Task retrieval methods, devices, and electronic equipment
By using a collaborative closed-loop adaptive task decomposition and retrieval reasoning between a central intelligent agent and distributed nodes, and by leveraging pragmatic relational graphs and adaptive multi-dimensional evaluation weights, the problems of inflexible task decomposition, low routing efficiency, and inaccurate evaluation are solved, thereby improving retrieval efficiency and result accuracy.
Patent Information
- Application Number
- CN202611306978.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies lack adaptability in task decomposition, have low retrieval routing efficiency, long distributed inference waiting times, and fixed evaluation weights that cannot adapt to the characteristics of subtasks, resulting in low retrieval efficiency and result bias.
A closed-loop adaptive task decomposition and retrieval reasoning method is adopted, which uses a central intelligent agent to coordinate distributed nodes. By constructing a preset pragmatic relation graph, semantic matching and node routing are performed. Query conditions are dynamically issued and adaptive multi-dimensional evaluation weights are applied. The execution status of subtasks is determined in real time, and highly relevant data nodes are accurately located while avoiding global broadcast communication.
It reduces network load, improves retrieval efficiency and result matching accuracy, accurately locates highly relevant data nodes, avoids invalid communication and waste of computing resources, and dynamically evaluates result quality.
Smart Images

Figure CN123114663A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and more specifically, to a task retrieval method, apparatus, and electronic device. Background Technology
[0002] With the rapid development of the Internet of Data, data resources are characterized by massive volume, heterogeneity, and distributed storage. A user's complex query often requires traversing multiple data sources and integrating various types of data objects to obtain a satisfactory answer.
[0003] Current retrieval methods typically employ a simple "broadcast query - independent retrieval by each node - result aggregation" model. However, this approach suffers from several problems: First, task decomposition lacks adaptability. Once the central agent breaks down the task into subtasks, it doesn't adjust the decomposition process, preventing intermediate retrieval results from being fed back to the decomposition strategy. This means that when the retrieval quality of some subtasks is low, timely correction is impossible. Second, retrieval task routing is inefficient. Most systems broadcast queries to all nodes, resulting in significant unnecessary communication overhead and ignoring the inherent relationships between data. Third, during distributed inference, all nodes must wait for the slowest node to complete its local inference before aggregation, severely slowing down the overall response time. Finally, the weights in the multi-dimensional evaluation system are usually pre-set and cannot adaptively adjust the importance of relevance dimensions based on the characteristics of different subtasks, leading to discrepancies between the evaluation results and the user's actual needs.
[0004] There is currently no effective solution to the above problems. Summary of the Invention
[0005] This invention provides a task retrieval method, apparatus, and electronic device to at least address the technical problem of low efficiency in task retrieval in related technologies.
[0006] According to one aspect of the present invention, a task retrieval method is provided, applied to an intelligent agent, comprising: receiving a query task initiated by a user, and splitting the query task into multiple subtasks, wherein each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector; determining the target data object corresponding to each subtask based on a preset pragmatic relational graph, and determining a candidate data object set based on the target data object, wherein the preset pragmatic relational graph includes: multiple data nodes, each data node representing a data object, and each data object corresponding to a distributed node; determining a target distributed node set for executing the subtask based on the candidate data object set, wherein each target distributed node in the target distributed node set is used to execute the subtask and return multiple retrieval results, the retrieval results including at least: retrieved data objects and object scores of the retrieved data objects; determining whether the subtasks have been completed based on the multiple retrieval results, and, if all subtasks have been completed, determining the target retrieval result corresponding to the query task based on all retrieval results.
[0007] Furthermore, before receiving a query task initiated by a user, the process includes: collecting data objects stored in each enterprise department system and representing each enterprise department system as a distributed node; representing each data object as a data node; for every two data nodes, if there is a relationship between the data objects corresponding to the two data nodes, constructing an edge between the two data nodes; determining the edge weight of each edge; and constructing a preset pragmatic relation graph based on all data nodes, all edges, and the edge weight of each edge.
[0008] Further, the steps for determining the edge weight of each edge include: collecting the number of associations between any two data objects, where the number of associations is divided into: reference frequency, co-occurrence frequency, and derivative frequency; determining the maximum reference frequency, maximum co-occurrence frequency, and maximum derivative frequency from all association counts; determining the time interval at which the latest association relationship is generated between any two data objects; for any two data objects, determining the degree of association between the two data objects based on the reference frequency, co-occurrence frequency, derivative frequency, maximum reference frequency, maximum co-occurrence frequency, maximum derivative frequency, and time interval; and determining the degree of association as the edge weight between the two data nodes indicated by the two data objects.
[0009] Further, the process of splitting the query task into multiple subtasks includes: parsing the query task to obtain multiple query targets and the corresponding data type, data range, and time constraints for each query target; for each query target, constructing a subtask based on the query target, data type, data range, and time constraints, wherein, based on the data type, a preset multi-dimensional evaluation weight vector is configured for the subtask, and the query conditions for the subtask are constructed based on the query target, data type, data range, and time constraints; calculating the complexity of each subtask; and if the complexity exceeds a preset complexity threshold, re-splitting the query task until the complexity of each subtask is less than or equal to the preset complexity threshold, resulting in multiple subtasks.
[0010] Further, the steps for determining the target data object corresponding to each subtask based on the preset pragmatic relation graph include: constructing a data semantic vector for each data object in the preset pragmatic relation graph, and constructing a query semantic vector for each subtask; for each query semantic vector, calculating the matching score between the query semantic vector and each keyword in each data object; for each subtask, traversing the preset pragmatic relation graph and representing the traversed data object as the current data object; calculating the initial matching degree between the subtask and the current data object based on the query semantic vector, the data semantic vector of the current data object, the matching score of each keyword in the current data object, and the number of keywords; determining the data distribution information entropy of the current data object and its operational activity at each time point, and determining the historical average activity of the current data object; calculating the matching score between the subtask and the current data object based on the initial matching degree, the data distribution information entropy, all operational activities, and the historical average activity; and determining the data object indicated by the maximum matching score as the target data object.
[0011] Further, the step of determining a candidate data object set based on the target data object includes: traversing a preset pragmatic relational graph using the target data object, filtering out all data objects whose edge weights are greater than a first preset weight threshold, to obtain an initial candidate data object set; determining whether the number of retrieved content based on the initial candidate data object set meets a preset content quantity; if the number of retrieved content does not meet the preset content quantity, continuing to filter out data objects whose edge weights are greater than a second preset weight threshold, and adding the data objects to the initial candidate data object set, until the number of retrieved content meets the preset content quantity, to obtain a candidate data object set, wherein the first preset weight threshold is greater than the second preset weight threshold.
[0012] Further, the step of determining whether a subtask has been completed based on multiple search results includes: determining the object score returned by the distributed nodes of completed tasks, wherein the object score is calculated by the distributed nodes based on the individual score of the data object in each dimension and a preset multi-dimensional evaluation weight vector; determining the maximum score returned, and determining the historical average score and historical score standard deviation of all distributed nodes with incomplete tasks; determining the global termination score based on the historical average score and historical score standard deviation; stopping the execution of all distributed nodes with incomplete tasks if the maximum score is greater than the global termination score; determining the average deviation based on all returned data objects; determining the task score of the subtask based on the object scores of all returned data objects, the number of all returned data objects, and the average deviation; determining that the subtask has been completed if the task score is greater than a preset task threshold; and reconfiguring the query conditions and preset multi-dimensional evaluation weight vector of the subtask and re-executing the subtask if the task score is less than or equal to the preset task threshold.
[0013] Further, after determining the global termination score, the process includes: determining the average object score of all object scores returned by each currently running distributed node when the maximum score is less than or equal to the global termination score; determining the average of the lowest object scores returned by all distributed nodes that have completed tasks, and determining the standard deviation of all distributed nodes that have completed tasks; determining the abnormal score based on the average and standard deviation; and identifying the distributed node as an abnormal node and stopping its execution when the average object score is less than the abnormal score.
[0014] According to another aspect of the present invention, a task retrieval device is also provided, applied to an intelligent agent, comprising: a splitting unit, configured to receive a query task initiated by a user and split the query task into multiple subtasks, wherein each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector; a first determining unit, configured to determine the target data object corresponding to each subtask based on a preset pragmatic relational graph, and determine a candidate data object set based on the target data object, wherein the preset pragmatic relational graph includes: multiple data nodes, each data node representing a data object, and each data object corresponding to a distributed node; a second determining unit, configured to determine a target distributed node set for executing the subtask based on the candidate data object set, wherein each target distributed node in the target distributed node set is used to execute the subtask and return multiple retrieval results, the retrieval results including at least: retrieved data objects and object ratings of the retrieved data objects; and a judging unit, configured to judge whether the subtask has been completed based on the multiple retrieval results, and, if all subtasks have been completed, determine the target retrieval result corresponding to the query task based on all retrieval results.
[0015] Furthermore, the task retrieval device is also used to: collect data objects stored in each enterprise department system before receiving a query task initiated by a user, and represent each enterprise department system as a distributed node; represent each data object as a data node; for every two data nodes, if there is a relationship between the data objects corresponding to the two data nodes, construct an edge between the two data nodes; determine the edge weight of each edge; and construct a preset pragmatic relation graph based on all data nodes, all edges, and the edge weight of each edge.
