feature extraction is performed using a three-layer encoding operator pre-trained or fine-tuned within the large language model. This not only ensures that the generated vectors are consistent with the semantic system of the large model in the multi-agent collaborative system, but also effectively eliminates the semantic ambiguity of general vectors in specific industries through weighted fusion of scene metadata. Finally, dimensionality reduction and compression reduce the computational consumption of vector retrieval and storage while preserving the core semantics, thus achieving high-precision, low-cost professional knowledge retrieval and application in resource-constrained, lightweight business scenarios.

[0049] Specifically, the three-layer coding operator is: Layer 1: Lexical and text embedding encoding operators, used to encode single data blocks Transform into a high-dimensional general semantic feature vector This completes the initial conversion from text to vector, using the following formula: ; In the formula, For large model word embedding operators.

[0050] The second layer consists of a long-distance contextual dependency operator, used to capture the position and preceding relationships of the sliced ​​data block within the entire document or configuration context. It integrates general semantic features with industry-specific features, balancing general expression with industry adaptability, and eliminating the semantic ambiguity of general vectors within specific industry contexts. The formula is as follows: ; In the formula, The scene feature vector is the target scene. ∈[0,1] is the preset fusion weight, which is empirically set to 0.6 in the tobacco scenario to take into account both general semantics and industry characteristics.

[0051] The fourth layer is the scene target alignment operator, which transforms high-dimensional enhanced features into low-dimensional dense vectors, reducing the computational cost of retrieval and storage, adapting to lightweight business scenarios, and ultimately generating a full set of vectors. The formula is: , ; In the formula, Principal component analysis compression operator, The dimension of the target vector is 128 or 256 in lightweight scenarios.

[0052] For example, using semantic data blocks extracted from tobacco product instructions as input, the system utilizes an industry-fine-tuned embedding model to map the text into high-dimensional vectors. This ensures that semantically similar data maintains proximity in the vector space, resulting in knowledge vectors. By fusing scene feature information, the vectors are enhanced to reflect not only textual semantics but also industry attributes, thereby improving the model's discriminative ability in professional scenarios. Sample vectorization results are shown in Table 2.

[0053] Table 2

[0054] In one embodiment, in step 103, the second intelligent agent generates a first prompt word based on the knowledge vector transmitted by the first intelligent agent and the user instruction requirement data. Specifically, the second intelligent agent restores the knowledge vector to the scene constraint rule text; the second intelligent agent concatenates the user instruction requirement data and the scene constraint rule text to form the first prompt word.

[0055] In this embodiment, the second agent transforms abstract knowledge vectors into explicit scene constraint rule text, which is then concatenated with user instructions to form the first prompt word. This not only breaks the black-box state of vector retrieval but also provides clear contextual boundaries and hard behavioral guidelines for the large model. Thus, it effectively suppresses the illusion of a large model at the source of task orchestration, improving the accuracy and controllability of subsequent planning and reasoning.

[0056] In one embodiment, after determining multiple sub-tasks, the multi-agent task orchestration method based on a vector knowledge base further includes: a third agent calling a first agent to retrieve historical task templates that match the multiple sub-tasks; if the matching degree between the multiple sub-tasks and the historical task templates is greater than a matching degree threshold, the third agent reuses the execution plan corresponding to the historical task templates; if the matching degree between the multiple sub-tasks and the historical task templates is less than or equal to the matching degree threshold, the third agent calling a first agent to obtain the performance parameters of available system resources and system tools.

[0057] In this embodiment, for subtasks with a matching degree higher than a threshold of historically mature tasks, the execution plan, which has been validated in practice, can be directly reused. This can significantly shorten the task planning cycle, reduce the frequency of large model calls, and lower system computing power overhead, while ensuring the stability and consistency of high-frequency repetitive business execution. Conversely, for new or complex tasks with a matching degree lower than the threshold, the third agent can obtain the performance parameters of available system resources and system tools through the first agent, and perform comprehensive planning and reasoning with long contexts through a large language model to generate a customized execution plan that includes precise time windows and computing power resource allocation schemes. In this way, task execution conflicts can be accurately avoided, the overall system resource utilization can be optimized, and various unknown business scenarios can be flexibly adapted. Ultimately, the optimal balance between efficiency, cost, and adaptability in multi-agent task planning is achieved, significantly improving the overall operating performance and business carrying capacity of the entire orchestration system.