[0016] Furthermore, the task retrieval device is also used to: collect the number of associations between any two data objects, wherein the number of associations is divided into: reference frequency, co-occurrence frequency, and derivative frequency; determine the maximum reference frequency, maximum co-occurrence frequency, and maximum derivative frequency from all association counts; determine the time interval at which the latest association relationship is generated between any two data objects; for any two data objects, determine the degree of association between the two data objects based on the reference frequency, co-occurrence frequency, derivative frequency, maximum reference frequency, maximum co-occurrence frequency, maximum derivative frequency, and time interval; and determine the degree of association as the edge weight between the two data nodes indicated by the two data objects.
[0017] Furthermore, the splitting unit is also used to: parse the query task to obtain multiple query targets and the data type, data range, and time constraints corresponding to each query target; for each query target, construct sub-tasks based on the query target, data type, data range, and time constraints, wherein, based on the data type, a preset multi-dimensional evaluation weight vector is configured for the sub-task, and the query conditions of the sub-task are constructed based on the query target, data type, data range, and time constraints; calculate the complexity of each sub-task; if the complexity is greater than a preset complexity threshold, re-split the query task until the complexity of each sub-task obtained by splitting is less than or equal to the preset complexity threshold, thus obtaining multiple sub-tasks.
[0018] Furthermore, the first determining unit is also used to: construct a data semantic vector for each data object in a preset pragmatic relation graph, and construct a query semantic vector for each subtask; for each query semantic vector, calculate the matching score between the query semantic vector and each keyword in each data object; for each subtask, traverse the preset pragmatic relation graph and represent the traversed data object as the current data object; calculate the initial matching degree between the subtask and the current data object based on the query semantic vector, the data semantic vector of the current data object, the matching score of each keyword in the current data object, and the number of keywords; determine the data distribution information entropy of the current data object and its operational activity at each time point, and determine the historical average activity of the current data object; calculate the matching score between the subtask and the current data object based on the initial matching degree, the data distribution information entropy, all operational activities, and the historical average activity; and determine the data object indicated by the maximum matching score as the target data object.
[0019] Furthermore, the first determining unit is also used to: traverse a preset pragmatic relation graph with the target data object, filter out all data objects whose edge weights are greater than a first preset weight threshold, and obtain an initial candidate data object set; determine whether the number of retrieved content based on the initial candidate data object set meets the preset content quantity; if the number of retrieved content does not meet the preset content quantity, continue to filter out data objects whose edge weights are greater than a second preset weight threshold, and add the data objects to the initial candidate data object set, until the number of retrieved content meets the preset content quantity, and obtain a candidate data object set, wherein the first preset weight threshold is greater than the second preset weight threshold.
[0020] Furthermore, the judgment unit is also used to: determine the object score returned by the distributed nodes that have completed the task, wherein the object score is calculated by the distributed nodes based on the individual score of the data object in each dimension and the preset multi-dimensional evaluation weight vector; determine the maximum score returned, and determine the historical average score and historical score standard deviation of all distributed nodes that have not completed the task; determine the global termination score based on the historical average score and historical score standard deviation; stop the execution of all distributed nodes that have not completed the task if the maximum score is greater than the global termination score; determine the average deviation based on all returned data objects; determine the task score of the subtask based on the object scores of all returned data objects, the number of all returned data objects, and the average deviation; determine that the subtask has been completed if the task score is greater than the preset task threshold; and reconfigure the query conditions and preset multi-dimensional evaluation weight vector of the subtask if the task score is less than or equal to the preset task threshold, and re-execute the subtask.
[0021] Furthermore, the task retrieval device is also used to: after determining the global termination score, if the maximum score is less than or equal to the global termination score, determine the average object score of all object scores returned by each currently running distributed node; determine the average of the lowest object scores returned by all distributed nodes that have completed tasks, and determine the standard deviation of all distributed nodes that have completed tasks; determine the abnormal score based on the average and standard deviation; if the average object score is less than the abnormal score, identify the distributed node as an abnormal node and stop the execution of the abnormal node.
[0022] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the task retrieval method described above.
[0023] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described task retrieval methods.
[0024] In this invention, a query task initiated by a user is received, and the query task is split into multiple subtasks. Based on a preset pragmatic relational graph, the target data object corresponding to each subtask is determined. Based on the target data object, a set of candidate data objects is determined. Based on the set of candidate data objects, a set of target distributed nodes for executing the subtask is determined. Based on multiple search results, it is determined whether the subtask has been completed. If all subtasks have been completed, the target search result corresponding to the query task is determined based on all search results. This solves the technical problem of low efficiency in task retrieval in related technologies.
[0025] In this invention, a closed-loop adaptive task decomposition and retrieval reasoning method is adopted, which coordinates a central intelligent agent with distributed nodes. By constructing a preset pragmatic relation graph covering the pragmatic association of data objects, semantic matching and node routing are performed. Subtasks containing query conditions and adaptive multi-dimensional evaluation weight vectors are dynamically issued. Based on the retrieval results and object scores returned asynchronously by distributed nodes, the execution status of subtasks is determined in real time. This achieves the goals of accurately locating highly relevant data nodes, avoiding global broadcast communication overhead, and realizing dynamic evaluation of results. As a result, the technical effects of reducing network load and improving retrieval efficiency and result matching degree are achieved. Attached Figure Description
[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0027] Figure 1 This is a flowchart of an optional task retrieval method according to an embodiment of the present invention;
[0028] Figure 2 This is a schematic diagram of an optional agent task decomposition and distributed enhanced retrieval reasoning process for data networks according to an embodiment of the present invention.
[0029] Figure 3 This is a schematic diagram of an optional intelligent agent task decomposition process for data networks according to an embodiment of the present invention;
[0030] Figure 4 This is a schematic diagram of an optional task decomposition and retrieval feedback closed-loop workflow according to an embodiment of the present invention;
[0031] Figure 5 This is a schematic diagram of an optional distributed retrieval task routing based on a pragmatic relation graph according to an embodiment of the present invention;
[0032] Figure 6 This is a schematic diagram of an optional task-driven multi-dimensional evaluation weight adaptive generation and usage process according to an embodiment of the present invention;
[0033] Figure 7 This is a schematic diagram of an optional task retrieval device according to an embodiment of the present invention;
[0034] Figure 8 This is a hardware structure block diagram of an electronic device (or mobile device) for a task retrieval method according to an embodiment of the present invention. Detailed Implementation
[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0036] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0037] It should be noted that all related information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, and necessary measures have been taken to ensure compliance with public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.
[0038] To address the following issues: (1) the task decomposition strategy lacks closed-loop feedback with the retrieval results, making it impossible to dynamically optimize the decomposition based on retrieval quality; (2) the retrieval routing does not fully utilize the pragmatic relationships between data objects, leading to a waste of communication and computing resources; (3) distributed reasoning has a long waiting time and lacks adaptive early termination methods; (4) the evaluation weights are fixed and do not adapt to the personalized needs at the sub-task level. This invention proposes an intelligent agent task decomposition and distributed enhanced retrieval reasoning method for the Internet of Things. This method relies on the collaborative operation of a central intelligent agent and multiple distributed retrieval reasoning nodes, adapting to the complex query scenarios of massive, heterogeneous, and distributed data in the Internet of Things environment. The overall design revolves around the entire task processing process, fully covering the four major stages of task parsing, distribution routing, retrieval reasoning, and result evaluation.
[0039] This invention abandons the traditional distributed retrieval model of static execution, global broadcasting, fixed weights, and passive waiting. It deeply integrates technologies such as large-scale model semantic understanding, graph structure routing, dynamic evaluation, and closed-loop feedback to construct a fully adaptive distributed retrieval reasoning system to ensure the efficient and accurate completion of complex query tasks.
[0040] The overall operation flow of this invention is as follows: After the user submits a query request, the central intelligent agent completes semantic parsing and subtask decomposition, and completes the routing allocation of retrieval nodes in combination with the pragmatic relation graph; each distributed node performs retrieval, scoring and reasoning in parallel and asynchronously returns the results; the central intelligent agent performs a comprehensive quality judgment on the results, and if the results do not meet the standards, it initiates a re-decomposition closed loop, and if the results meet the standards, it summarizes all valid results and outputs them to the outside world.
[0041] The present invention will now be described in detail with reference to various embodiments.
[0042] Example 1
[0043] According to an embodiment of the present invention, an embodiment of a task retrieval 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.
[0044] Figure 1 This is a flowchart of an optional task retrieval method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:
[0045] Step S101: Receive the query task initiated by the user and split the query task into multiple subtasks. Each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector.
[0046] In this embodiment of the invention, the central intelligent agent receives natural language query instructions submitted by the user, and uses a large language model to perform semantic parsing, intent recognition, and requirement decomposition, mapping complex query logic into subtasks that can be executed independently or sequentially. For example, "find data security reports from the past two years" can be decomposed into the subtask of "retrieve reports that meet the topic constraints." During this process, the complexity of the subtasks can be quantified, and standardized execution parameters can be generated.
[0047] Here, the query conditions refer to the search scope, keywords, time constraints, and data filtering rules determined after semantic extraction; the preset multi-dimensional evaluation weight vector refers to the set of coefficients for each evaluation dimension dynamically generated according to the characteristics of the sub-task (such as timeliness and semantic relevance), which satisfies the normalization constraint.
[0048] For example, in the context of the Internet of Things (IoT), user queries are mostly complex tasks described in natural language, characterized by semantic ambiguity, multi-level requirements, and diverse data types. Traditional manual or fixed-rule decomposition methods are ill-suited to dynamically changing query needs. Therefore, a task decomposition module can be built by leveraging the natural language understanding, semantic parsing, and task decomposition capabilities of a large language model. After receiving the user's original query command, the central intelligent agent first performs word segmentation, semantic extraction, intent recognition, and requirement parsing on the query text, accurately identifying key information such as the query target, required data type, data range, time constraints, and result format. Based on this, the overall complex task is decomposed into several independently executable subtasks that can be run serially or in parallel, according to business logic.
[0049] Each completed subtask is accompanied by standardized execution parameters, primarily including precise search and query conditions, the expected type of digital object to be returned, data filtering rules, and an initial multi-dimensional evaluation weight vector. The parameter format is standardized and can be directly recognized and invoked by each distributed node. Once all subtasks are generated, the central agent distributes the subtasks and their associated parameters to the corresponding distributed nodes according to a pre-defined scheduling logic to perform retrieval and inference tasks.
[0050] Step S102: Based on the preset pragmatic relational graph, determine the target data object corresponding to each subtask, and based on the target data object, determine the candidate data object set. The preset pragmatic relational graph includes: multiple data nodes, each data node represents a data object, and each data object corresponds to a distributed node.
[0051] In this embodiment of the invention, a decentralized graph structure (i.e., a preset pragmatic relation graph) covering the entire data domain is pre-constructed. Nodes represent digital objects such as documents and datasets, and edges represent pragmatic associations and weights such as citations and co-occurrences. A central agent calculates the semantic matching degree between the query statement and the nodes in the graph, locating the "core seed object" (i.e., the target data object) with the highest matching degree. Then, based on this seed object, the preset pragmatic relation graph is traversed to select neighboring nodes whose edge weights are greater than a preset threshold, forming a "candidate data object set". For example, using industry standard documents as the core seed, research reports strongly associated with them are selected as candidate objects.
[0052] Here, the preset pragmatic relation graph is a graph structure that represents the pragmatic associations and weights between data objects; the target data object refers to the core seed object determined through semantic matching; and the "candidate data object set" refers to the set of digital objects that have strong pragmatic associations with the target data object.
[0053] Step S103: Based on the candidate data object set, determine the target distributed node set for executing the sub-task. Each target distributed node in the target distributed node set is used to execute the sub-task and return multiple search results. The search results include at least: the retrieved data object and the object score of the retrieved data object.
[0054] In this embodiment of the invention, the "candidate data object set" is mapped to its physical storage location, and the "target distributed node set" that carries these objects is determined. Subtasks and weight vectors are only issued to nodes in this set. Each node performs local retrieval, weighted scoring based on dynamic weights, and sorting in parallel, and asynchronously returns the Top-K high-quality results. The returned data includes the specific digital objects retrieved and their comprehensive "object score" calculated across various dimensions.
[0055] Here, the target distributed node set refers to the physical storage and computing nodes that carry candidate data objects; the object score refers to the quantitative score obtained by the node after performing multi-dimensional weighted calculation on the search results based on the dynamic weight vector issued.
[0056] For example, the physical storage nodes and computing nodes corresponding to these neighboring digital objects in the candidate data object set are mapped, and subtasks are preferentially routed to these nodes, allowing only highly correlated nodes to participate in the retrieval work. If the number of results returned by the priority nodes in the first round is insufficient or the content coverage is too narrow to meet the requirements of the subtask, the routing range will be gradually expanded in order of decreasing correlation strength, successively connecting the next strongest correlated nodes and then the weakest correlated nodes, until the retrieval results meet the requirements. This routing method makes full use of the natural pragmatic correlation between data, abandons the indiscriminate broadcasting mode, reduces invalid communication and redundant computation from the source, effectively improves retrieval routing efficiency, and reduces the overall operating pressure of the data network.
[0057] Step S104: Based on multiple search results, determine whether the subtask has been completed, and if all subtasks have been completed, determine the target search result corresponding to the query task based on all search results.
[0058] In this embodiment of the invention, the central intelligent agent collects the "object scores" returned by each node in real time and calculates the global quality score of the subtask based on the comprehensive quality formula. If the score is lower than the threshold, a re-decomposition mechanism is triggered to optimize the query conditions and weights before resubmitting the task; if the score meets the threshold, the subtask is considered to have been completed. When all subtasks are completed, the valid results are aggregated, merged, and sorted to generate the final "target retrieval result". In addition, based on the historical score statistical distribution, an early termination instruction can be executed on slow nodes that have not returned high-scoring results.
[0059] Here, the target retrieval results refer to the set of valid answers that are finally output to the user after all subtasks have been processed, merged, and sorted.
[0060] In summary, a closed-loop adaptive task decomposition and retrieval reasoning approach can be adopted, which involves a central intelligent agent coordinating with distributed nodes. By constructing a pre-defined pragmatic relation graph covering the pragmatic associations of data objects, semantic matching and node routing are performed. Subtasks containing query conditions and adaptive multi-dimensional evaluation weight vectors are dynamically distributed. Based on the retrieval results and object scores asynchronously returned by distributed nodes, the execution status of subtasks is determined in real time. This achieves the goals of accurately locating highly relevant data nodes, avoiding global broadcast communication overhead, and realizing dynamic evaluation of results. As a result, the technical effects of reducing network load and improving retrieval efficiency and result matching degree are achieved.
[0061] To improve the accuracy of constructing a preset pragmatic relation graph, in the task retrieval method provided in Embodiment 1 of this application, before receiving a query task initiated by a user, data objects stored in each enterprise department system are collected, and each enterprise department system is represented as a distributed node; each data object is represented as a data node; for every two data nodes, if there is a relationship between the data objects corresponding to the two data nodes, an edge is constructed between the two data nodes; the edge weight of each edge is determined; and a preset pragmatic relation graph is constructed based on all data nodes, all edges, and the edge weight of each edge.
[0062] In this embodiment of the invention, the number of data objects within the data network is enormous and their storage locations are scattered. Traditional retrieval generally adopts a global broadcast mode, pushing query tasks to all nodes. This not only generates massive amounts of invalid network communication and consumes bandwidth resources, but also forces a large number of irrelevant nodes to perform useless calculations, significantly increasing the overall system load. Based on this, this embodiment pre-constructs a pragmatic relation graph of digital objects covering the entire data domain as the core basis for retrieval routing. The pragmatic relation graph is a decentralized graph structure, where each graph node corresponds to a different type of digital object within the data network, including documents, datasets, reports, service interfaces, technical materials, etc. The directed and undirected edges in the graph represent pragmatic relationships between different digital objects, specifically divided into three categories: reference relationships, derivation relationships, and co-occurrence relationships.
[0063] Specifically, all digital objects stored in the systems of various business departments within the enterprise (such as R&D, legal, and marketing departments) are collected through data access interfaces, including research reports, documents, and technical specifications. Each enterprise department system is mapped as an independent distributed node to carry and manage the department's data.
[0064] Each collected data object is uniquely identified and represented as a data node in a pragmatic relation graph, establishing a mapping relationship between the data object and the graph node.
[0065] Analyze the interaction history between the data objects corresponding to any two data nodes. If a relationship such as reference, co-occurrence, or derivation is found between the two data nodes, a directed or undirected edge is constructed between the two data nodes to represent their pragmatic relationship.
[0066] The numerical weight of each edge is calculated based on factors such as association frequency and time decay, quantifying the degree of association between data objects.
[0067] Integrate all data nodes, constructed edges, and corresponding edge weights to form a complete preset pragmatic relation graph.
[0068] Here, "enterprise department system" refers to a computer system or data cluster within an enterprise that undertakes specific business functions and stores relevant data resources; "distributed node" refers to a physical or logical computing unit in a distributed computing environment that is responsible for storing some data objects and performing local retrieval and reasoning tasks, which corresponds to the node in the enterprise department system.
[0069] Data objects refer to various forms of data entities stored in the network, such as documents, datasets, reports, or service interfaces; data nodes refer to graph structure units in a predefined pragmatic relational graph used to represent a single data object.
[0070] Association refers to the semantic or logical connection between data objects, specifically including reference relationship, derivation relationship and co-occurrence relationship; edge refers to the line connecting two data nodes in the predefined pragmatic relationship graph, used to express the above association relationship between the two data objects.
[0071] Edge weight refers to a quantitative value that measures the strength of pragmatic association between two data objects. The value range is usually [0,1]. The larger the value, the stronger the association. Its calculation can take into account citation frequency, co-occurrence frequency and time decay factor.
[0072] In this embodiment, by constructing a preset pragmatic relation graph containing data nodes, edges, and edge weights, the pragmatic association relationships among massive heterogeneous data objects within the data network are made explicit and quantified. This enables the retrieval system to perform intelligent routing by utilizing the natural association structure between data, thereby effectively avoiding the communication overhead caused by global broadcasting and improving the accuracy of retrieval tasks and the overall operating efficiency of the system.