[0058] In one embodiment, the fourth intelligent agent includes a control and scheduling intelligent agent and at least one subtask execution intelligent agent. In step 105, the fourth intelligent agent dynamically invokes system tools to execute multiple subtasks based on the execution plan, specifically including: Step 105-1: The control and scheduling agent parses the execution plan and generates scheduling instructions based on the dependencies of the parsed subtasks.

[0059] Step 105-2: The control and scheduling agent assigns different subtasks to the corresponding subtask execution agents based on scheduling instructions.

[0060] Step 105-3: The subtask execution agent responds to the received subtask by calling the matching system tool to perform the task operation.

[0061] In this embodiment, after the fourth agent receives the execution plan output by the large language model, its internal control and scheduling agent performs structured parsing of the plan using a lightweight lexical parser, identifies the directed graph topology of subtasks, generates scheduling directed graph instructions, and performs global task scheduling. Subsequently, based on the dependency order of the scheduling instructions, the control and scheduling agent dynamically distributes different subtasks to corresponding specialized subtask execution agents when the preconditions of each subtask are met. The subtask execution agents dynamically call and match the best underlying system tools to execute specific atomic-level operations according to their corresponding task operation semantics. This solves the problems of excessive load on a single execution agent and chaotic task scheduling, clearly defines the boundaries of responsibility between scheduling decisions and task execution, ensures that subtasks are executed in an orderly manner according to their dependencies, and avoids task execution conflicts and logical errors. It also supports parallel scheduling and execution of subtasks without dependencies, significantly improving overall task execution efficiency, and can dynamically expand the number of subtask execution agents according to business needs, enhancing the system's concurrent processing capabilities and business carrying capacity. In addition, through precise matching and calling of subtasks and system tools, the accuracy and stability of task execution can also be effectively improved.

[0062] In one embodiment, such as Figure 1 As shown, the orchestration system also includes a fifth agent. The multi-agent task orchestration method based on a vector knowledge base also includes: Step 106: The fifth agent monitors the execution status of the system tool. If any subtask execution abnormality is detected, the context stack information of any subtask is extracted and transmitted to the third agent.

[0063] The context stack information includes the exception-throwing code, current system environment variables, input / output parameter streams, and current physical resource consumption residuals.

[0064] Step 107: The third agent calls the large language model again to adjust the plan for multiple subtasks based on the context stack information, and obtains the revised execution plan.

[0065] Step 108: In response to the revised execution plan, the fourth agent re-invokes the corresponding system tool to execute any subtask until the execution is successful or the preset maximum retry threshold is reached.

[0066] In this embodiment, during the process of the fourth agent calling the tool to execute a subtask, the fifth agent can monitor the execution status in real time. Once any subtask execution anomaly is detected, the fifth agent will intervene in the task execution and extract the current context stack information of the subtask. The fifth agent will then transmit this context stack information back to the third agent in real time. The third agent will re-encode the current failure stack as a negative feedback example into the second prompt word and trigger the large language model to adjust the plan. The large language model, combined with the cause of failure, will dynamically adjust or regenerate a targeted corrected execution plan while maintaining the progress of other unaffected subtasks. In response to the corrected execution plan, the fourth agent will call the adjusted tool again to attempt execution, achieving runtime online fault self-healing until successful execution or the system's preset maximum retry threshold is reached. Thus, without requiring manual intervention for troubleshooting and repair, and without interrupting the execution progress of unaffected subtasks, the runtime online fault self-healing and closed-loop anomaly handling of the multi-agent task orchestration system can be completed, helping to reduce unnecessary computing power consumption and task execution latency.