[0073] To improve the accuracy of determining the edge weight of each edge, the task retrieval method provided in Embodiment 1 of this application collects the number of associations between every two data objects. The number of associations is divided into: reference frequency, co-occurrence frequency, and derivative frequency. The maximum reference frequency, maximum co-occurrence frequency, and maximum derivative frequency are determined from all association counts. The time interval at which the latest association relationship is generated between any two data objects is determined. For any two data objects, the degree of association between the two data objects is determined based on the reference frequency, co-occurrence frequency, derivative frequency, maximum reference frequency, maximum co-occurrence frequency, maximum derivative frequency, and time interval. The degree of association is determined as the edge weight between the two data nodes indicated by the two data objects.
[0074] In this embodiment of the invention, the data interaction logs within the data network are traversed to count the number of associations between any two data objects, and the associations are further subdivided into reference frequency, co-occurrence frequency, and derivation frequency, respectively recording direct references, co-occurrences, and derivation generation behaviors between objects.
[0075] A global scan of all data pairs is performed to extract the highest citation frequency, highest co-occurrence frequency, and highest derivative frequency, which are used as normalization benchmarks. Then, for any pair of data objects, the difference between the time point of their most recent association and the current time point is calculated; this is the time interval. Finally, the aforementioned frequencies, the corresponding highest frequency benchmark values, and the time intervals are substituted into a preset association degree calculation model to obtain a quantified association degree value, which is then directly set as the edge weight of the connection between the corresponding data nodes.
[0076] Here, the number of associations refers to the total statistical sum of historical interactions between data objects, specifically including "citation frequency," which is the number of times one object explicitly cites another object in the content; "co-occurrence frequency," which is the number of times two objects appear simultaneously in the same context or document; and "derivation frequency," which is the number of times one object generates or evolves into another object.
[0077] "Maximum Citation Frequency," "Maximum Co-occurrence Frequency," and "Maximum Derivation Frequency" refer to the peak values of various association frequencies statistically analyzed across all data object pairs globally. These are used for normalization calculations to eliminate magnitude deviations caused by differences in the overall activity of different objects. "Time Interval" refers to the difference between the timestamp of the most recent association between any two data objects and the current system time, used to characterize the timeliness or freshness of the association relationship.
[0078] For example, by combining data interaction characteristics and association frequency, a pragmatic relation edge weight calculation model is constructed to quantify the degree of association between two digital objects:
[0079] ;
[0080] in, The pragmatic relation weight between digital object u and digital object v, with a value range of... ; These represent the frequency of references between digital objects, the frequency of co-occurrence, and the number of derived associations. This represents the maximum value of the corresponding global indicator. The time interval between the most recent association between the two data objects; This represents the time decay coefficient. Thus, by employing multi-factor product normalization combined with exponential time decay, linear weighting is abandoned. The comprehensive correlation strength is characterized by the coupled product of correlation indicators, while the exponential function is used to achieve the natural decay of correlation over time.
[0081] In this embodiment, by introducing three fine-grained association frequency statistics—reference, co-occurrence, and derivation—and combining global maximum value normalization and time exponential decay mechanisms, a dynamic and multidimensional pragmatic relation edge weight calculation model is constructed. This model can accurately quantify the association strength between data objects over time, thereby ensuring that the preset pragmatic relation graph can truly reflect the real-time relevance and historical association value between data, and improving the accuracy and timeliness of graph routing-based retrieval.
[0082] To improve the accuracy of task decomposition for query tasks, the task retrieval method provided in Embodiment 1 of this application parses the query task to obtain multiple query targets and the data type, data range, and time constraints corresponding to each query target. For each query target, a subtask is constructed based on the query target, data type, data range, and time constraints. A preset multi-dimensional evaluation weight vector is configured for the subtask based on the data type, and query conditions for the subtask are constructed based on the query target, data type, data range, and time constraints. The complexity of each subtask is calculated. If the complexity exceeds a preset complexity threshold, the query task is re-decomposed until the complexity of each subtask is less than or equal to the preset complexity threshold, resulting in multiple subtasks.
[0083] In this embodiment of the invention, the central intelligent agent receives user query commands, uses a large language model for intent recognition and demand decomposition, extracts multiple independent query targets, and clarifies the corresponding data type, data range, and time constraints for each target. Then, it constructs initial subtasks based on these elements. According to the data type involved in the subtask (e.g., time-sensitive data or static documents), a preset multi-dimensional evaluation weight vector is dynamically configured, and the query target, data type, data range, and time constraints are integrated into the query conditions of the subtask. Next, the complexity of each subtask is calculated, taking into account the query text length, the expected number of objects, and the time constraint coefficient. If the calculated subtask complexity exceeds a preset complexity threshold, the current split is deemed unreasonable, and a re-splitting mechanism is triggered to perform a finer-grained decomposition or logical adjustment of the original query task until the complexity of all generated subtasks is less than or equal to the preset complexity threshold, ultimately outputting a stable and executable set of multiple subtasks.
[0084] For example, to ensure the rationality of task decomposition, a formula for quantifying the complexity of subtasks is introduced to quantitatively control the difficulty and computational load of the decomposed subtasks:
[0085] ;
[0086] in, The complexity is that of a single subtask; The valid length of the query text; The expected number of digital objects to be retrieved; The time constraint coefficients are used. Logarithmic, hyperbolic tangent, and exponential composite operations are used instead of linear weighting, and the coupling effect of each factor on the complexity of the subtask is characterized by a nonlinear mapping relationship.
[0087] In some alternative embodiments, traditional distributed retrieval uses globally uniform and fixed evaluation weights, which cannot distinguish the retrieval focus of different sub-tasks and is difficult to match the diverse and personalized query needs of users, ultimately leading to unreasonable result ranking and the burying of high-quality content. Based on this, this embodiment adopts a task-driven multi-dimensional adaptive evaluation weight method to achieve one weight per task, so that the evaluation system is deeply bound to the needs of sub-tasks.
[0088] A standardized multi-dimensional evaluation index system is predefined, with a total of M evaluation dimensions, and the dimension set is as follows. ( (Different dimensions, respectively), corresponding to the dynamic weight vector. The weights must satisfy normalization and non-negativity constraints:
[0089] ;
[0090] While the central intelligent agent decomposes subtasks using a large model, it simultaneously analyzes the core needs and user preferences of each subtask: identifying explicit preferences such as "latest data" and "authoritative documents" in the query text, and inferring implicit user needs based on the query context and business scenario. Based on the analysis results, the large model independently generates a dedicated weight vector for each subtask. .
[0091] The generated weight vector and query conditions are synchronously distributed to each distributed retrieval node. Nodes perform weighted and time-series correction calculations based on the scoring formula to obtain the final comprehensive score, and then sort the results according to the comprehensive score. Different subtasks can flexibly adjust the weight ratio of each dimension according to their own positioning; for example, timeliness queries increase the weight of freshness, content retrieval increases the weight of semantic relevance, and related data queries increase the weight of pragmatic relation strength. This approach allows the evaluation rules to accurately adapt to the characteristics of each subtask, significantly improving the matching degree of search results and user satisfaction.
[0092] In this embodiment, by introducing a complexity-based dynamic decomposition mechanism, the limitations of traditional static task decomposition are broken, and the adaptability of the task decomposition process is realized. It can automatically adjust the granularity of subtasks according to the semantic complexity of the query, ensuring that each issued subtask satisfies both business logic and system performance boundaries. This effectively avoids retrieval failures or response delays caused by task overload, and improves the execution stability and efficiency of complex query tasks in a distributed environment.
[0093] To improve the accuracy of determining the target data object corresponding to each subtask, the task retrieval method provided in Embodiment 1 of this application constructs a data semantic vector for each data object in a preset pragmatic relation graph and constructs a query semantic vector for each subtask. For each query semantic vector, the matching score between the query semantic vector and each keyword in each data object is calculated. For each subtask, the preset pragmatic relation graph is traversed, and the traversed data objects are represented as the current data object. Based on the query semantic vector, the data semantic vector of the current data object, the matching score of each keyword in the current data object, and the number of keywords, the initial matching degree between the subtask and the current data object is calculated. The data distribution information entropy and the running activity of the current data object at each time point are determined, and the historical average activity of the current data object is determined. Based on the initial matching degree, the data distribution information entropy, all running activities, and the historical average activity, the matching score between the subtask and the current data object is calculated. The data object indicated by the maximum matching score is determined as the target data object.
[0094] In this embodiment of the invention, a data semantic vector is constructed for each data object in a preset pragmatic relational graph, and a query semantic vector is constructed for each subtask. For each subtask's query semantic vector, its local matching score with each keyword in the data object is calculated. Here, the data semantic vector refers to a high-dimensional vector mapped from the content of the data object using natural language processing techniques, used to represent the semantic features of the data object. The query semantic vector refers to a high-dimensional vector mapped from the query conditions of the subtask using natural language processing techniques, used to represent the semantic intent of the subtask.
[0095] During the traversal of the predefined pragmatic relation graph, each traversed object is designated as the current data object. Using the query semantic vector, the current data object's data semantic vector, the local matching score, and the number of keywords, the initial matching degree between the subtask and the current data object is calculated. Based on this, the data distribution information entropy of the current data object, as well as its operational activity and historical average activity at various times, are obtained. The initial matching degree, data distribution information entropy, operational activity, and historical average activity are then substituted into a composite formula to calculate the matching score between the subtask and the current data object. Here, data distribution information entropy refers to an information-theoretic index calculated based on the probability distribution of data types within the current data object, used to quantify the diversity or disorder of the data stored in the object. Operational activity refers to the processing request volume or resource consumption of the current data object at a specific time, reflecting its real-time load status. The historical average activity refers to the average activity level of the current data object over a long period, used to characterize the node's typical load characteristics.
[0096] Compare the matching scores of all traversed objects, and determine the data object with the highest matching score as the target data object, which is the core seed for subsequent routes.
[0097] For example, after the central agent completes the subtask decomposition, it first calculates the matching degree based on the query conditions of the subtask by fusing word vectors and contextual semantic features. Then, it locates the node with the highest matching degree to the query requirement in the pragmatic relation graph and marks it as the core seed object. The semantic matching calculation method is as follows:
[0098] ;
[0099] in, The value represents the overall semantic match between the query statement and the numeric object (i.e., the initial match). ; These are the global semantic vectors for query statements and numeric objects, respectively. The matching score is the score of the k-th local keyword in the query and the numeric object, where K is the total number of keyword groups in the numeric object; This is the global semantic weight coefficient, with values ranging from [value missing]. .
[0100] To further refine the selection of candidate nodes, a comprehensive formula for node matching is constructed by combining semantic matching degree, pragmatic weight, and node activity. This formula integrates information entropy and weighted bias calculations to achieve multi-dimensional joint judgment.
[0101] ;
[0102] in, The comprehensive matching score for physical nodes (i.e., the matching score); Entropy is the information entropy of data distribution within a node, used to characterize data diversity. This represents the probability distribution of data point x within a data node; This represents the activity level of the node at time k. This represents the average historical activity level of the nodes. The fraction and the bias term together correct the matching results, reducing the probability of abnormal nodes being selected.
[0103] In this embodiment, a multi-dimensional node comprehensive matching evaluation model is constructed by integrating global semantic vectors, local keyword matching, data distribution information entropy, and node dynamic activity. This model enables accurate scoring of data objects from two dimensions: semantic relevance and system load. As a result, it can intelligently select the most suitable subtask execution starting point, effectively avoiding the problem of low retrieval efficiency caused by uneven node load or data quality differences, and improving the rationality of distributed retrieval task allocation and overall execution efficiency.
[0104] To improve the accuracy of determining the candidate data object set, in the task retrieval method provided in Embodiment 1 of this application, the target data object is used to traverse a preset pragmatic relation graph to filter out all data objects whose edge weights are greater than a first preset weight threshold, thus obtaining an initial candidate data object set; it is then determined whether the number of retrieved content based on the initial candidate data object set meets a preset content quantity; if the number of retrieved content does not meet the preset content quantity, data objects whose edge weights are greater than a second preset weight threshold are further filtered out and added to the initial candidate data object set, until the number of retrieved content meets the preset content quantity, thus obtaining a candidate data object set, wherein the first preset weight threshold is greater than the second preset weight threshold.
[0105] In this embodiment of the invention, starting with a defined target data object, a graph traversal is performed in a preset pragmatic relational graph to filter out all data objects connected to the target data object with edge weights greater than a first preset weight threshold. These highly correlated objects are then grouped into an initial candidate data object set. Next, a retrieval request is initiated to the corresponding distributed nodes based on this initial set, and the number of returned retrieval contents is counted. If the count shows that the number of retrieval contents does not reach the preset content count, the initial set is deemed insufficiently covered. The filtering criteria are then lowered, and data objects with edge weights greater than a second preset weight threshold (less than the first preset weight threshold) are included in the filtering scope and added to the initial candidate data object set, forming an expanded set. This process is iterated until the number of returned retrieval contents meets the preset content count requirement. The final determined set is the candidate data object set.
[0106] For example, setting a route filtering threshold Only when When a strong pragmatic relationship is identified between two objects, they are included in the priority routing scope. Specifically, the pragmatic relationship graph can be traversed around the core seed object to filter out all edge weights greater than a preset threshold. The neighboring nodes are the set of digital objects with strong pragmatic associations to the core seed object. These neighboring digital objects are further mapped to their corresponding physical storage and computing nodes, and subtasks are preferentially routed to these nodes, allowing only highly associated nodes to participate in the retrieval process. If the number of results returned by the first-round priority nodes is insufficient or the content coverage is too narrow to meet the requirements of the subtask, the routing range can be gradually expanded in descending order of association strength, successively connecting the next strongest associated nodes and then the weakest associated nodes, until the retrieval results meet the requirements. This routing method fully utilizes the natural pragmatic associations between data, abandons the indiscriminate broadcasting mode, reduces invalid communication and redundant computation from the source, effectively improves retrieval routing efficiency, and reduces the overall operational pressure on the data network.
[0107] In this embodiment, an adaptive adjustment of the retrieval route range is achieved by setting a hierarchical filtering and dynamic expansion mechanism based on edge weight thresholds. First, precise retrieval is performed within the smallest possible range using strong pragmatic associations. Only when results are insufficient is the range gradually expanded to weakly related nodes. This ensures high relevance of retrieval results while effectively avoiding data loss due to an excessively narrow route range, balancing retrieval efficiency and result coverage, and improving task completion capabilities in complex query scenarios.
[0108] To improve the accuracy of determining whether a subtask has been completed, the task retrieval method provided in Embodiment 1 of this application determines the object score returned by the distributed node for the completed task. The object score is calculated by the distributed node based on the individual score of the data object in each dimension and a preset multi-dimensional evaluation weight vector. The method then determines the maximum returned score and the historical average score and historical score standard deviation of all distributed nodes with incomplete tasks. Based on the historical average score and historical score standard deviation, a global termination score is determined. If the maximum score is greater than the global termination score, the execution of all distributed nodes with incomplete tasks is stopped. Based on all returned data objects, the average deviation is determined. Based on the object scores of all returned data objects, the number of returned data objects, and the average deviation, the task score of the subtask is determined. If the task score is greater than a preset task threshold, the subtask is determined to be completed. If the task score is less than or equal to the preset task threshold, the query conditions and preset multi-dimensional evaluation weight vector of the subtask are reconfigured, and the subtask is re-executed.
[0109] In this embodiment of the invention, distributed retrieval inference is completed collaboratively by multiple heterogeneous nodes. These nodes exhibit significant differences in hardware performance, data volume, and computational load. Traditional methods require waiting for all nodes to complete their calculations before aggregating the results, with the slowest node directly determining the overall task time, easily leading to a waste of computing power and time resources. Therefore, this embodiment proposes an early termination mechanism for distributed inference based on multi-dimensional evaluation results to achieve asynchronous node execution, dynamic judgment, and flexible termination.
[0110] After receiving the subtask, all distributed nodes independently and in parallel perform local retrieval, feature matching, multi-dimensional scoring, and result ranking. Nodes do not wait for or interfere with each other. Once any node completes its local processing, it immediately reports the complete score distribution of the Top-K objects it has selected to the central agent without waiting for other nodes. The reported content includes the individual scores of each digital object in each evaluation dimension, as well as the weighted overall score.
[0111] Specifically, the system collects object scores reported by distributed nodes that have completed their tasks in real time. These scores are calculated by each node by weighting the individual scores of the retrieved data objects in each dimension based on the preset multi-dimensional evaluation weight vector.
[0112] For example, when each node scores local candidate digital objects, it calculates a comprehensive score by combining dimensional scores, weight coefficients, and time-series decay factors. The scoring formula is as follows:
[0113] ;
[0114] in, The final overall score (object score) for a single digital object; For the j-th evaluation dimension, use the adaptive weights. Give the individual score of the digital object in the j-th dimension; The interval between the creation time of the digital object and the current query time; This is the time-series decay coefficient, used to weaken the weight of scores for older data.
[0115] The highest object score returned by the completed node is extracted as the maximum score. The historical average score and standard deviation of all distributed nodes with incomplete tasks are retrieved from the historical database. These scores are then combined with confidence interval coefficients to calculate the global termination score. When the maximum score exceeds the global termination score, it is determined that the current result is significantly better than the expected output of the remaining nodes in terms of statistical confidence. A stop command is then issued to all incomplete nodes, terminating their execution.
[0116] For example, the central intelligent agent collects and summarizes the score data reported by each node in real time, and combines it with historical operating data, score intervals, and confidence intervals to complete a comprehensive judgment. The inference termination judgment model is as follows:
[0117] ;
[0118] in, The highest overall score among the results returned by the completed nodes; This represents the average historical score of incomplete nodes. The standard deviation of the historical scores of incomplete nodes; This is the confidence interval coefficient, used to control the stringency of the judgment.
[0119] When the global termination condition is met, it means that the best result obtained so far is significantly better than the result that the remaining nodes can output within the statistical confidence interval, and the overall confidence level fully meets the task requirements. Then, a global termination command is issued.
[0120] In this embodiment of the invention, based on all returned data objects, the average semantic deviation is calculated, and combined with the object score, the number of data objects, and the average deviation, the task score of the subtask is calculated using a comprehensive quality formula. If the task score is higher than a preset task threshold, the subtask is determined to be completed; if it is lower than or equal to the threshold, a re-decomposition mechanism is triggered, and the central agent optimizes the query conditions and weight vector before re-issuing the subtask.