[0067] Furthermore, once the entire subtask flow has been successfully executed under the drive of the fourth agent, i.e., when the fifth agent detects that the subtask has been completed, the fifth agent extracts features from the successful execution log stream and transmits the obtained task completion features to the first agent. The first agent can then use the task completion features as new experience samples, updating and asynchronously appending them to the vector knowledge base in the same way as in step 101. This allows for continuous optimization of the execution plan and anomaly handling strategies for subsequent similar tasks, achieving iterative evolution of system capabilities.

[0068] In one embodiment, the orchestration system further includes a sixth agent. The multi-agent task orchestration method based on a vector knowledge base further includes: the sixth agent responding to user instructions by inputting the user instructions into a large oracle model for intent parsing to obtain requirement information; the sixth agent obtaining user feedback content after the subtask is completed; and the first agent updating the vector knowledge base with the user feedback content transmitted by the sixth agent.

[0069] In this embodiment, a sixth agent serves as the human-computer interaction interface. A large model is used to accurately transform unstructured user commands into structured requirements that the system can understand. User feedback is dynamically collected after task execution, and the first agent incrementally updates the vector knowledge base with this feedback. This not only improves the system's accuracy in understanding the user's true intent but also enables continuous iteration and adaptive optimization of system knowledge, significantly enhancing the personalization and retrieval accuracy of subsequent task orchestration.

[0070] For example, the roles and collaboration framework of the intelligent agents are pre-constructed. The first intelligent agent manages short-term conversation memory and long-term historical memory. It stores dialogue context, user preferences, domain knowledge, etc., and can efficiently retrieve information based on current needs, providing information support to other agents. The second intelligent agent receives the user's original instructions and decomposes and refines them into a series of clear, specific, and executable sub-tasks. The third intelligent agent dynamically adjusts and optimizes the execution plan. The fourth intelligent agent has the ability to call external tools, APIs, or functions to perform specific operations according to instructions, such as obtaining real-time information, running code, and manipulating databases, transforming abstract instructions into concrete results. The fifth intelligent agent monitors the task execution process, dynamically adjusting and optimizing the original execution plan based on intermediate results, unexpected situations, or feedback. It also reviews the output results of other agents, enabling the system to cope with uncertainty. The sixth intelligent agent serves as the interface between the user or other systems and the agent system. It receives user input, performs operations on behalf of the user (such as confirmation and authorization), and returns the system's final results to the user in a user-friendly format.

[0071] The operation process of this orchestration system is broken down into six stages, as detailed below: Phase A: Through the collaboration of the sixth intelligent agent, the first intelligent agent, and the big oracle model, unstructured user input requests are transformed into structured tasks that the system can understand. : ; in, This indicates that the first agent responds to the request of the sixth agent by retrieving a knowledge vector from the knowledge base. KB U ; This indicates that demand information is fused through a large model. U With knowledge vectors KB U .

[0072] Phase B: Through the collaboration of the second intelligent agent, the first intelligent agent, and the big oracle model, structured tasks are... Split into a set of subtasks Dependency D This allows complex tasks to be broken down into a combination of subtasks, which can be represented as: ; in, This indicates that the first intelligent agent provides a vector knowledge base. ; This indicates that subtasks are generated from a large model. 。

[0073] Phase C: Through the collaboration of the third agent, the first agent, and the big oracle model, based on the set of sub-tasks. Dependency D Generate an execution plan based on system parameters. This enables the system to complete work tasks according to a given plan and execute the plan. It can be represented as: ; in, This indicates that the first intelligent agent provides the system parameters; KB res Indicates computing resource parameters; KB tool-perf Indicates tool performance parameters; This indicates that an execution plan is generated using a large model. .