[0121] For example, after each distributed node completes local retrieval, sorting, and preliminary inference, it uniformly sends the selected Top-K high-quality digital object results back to the central agent. The central agent retrieves a preset multi-dimensional evaluation index system and, combined with result dispersion and matching deviation, comprehensively calculates the overall quality of the subtask to quantify the subtask execution effect. The comprehensive quality calculation formula is as follows:
[0122] ;
[0123] Where Q is the global overall quality score (task score) of the current subtask, and its value range is... N represents the total number of numeric objects returned under this subtask. For the i-th digital object, a single comprehensive score (object score); The arithmetic mean of the scores for all digital objects; The overall semantic deviation coefficient represents the average degree of deviation between the returned content and the query topic, and its value ranges from [value missing]. ; The component weights are coefficients that satisfy the constraint relationships. ,and .
[0124] Preset fixed quality threshold When the calculated overall quality score When the current search results fail to meet the user's query needs, a re-decomposition mechanism is automatically triggered. The central agent combines information such as the distribution characteristics, data deviation direction, and semantic matching degree of the returned results in this round to optimize the sub-task configuration in a targeted manner: on the one hand, it adjusts the query conditions, such as expanding the keyword search range, supplementing synonym search terms, relaxing or tightening data filtering rules, and correcting semantic similarity matching parameters; on the other hand, it reallocates the multi-dimensional evaluation weight vector to adapt to the retrieval focus of the current sub-task. After the parameter optimization is completed, the updated sub-task is redeployed to the distributed nodes to re-execute the retrieval and inference process. This iterative process continues until the sub-task quality score reaches the preset standard, or automatically terminates after reaching the maximum number of retries set by the system. This closed-loop mechanism allows the task decomposition strategy to continuously learn and optimize from the search results, solving the problems of traditional static task decomposition not being able to dynamically adjust and sub-task execution failures not being able to be remedied, and comprehensively improving the reliability of complex task decomposition and execution.
[0125] In some optional embodiments, to achieve dynamic iterative optimization of weights, a weight update and iteration formula is constructed, and the weight configuration is continuously optimized based on historical task data:
[0126] ;
[0127] in, Let be the weight of the j-th dimension in the t-th round of the task; The weights after iteration; To learn step size; This is the weight correction amount calculated based on historical errors.
[0128] In this embodiment, by introducing an early termination mechanism based on statistical confidence and a task feedback loop based on multi-dimensional quality indicators, dynamic scheduling of distributed retrieval resources and adaptive control of task quality are achieved. This not only effectively avoids waiting delays caused by differences in node performance and shortens the overall inference time, but also solves the problem that static decomposition cannot handle retrieval failures through a result quality-driven task re-decomposition mechanism. Thus, while ensuring high accuracy of retrieval results, it improves the execution efficiency and robustness of complex query tasks in the Internet of Things environment.
[0129] To achieve efficient task retrieval, the task retrieval method provided in Embodiment 1 of this application includes the following steps after determining the global termination score: if the maximum score is less than or equal to the global termination score, determining the average object score of all object scores returned by each currently running distributed node; determining the average of the lowest object scores returned by all distributed nodes that have completed tasks, and determining the standard deviation of all distributed nodes that have completed tasks; determining abnormal scores based on the average and standard deviation; and identifying the distributed node as an abnormal node and stopping its execution if the average object score is less than the abnormal score.
[0130] In this embodiment of the invention, all distributed nodes currently executing tasks are monitored in real time. All object scores reported by each node are collected, and the arithmetic mean of the object scores returned by each node is calculated, i.e., the average object score. Furthermore, the lowest object scores returned by all distributed nodes that have completed tasks are obtained, the average of these lowest scores is calculated, and the standard deviation of all object scores for nodes that have completed tasks is determined. Then, using this average and standard deviation, combined with a preset confidence interval coefficient, an anomaly score threshold for determining whether a node's performance is abnormal is calculated. When the average object score of a currently running node is lower than this anomaly score, the node is determined to be an abnormal node, and a stop command is issued to it, terminating its subsequent execution process and releasing computing resources.
[0131] For example, the central intelligent agent collects and summarizes the score data reported by each node in real time, and combines it with historical operating data, score intervals, and confidence intervals to complete a comprehensive judgment. The inference termination judgment model is as follows:
[0132] ;
[0133] in, The average score of the results already output by the currently running node; The average of the lowest scores of the completed nodes. The standard deviation of the scores of the completed nodes; This is the confidence interval coefficient, used to control the stringency of the judgment.
[0134] Introducing a formula for estimating the overall time consumption of a task, and dynamically evaluating the efficiency improvement brought about by the termination strategy:
[0135] ;
[0136] in, The total time taken to complete the task; Time taken to complete the first node; The computation time for the remaining nodes; This is a correction factor for early termination efficiency, with a range of values. .
[0137] When the termination criteria for a single abnormal node are met, it indicates that the overall quality of the node's output is low and deviates too much from the valid result range. Subsequent computational processes such as depth sorting and fine-grained matching for that node are terminated prematurely. Based on the judgment result, the central agent issues a termination command to the corresponding node, which immediately stops all computational tasks. This mechanism accurately avoids the time loss caused by node waiting, shortens the overall latency of distributed inference, and is particularly suitable for heterogeneous data networking environments with large differences in node performance and uneven data distribution.
[0138] In this embodiment, by establishing an anomaly score judgment model based on the statistical characteristics of completed task nodes, real-time monitoring and anomaly removal of individual nodes in operation are realized. It can accurately identify and interrupt abnormal nodes that output low-quality results, effectively prevent inferior data from interfering with the final result summary, save computing resources, and improve the fault tolerance and overall execution efficiency of the distributed inference system in heterogeneous environments.
[0139] The following describes in detail another optional implementation method.
[0140] In this embodiment of the invention, a method for agent task decomposition and distributed enhanced retrieval reasoning for the Internet of Things is proposed. Figure 2 This is a schematic diagram of an optional agent task decomposition and distributed enhanced retrieval reasoning process for data networks according to an embodiment of the present invention, such as... Figure 2As shown, after a user submits a query request, the central agent performs semantic parsing and task decomposition, and completes node routing based on the pragmatic relation graph. Each distributed retrieval node executes retrieval inference in parallel, and the nodes asynchronously report Top-K results and scores. The central agent performs a comprehensive quality assessment of the results. If the results do not meet the standards, a re-decomposition closed loop is initiated, i.e., optimizing the query conditions / weights, re-decomposing the task, and re-completing node routing based on the pragmatic relation graph; if the results meet the standards, the results are aggregated to generate the final output.
[0141] This paper details the application scenario of a large enterprise's internal intelligent question-and-answer and data retrieval system in a data network environment. Assume the enterprise has built a comprehensive data network platform that integrates digital objects such as business documents, research reports, industry standards, technical papers, materials, and operational logs from various departments and branches. All data is distributed and stored on physical nodes corresponding to different departments, with varying node hardware configurations, data storage volumes, and operational loads—a typical heterogeneous distributed data network architecture.
[0142] The simulated user query is: "Find the most cited research reports related to 'data security' in the past two years and summarize their core viewpoints." This query belongs to a multi-level complex natural language task, which includes multiple requirements such as data filtering, indicator statistics, result sorting, and content summarization, making it suitable for verifying the complete process of this embodiment.
[0143] The central intelligent agent uses a mainstream large language model as its core semantic parsing and reasoning engine, possessing powerful natural language understanding, task decomposition, weight generation, and logical judgment capabilities. Ten distributed retrieval and reasoning nodes are deployed, corresponding to different departments such as the enterprise's R&D department, operations and maintenance department, legal department, marketing department, and research institute. Each node independently stores various digital objects generated and archived by its department, including research reports, academic papers, documents, industry standards, and technical documents. Nodes interact and transmit instructions with each other through a data network communication protocol.
[0144] (1) Implementation of task decomposition and retrieval feedback closed loop.
[0145] Users input their natural language queries into the central AI agent, initiating the entire task processing. The central AI agent first performs comprehensive semantic analysis, intent recognition, and requirement breakdown on the user's query text, extracting the constraints and execution objectives one by one: First, the data topic constraint is "data security"; second, the time range constraint is "the last two years"; third, the sorting rule constraint is "most cited"; fourth, the final output requirement is "a filtered report + a summary of core viewpoints". Combining business execution logic with the operational characteristics of data network nodes, the central intelligent agent breaks down this complex overall task into three logically progressive and sequentially executed independent subtasks. These three subtasks have clear divisions of labor and are executed in sequence, as follows: Subtask A: Retrieve all research reports from the past two years whose titles, keywords, or abstracts contain the semantics of "data security," completing preliminary data screening; Subtask B: For the candidate report set output by Subtask A, count the total number of citations for each report across the entire network and sort them from highest to lowest citation count; Subtask C: Select the top five research reports by citation count, perform content analysis and key information extraction on the document text, summarize the core viewpoints of each report, and finally generate standardized abstract content.
[0146] After task decomposition, the central agent generates corresponding query conditions and initial multi-dimensional evaluation weight vectors based on the different positioning of the three subtasks. Considering the execution logic, subtask A and subtask B have a strong dependency relationship; therefore, the two subtasks are merged into a joint retrieval task and uniformly distributed to each distributed node for execution. For this joint task, the system sets the initial weight vector as follows: semantic relevance weight 0.6, pragmatic relation strength weight 0.2, metadata matching degree weight 0.2, and strong weighting of the time dimension is not yet enabled.
[0147] The central intelligent agent distributes the joint retrieval task, query keywords, time constraints, and initial weight vector to 10 distributed nodes within the enterprise. All nodes simultaneously initiate local retrieval operations. Each node traverses all digital objects stored locally, filters research reports that meet the query criteria of "recent two years + data security theme," extracts basic information such as report basics and citation counts, compiles a local Top-10 results set, and uniformly transmits it back to the central intelligent agent.
[0148] Figure 3 This is a schematic diagram of an optional data network-oriented intelligent agent task decomposition process according to an embodiment of the present invention, such as... Figure 3 As shown, the user inputs a complex natural language query, the central large model intelligent agent performs semantic parsing, breaks down the query into multiple sets of serial / parallel subtasks, generates query conditions and initial evaluation weight vectors for each subtask, and distributes the subtasks to distributed nodes of the data network.
[0149] After the central agent collects the results from all nodes, it initiates a multi-dimensional comprehensive evaluation process. Each returned report is scored according to preset scoring rules, and the overall quality score of the current joint subtask is calculated based on the overall results. In this case, one department node returned reports that largely deviated from the core theme of "data security" due to incomplete local document tags and a narrow keyword matching range. The semantic relevance scores of these reports were generally below 0.3, dragging down the overall task quality score. The final comprehensive quality score for this subtask was 0.45. The system's preset task quality threshold is 0.6. Since the current score is significantly below the standard, the system automatically determines that the first round of search results is substandard and immediately triggers a re-decomposition and parameter optimization mechanism.
[0150] The central intelligent agent made targeted adjustments based on the issues exposed in the first round of results: First, it optimized the query conditions by semantically expanding the original single keyword "data security" to "data security OR information security OR data privacy," thereby broadening the scope of search matching and avoiding missing valid documents due to keyword limitations; Second, it readjusted the multi-dimensional evaluation weight vector, increasing the semantic relevance weight to 0.7 to further strengthen the importance of topic matching, while lowering the pragmatic relation strength weight to 0.1, and keeping the metadata matching degree unchanged at 0.2.
[0151] After parameter optimization, the central agent reissues the updated query conditions and new weight vectors to the corresponding distributed nodes, initiating the second round of retrieval inference. Each node re-filters, sorts, and returns the results to its local reports according to the new rules. After the second round of results is summarized, the central agent recalculates the task quality score. This time, the score improves to 0.78, exceeding the preset threshold of 0.6, indicating that the retrieval results are qualified, and the task proceeds normally to the next stage.
[0152] Figure 4 This is a schematic diagram of an optional task decomposition and retrieval feedback closed-loop workflow according to an embodiment of the present invention, such as... Figure 4 As shown, the central agent issues sub-tasks, distributed nodes execute retrieval inference, and the nodes return Top-K numerical object results. The central agent calculates the quality score of the sub-task from multiple dimensions and determines whether the score is greater than or equal to a preset threshold. If it is, the results are summarized and proceed to the next process; otherwise, a re-decomposition mechanism is triggered to adjust the query conditions / weight vector.
[0153] The above complete process constitutes a closed-loop feedback system. In actual engineering operation, the system will also be configured with a maximum retry limit (e.g., setting a maximum of 3 retries). If the task score still fails to meet the target after multiple consecutive re-decompositions, the system will automatically terminate the task and push an anomaly alarm to the administrator, indicating that there are problems with the data source, retrieval rules, or node status, thus achieving proactive anomaly detection. This closed-loop mode completely changes the static defects of the traditional task "one-time decomposition, lifelong execution," enabling the entire retrieval process to have self-optimization and self-correction capabilities, significantly improving the success rate of complex query tasks.
[0154] (2) Implementation of retrieval task routing based on pragmatic relation graph.
[0155] Prior to the official launch of the enterprise's digital network platform, technical personnel completed a comprehensive inventory of all digital objects and constructed a complete pragmatic relationship graph. This graph covers all digital objects stored in all departmental nodes within the enterprise, including reports, papers, and standards, and is continuously and dynamically updated to ensure the real-time validity of object, relationship, and weight information. The pragmatic relationship graph is strictly constructed according to three types of pragmatic relationships: citation, derivation, and co-occurrence. Each related edge is assigned a weight value between 0 and 1 based on the degree of association; the closer the weight is to 1, the stronger the association between the two digital objects.
[0156] To illustrate, consider a typical object in the diagram: Research report R1 is a data security analysis report developed by the company itself. The document directly references several normative clauses from industry standard document S2. Therefore, a directed reference edge from R1 to S2 is constructed in the pragmatic relation graph, with a weight of 0.9 based on the length and core nature of the reference. Simultaneously, research report R1 and academic paper P3 frequently appear simultaneously in company technical seminars and comprehensive document compilations, indicating a high-frequency co-occurrence relationship. Therefore, an undirected co-occurrence edge is constructed between R1 and P3, with a weight of 0.6. Similar relationships are abundant throughout the pragmatic relation graph, forming a complete object association network.
[0157] After the central agent completes the breakdown and parameter configuration of subtask A, the retrieval routing process is officially initiated. The first step, based on the query condition of the subtask "retrieve research reports related to data security in the past two years," employs a dual approach of keyword matching and semantic vector similarity calculation to traverse all nodes in the pragmatic relationship graph, selecting the object with the highest matching degree as the core seed object. In this retrieval, industry standard document S2, a fundamental data security specification recognized internally by the enterprise, has a semantic vector similarity score as high as 0.95 and is therefore identified as the core seed object for this routing.
[0158] The second step involves traversing all directly connected neighbor nodes in the pragmatic relation graph, centered on the core seed object S2, and extracting the weight values of the corresponding associated edges. The system presets a strong association threshold of 0.5, and all neighbor nodes with edge weights greater than 0.5 are considered strong pragmatic relation nodes. After screening, R1 (reference edge, weight 0.9) and P3 (co-occurrence edge, weight 0.6), which are directly associated with S2, are both considered strong relation nodes.
[0159] The third step is to complete the mapping between logical objects and physical nodes: digital object R1 is stored in the physical node Node1 corresponding to the R&D department, and digital object P3 is stored in the physical node Node2 corresponding to the research institute. The central agent prioritizes sending the current subtask to the two highly correlated nodes Node1 and Node2 according to the routing rules, instead of broadcasting the task to all 10 nodes in the traditional mode, thus reducing invalid communication from the source.
[0160] The first priority node begins executing the retrieval task. The system monitors the quantity and quality of the results returned by the two nodes in real time. If Node1 and Node2 return a sufficient number of research reports with comprehensive content coverage, fully meeting the requirements of the subtask, the routing process ends and no other nodes are connected. If the number of results returned by the first-round nodes is insufficient or the topic coverage is too narrow to support subsequent sorting and abstracting work, the system will gradually expand the routing scope according to the association weight from high to low, sequentially connecting secondary nodes and weakly associated nodes that are indirectly related to R1 and P3, until the retrieval results meet the requirements.
[0161] Figure 5 This is a schematic diagram of an optional distributed retrieval task routing based on a pragmatic relation graph according to an embodiment of the present invention, such as... Figure 5 As shown, a pre-constructed pragmatic relation graph of digital objects (reference / derivation / co-occurrence relations + weights) is used to perform sub-task query condition matching, locate the core seed object, filter strong pragmatic relation neighbor objects, prioritize routing to the physical node where the neighbor object is located, and determine whether the search results are sufficient. If so, the node executes the search task; otherwise, it gradually expands to the second strongest relation nodes.
[0162] In this case, Node1 and Node2 returned multiple high-quality data security research reports, providing sufficient results and thus eliminating the need to expand the routing scope. This routing method fully leverages the inherent pragmatic relevance of data to accurately locate high-value nodes. Compared to global broadcast mode, it significantly reduces network bandwidth consumption and computational pressure on idle nodes. Furthermore, because node data is highly relevant to query requirements, the overall quality of search results is significantly improved. In the daily operation of enterprise data networks, facing tens of thousands of search requests daily, the resource optimization effect of this routing mechanism is particularly outstanding.
[0163] (3) Early termination of distributed reasoning driven by multi-dimensional evaluation results.
[0164] This section focuses on the implementation of the merged subtask B. The core task of subtask B is to sort the selected research reports according to their citation count, which is a typical distributed statistics and sorting task. Based on the routing results, the central agent distributes this subtask to four distributed nodes: Node1, Node2, Node3, and Node4. These four nodes perform local data statistics, citation count calculation, result sorting, and scoring in parallel. The hardware performance of the four nodes varies: Node1 and Node2 have high-performance hardware and small amounts of local data, resulting in fast processing speeds; Node3 and Node4 have large data storage requirements and generally lower hardware performance, requiring approximately 1.8 seconds to complete a full deep sort.
[0165] Upon receiving the task, all nodes immediately and independently begin local computation, operating independently without data interaction or waiting. Once any node completes the retrieval, scoring, and sorting of its local Top-5 objects, it immediately reports the complete score distribution data to the central agent without waiting for other nodes. The reported content includes the semantic score, pragmatic score, metadata score, and weighted comprehensive score for each report.