[0074] Phase D: Through the collaboration of the fifth, fourth, third, and first agents and the Grand Oracle model, sub-tasks are executed according to the execution plan, exceptions are handled, and a task execution result set is generated. Execution functions are divided into those with and without exceptions. (1) No abnormal execution: ; (2) Abnormal execution: ; in, This indicates that the first intelligent agent provides tool execution parameters. KB tool-para ; E This indicates a task execution error. This is represented as a large model diagnostic function, which generates corrected parameters and re-executes the function. R This is the set of results from the subtasks.

[0075] Phase E: Through the collaboration of the fifth agent, the first agent, and the big oracle model, a qualification check is provided to verify whether the task execution results meet the standards and to optimize the results that do not meet the standards.

[0076] ; right For the evaluation result set, E val The formula for making the judgment is as follows: ; in, This represents the evaluation criteria dataset provided by the first intelligent agent; This indicates that optimization suggestions are generated using a large model; This represents an iterative function that adjusts the task set. With the execution plan Re-execute until satisfied. ; θ The system defines a compliance threshold between 0 and 1, and outputs the compliance result. .

[0077] Phase F: Through the collaboration of the sixth agent, the first agent, and the big oracle model, the final result is returned to the user. The results are formatted, and the knowledge base of the first agent is updated.

[0078] ; ; in, Format ind Indicates the formatting rules; F This indicates user feedback content; This indicates that user feedback content is integrated with a vector knowledge base through a large model; KB' This represents the updated vector knowledge base, which can be reused next time.

[0079] Let's take the user-input command "Query the core indicators and corresponding clauses of GB 5606.4-2022 for sensory evaluation of cigarettes, for today's quality inspection report" as an example. The sixth intelligent agent drives the big oracle model to parse the demand intent and encapsulate it into an unstructured JSON command. At the same time, it requests the first intelligent agent to access knowledge in the tobacco national standard domain and quality inspection business constraints. The first intelligent agent uses the semantic feature vector of the user command to retrieve relevant industry constraint knowledge of the tobacco national standard and quality inspection report scenario.

[0080] The second agent uses the knowledge vectors and user instruction data obtained by the first agent to generate the first prompt word. It then uses a large language model to parse the task and obtain a standardized structured task: json{"structured_task_Tstr":{"task_objective":"Extract core indicators from chapters of GB5606.4-2022 'Sensory Evaluation of Cigarettes'","constraints":["Covering all indicators","Clauses accurate to chapter-section-sub-clause"],"output_format":["Indicator name","Clause number","Clause summary within 10 characters"]}}. The second agent uses the structured task as input to drive the large model... Based on domain knowledge, the model is decomposed into 5 ordered subtasks: json{"subtask_set_Tsub":[{"t1":"Locate document, tool: document retrieval tool"},{"t2":"Parse pages 3-4, tool: tobacco OCR"},{"t3":"Extract indicators, tool: tobacco NER"},{"t4":"Associate clauses, tool: clause matching tool"},{"t5":"Integrate results, tool: formatting tool"}],"dependency_matrix_D":[0,0,0,0,0],[1,0,0,0,0],[0,1,0,0,0],[0,0,1,0,0],[0,0,0,1,0]]}.

[0081] The third agent requests computing resources (CPU / memory), performance parameters of various tobacco tools, and business time constraints from the first agent. It then leads the large model in optimizing time allocation and computing power allocation based on dependencies and resource consumption constraints, generating a time-series execution plan: json{ "optimal_execution_plan_Popt": [ {"t1":"00:00-00:01,Tool: Document Retrieval, 0.5 cores"}, {"t2":"00:01-00:03,Tool: Tobacco OCR, 1 core"}, {"t3":"00:03-00:04,Tool: Tobacco NER, 1 core"}, {"t4":"00:04-00:05,Tool: Clause Matching, 0.5 cores"},{"t5":"00:05-00:06,Tool: Formatting, 0.5 cores"} ], "plan_note": "Total time 6 minutes, no computing power conflict"}

[0082] The fourth intelligent agent issues scheduling instructions and calls various tools to execute each task, obtaining the execution result: json {"subtask_result_set_R": [ {"r1":"Document path: / Tobacco National Standard / GB5606.4-2022.pdf"},{"r2_corrected":"Parsed text contains 5 indicator clauses (including off-flavors 3.2.5)"}, {"r3":"Indicator list: [Aroma, Harmony, Aftertaste, Irritation, Off-flavors]"}, {"r4":"Relationship table: Aroma → 3.2.1 etc."}, {"r5":"Integrated draft: Aroma 3.2.1 Aroma intensity etc."} ],"execution_status":"Total time 7 minutes, anomaly fixed"}.

[0083] Furthermore, using the same input parameters, the method provided in this application embodiment was compared with traditional single-agent technology and general RAG intelligent agent orchestration technology. The core performance indicators are shown in Table 3. As can be seen from Table 3, this application, through dual-feature three-layer operator encoding of the first intelligent agent and resource-aware first prompt word orchestration of the second intelligent agent, directly reduces the illusion rate from over 20% to below 1.5%. Moreover, relying on the stack self-healing of the fifth intelligent agent, the success rate of task anomaly self-healing soars to 89.5%. The experimental data fully and conclusively demonstrate the significant technological breakthroughs and substantial technological advancements of this application.

[0084] Table 3

[0085] The multi-agent task orchestration method based on vector knowledge base provided in this application can be applied to terminals, servers, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0086] It should be noted that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0087] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data that have been permitted by the user or fully permitted by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to allow or refuse.

[0088] Based on the above, Figure 1 Accordingly, embodiments of this application also provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method shown is a multi-agent task orchestration method based on a vector knowledge base.

[0089] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0090] Based on the above, Figure 1 The method shown is for achieving the above objectives, such as... Figure 2 As shown in the figure, this application embodiment also provides a computer device 400, which includes a processor 401 and a memory 402. The memory 402 stores a program or instructions that can run on the processor 401. When the program or instructions are executed by the processor 401, they implement the above-mentioned... Figure 1 The method shown is a multi-agent task orchestration method based on a vector knowledge base.

[0091] The memory 402 can be used to store software programs and various data. The memory 402 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 402 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 402 in this embodiment includes, but is not limited to, these and any other suitable types of memory.

[0092] Processor 401 may include one or more processing units; optionally, processor 401 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 401.

[0093] Computer equipment can specifically include personal computers, servers, network devices, etc.

[0094] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB ports, card reader ports, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0095] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0097] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0098] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A multi-agent task orchestration method based on a vector knowledge base, characterized in that, The method is applied to an orchestration system, which includes: a lightweight large language model and an intelligent agent collaboration unit. The intelligent agent collaboration unit drives the large language model to complete automated task orchestration and execute scheduling. The intelligent agent collaboration unit includes a first intelligent agent, a second intelligent agent, a third intelligent agent, and a fourth intelligent agent. The method includes: The first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags, calls a large language model to transform the semantic data blocks into low-dimensional knowledge vectors, and constructs a vector knowledge base that integrates semantic features and scene features; wherein, the vector knowledge base contains a tracing channel from the knowledge vectors to the unstructured data; The first intelligent agent performs a similarity search on the vector knowledge base based on the semantic feature vector of the user instruction to obtain a knowledge vector that matches the user instruction; The second agent generates a first prompt word in response to the knowledge vector and user instruction data transmitted by the first agent, and inputs the first prompt word into the large language model. The large language model is then used to decompose the task to obtain multiple sub-tasks with dependencies. The third agent responds to the multiple sub-tasks transmitted by the second agent, calls the first agent to obtain the available system resources and the performance parameters of the system tools, and inputs the available system resources, the performance parameters, and the second prompt words generated collaboratively by the multiple sub-tasks into the large language model. Through the large language model, planning and reasoning are performed to obtain the execution plan that includes time windows and computing resource allocation schemes. The fourth agent responds to the execution plan and dynamically invokes system tools to execute the multiple sub-tasks based on the execution plan.

2. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The first intelligent agent transforms multi-source unstructured data in the target scene into semantic data blocks with scene metadata tags, including: The first intelligent agent performs standardization processing on the multi-source unstructured data to obtain semi-structured data; The first intelligent agent identifies scene metadata in the semi-structured data and binds the scene metadata to the semi-structured data to obtain tagged structured data; The first intelligent agent performs semantic slicing on the structured data to obtain the semantic data block, and establishes a mapping relationship between the semantic data block, the structured data, and the multi-source unstructured data.

3. The multi-agent task orchestration method based on vector knowledge base according to claim 2, characterized in that, The first intelligent agent invokes a large language model to transform the semantic data block into a low-dimensional knowledge vector, constructing a vector knowledge base that integrates semantic features and scene features, including: The first intelligent agent inputs the semantic data block and its scene metadata tag into a lightweight large language model, and vectorizes the semantic data block into a semantic feature vector through the three-layer encoding operator of the large language model. The semantic feature vector of the semantic data block and the scene feature vector corresponding to the scene metadata tag are weighted and fused with a preset fusion weight to generate an enhanced feature vector. The enhanced feature vector is then compressed to obtain a low-dimensional dense feature vector corresponding to the semantic data block. The first intelligent agent constructs the vector knowledge base using multiple low-dimensional dense feature vectors as knowledge vectors.

4. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The second agent generates a first prompt word based on the knowledge vector transmitted by the first agent and the user instruction's requirement data, including: The second agent converts the knowledge vector back into scene constraint rule text; The second intelligent agent concatenates the user instruction's requirement data with the scene constraint rule text to form the first prompt word.

5. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The method further includes: The third agent invokes the first agent to retrieve historical task templates that match the multiple sub-tasks; If the matching degree between the multiple subtasks and the historical task template is greater than the matching degree threshold, the third intelligent agent reuses the execution plan corresponding to the historical task template. If the matching degree between the multiple subtasks and the historical task template is less than or equal to the matching degree threshold, the third agent calls the first agent to obtain the performance parameters of the system's available resources and system tools.

6. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The fourth intelligent agent includes a control and scheduling intelligent agent and at least one subtask execution intelligent agent; The fourth intelligent agent dynamically invokes system tools to execute the multiple sub-tasks based on the execution plan, including: The control and scheduling agent parses the execution plan and generates scheduling instructions based on the dependencies of the parsed subtasks; The control and scheduling agent assigns different subtasks to the corresponding subtask execution agents based on the scheduling instructions; The subtask execution agent responds to the received subtask by invoking the matching system tool to perform the task operation.

7. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The orchestration system further includes a fifth intelligent agent; the method further includes: The fifth agent monitors the execution status of the system tools. If any subtask execution abnormality is detected, the fifth agent extracts the context stack information of the subtask and transmits the context stack information to the third agent. The third agent, based on the context stack information, calls the large language model again to adjust the plan for the multiple subtasks, and obtains a revised execution plan. The fourth agent responds to the modified execution plan by re-invoking the corresponding system tool to execute any of the sub-tasks until execution is successful or the preset maximum retry threshold is reached.

8. The multi-agent task orchestration method based on vector knowledge base according to claim 7, characterized in that, The method further includes: When the subtask is completed, the fifth agent transmits the task completion feature obtained by feature extraction of the execution result to the first agent. The first agent updates the task completion features to the vector knowledge base.

9. The multi-agent task orchestration method based on vector knowledge base according to claim 1, characterized in that, The orchestration system further includes a sixth intelligent agent; the method further includes: The sixth intelligent agent responds to the user command by inputting the user command into the big oracle model for intent parsing and obtaining the demand information. After the subtask is completed, the sixth intelligent agent obtains the user feedback content; The first intelligent agent updates the user feedback content transmitted by the sixth intelligent agent to the vector knowledge base.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-agent task orchestration method based on a vector knowledge base as described in any one of claims 1 to 9.

Images