[0166] The task execution sequence is as follows: From 0 to 0.5 seconds, Node1 completes all calculations first, compiling the comprehensive score distribution of the local Top-5 reports: [0.92, 0.88, 0.85, 0.80, 0.75], and reports it to the central agent in real time; From 0 to 0.7 seconds, Node2 completes its local calculations and reports the Top-5 score distribution: [0.90, 0.86, 0.82, 0.78, 0.72]. At this time, Node3 and Node4 are still performing depth sorting calculations and have not yet completed their work.
[0167] After collecting the two sets of reported data in real time, the central agent immediately initiates the early termination judgment logic. First, it extracts the highest comprehensive score of the completed nodes: Node1's highest score is 0.92, and Node2's highest score is 0.90, both significantly higher than 0.9. Then, it retrieves the system's historical operation database to query the historical score records of Node3 and Node4 in similar data security retrieval tasks, finding that the historical average score ceiling for these two nodes is only around 0.6. The overall judgment is: the optimal value obtained so far far exceeds the upper limit of the results that the remaining two nodes can produce; the existing data is sufficient to support subsequent sorting, statistics, and summarization work; the overall result confidence meets the task requirements, and there is no need for the remaining nodes to continue computation.
[0168] Based on this judgment, the central agent sends an early termination command to Node3 and Node4, the two still-running nodes, at 0.9 seconds. Upon receiving the command, the two nodes immediately stop all subsequent computation processes, including depth sorting and fine matching, and release their local computing resources. Previously, if the traditional method were followed to wait for the slowest node to complete its computation, the overall task would take 1.8 seconds. With the early termination mechanism of this invention, the overall inference time is reduced to 0.9 seconds, a reduction of half the time.
[0169] (4) Implementation of a task-driven, multi-dimensional evaluation weight adaptive method.
[0170] Multi-dimensional adaptive weighting is a core element in ensuring that search results accurately match user needs. Differentiated dynamic weight vectors are configured for the different requirements of the three sub-tasks, intuitively demonstrating the application logic of adaptive weighting. The system's preset unified evaluation dimensions include four dimensions: semantic relevance, pragmatic relation strength, metadata matching degree, and time freshness. The scoring criteria for all nodes are completely consistent.
[0171] First, configuration is performed for subtask A: The user query explicitly states a "last two years" time constraint, indicating that data timeliness is a crucial requirement, but it does not specify the report author, publishing organization, file format, etc. After the large model analyzes this requirement, it generates a weight vector specific to this subtask: Semantic Relevance = 0.5, Pragmatic Relationship Strength = 0.2, Metadata Matching Degree = 0.1, Time Freshness = 0.2. This weight combination takes into account topic matching, data timeliness, and object relevance, adapting to the business objective of initially screening reports.
[0172] After receiving the weight vector, each distributed node scores each local candidate report across different dimensions, then calculates the overall score by summing the scores according to the dynamic weights. A typical research report is selected as an example: this report has a semantic score of 0.8, a pragmatic relation score of 0.6, a metadata score of 0.7, and a time freshness score of 0.9. The overall score is calculated according to the weights of subtask A:
[0173] ;
[0174] Next, configuration is performed for subtask C: Subtask C aims to extract the core viewpoints of highly cited reports and generate a content summary. This step requires extremely high semantic understanding and content accuracy; time, metadata, and other dimensions are no longer the primary considerations. The large model generates a new weight vector based on the task characteristics: semantic relevance = 0.7, pragmatic relation strength = 0.2, metadata matching degree = 0.05, and time freshness = 0.05. The same report is then substituted into the weight rules of subtask C to calculate the score.
[0175] ;
[0176] The scores calculated in the two trials showed a significant difference, fully reflecting the different evaluation focuses of the various subtasks. The weight vector was distributed to all nodes synchronously along with the subtasks, and the nodes used dynamic weights to complete the scoring and sorting throughout the process, no longer using fixed global weights.
[0177] From user query input, preference parsing, weight generation, execution, to result output, a complete weight-adaptive workflow is formed. Throughout the entire process, all evaluation dimension names, calculation rules, and scoring standards remain consistent, with only the weight values dynamically adjusted according to sub-tasks. This ensures both the standardization of the scoring system and the personalization of evaluation rules. In practical applications of enterprise data networks, this method can flexibly adapt to various differentiated needs such as time-based queries, content-based queries, relational queries, and source-based queries, ensuring that search results accurately match the user's true intent.
[0178] Figure 6 This is a schematic diagram of an optional task-driven, multi-dimensional evaluation weight adaptive generation and usage process according to an embodiment of the present invention, such as... Figure 6 As shown, based on the user query, the large model agent analyzes the task preference, generates a dynamic weight vector specific to the subtask, and synchronously sends the weight vector and query conditions to the nodes. The nodes extract the original scores of each dimension of the candidate digital objects, calculate the comprehensive score by weighting and summing according to the dynamic weights, sort the results according to the comprehensive scores, and output the Top-K results.
[0179] The implementation process of the above four stages demonstrates that the deep integration of four key technologies—task decomposition closed-loop, pragmatic relation routing, early inference termination, and adaptive weighting—fully covers the entire lifecycle of complex queries in the Internet of Things. The entire methodology is simple to deploy and logically clear, facilitating easy setup, configuration, and deployment. Furthermore, it outperforms traditional distributed retrieval solutions in terms of retrieval accuracy, operational efficiency, resource utilization, and user experience, possessing significant engineering and industry application value.
[0180] In this embodiment of the invention, by introducing quality feedback for search results, the central agent can dynamically adjust the query conditions and weights of subtasks, improving the accuracy of complex query decomposition and avoiding partial subtask failures caused by one-time decomposition. Utilizing pragmatic relational graphs to guide routing reduces broadcast communication volume, lowers network load, and improves the relevance of search results. Compared to traditional broadcast methods, communication overhead is reduced, and search latency is further lowered. Dynamically determining whether to prematurely terminate slow nodes based on the quality of returned results significantly shortens the overall waiting time for distributed inference. Each subtask has an independent weight vector, enabling flexible adaptation to different information needs and improving the relevance of result ranking and user satisfaction.
[0181] The following is a detailed description with reference to another embodiment.
[0182] Example 2
[0183] The task retrieval device provided in this embodiment includes multiple implementation units, each of which corresponds to a specific implementation step in Embodiment 1 above.
[0184] Figure 7 This is a schematic diagram of an optional task retrieval device according to an embodiment of the present invention, such as... Figure 7 As shown, the task retrieval device may include: a splitting unit 70, a first determining unit 71, a second determining unit 72, and a judging unit 73.
[0185] The splitting unit 70 is used to receive the query task initiated by the user and split the query task into multiple subtasks. Each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector.
[0186] The first determining unit 71 is used to determine the target data object corresponding to each subtask based on a preset pragmatic relational graph, and to determine a set of candidate data objects based on the target data objects. The preset pragmatic relational graph includes: multiple data nodes, each data node representing a data object, and each data object corresponding to a distributed node.
[0187] The second determining unit 72 is used to determine the target distributed node set for executing sub-tasks based on the candidate data object set, wherein each target distributed node in the target distributed node set is used to execute sub-tasks and return multiple search results, and the search results include at least: the retrieved data objects and the object scores of the retrieved data objects;
[0188] The judgment unit 73 is used to determine whether a subtask has been completed based on multiple search results, and, if all subtasks have been completed, to determine the target search result corresponding to the query task based on all search results.
[0189] The aforementioned task retrieval device employs a closed-loop adaptive task decomposition and retrieval reasoning approach, where a central intelligent agent collaborates with distributed nodes. By constructing a pre-defined pragmatic relational graph encompassing the pragmatic associations of data objects, it performs semantic matching and node routing. Subtasks containing query conditions and adaptive multi-dimensional evaluation weight vectors are dynamically distributed. Based on the retrieval results and object scores asynchronously returned by distributed nodes, the device determines the execution status of subtasks in real time. This achieves the goals of accurately locating highly relevant data nodes, avoiding global broadcast communication overhead, and realizing dynamic evaluation of results. Consequently, it reduces network load and improves retrieval efficiency and result matching accuracy.
[0190] Optionally, the task retrieval device is further configured to: collect data objects stored in each enterprise department system before receiving a query task initiated by a user, and represent each enterprise department system as a distributed node; represent each data object as a data node; for every two data nodes, if there is a relationship between the data objects corresponding to the two data nodes, construct an edge between the two data nodes; determine the edge weight of each edge; and construct a preset pragmatic relation graph based on all data nodes, all edges, and the edge weight of each edge.
[0191] Optionally, the task retrieval device is further configured to: collect the number of associations between any two data objects, wherein the number of associations is divided into: reference frequency, co-occurrence frequency, and derivative frequency; determine the maximum reference frequency, maximum co-occurrence frequency, and maximum derivative frequency from all association counts; determine the time interval at which the latest association relationship is generated between any two data objects; for any two data objects, determine the degree of association between the two data objects based on the reference frequency, co-occurrence frequency, derivative frequency, maximum reference frequency, maximum co-occurrence frequency, maximum derivative frequency, and time interval; and define the degree of association as the edge weight between the two data nodes indicated by the two data objects.
[0192] Optionally, the splitting unit is also used to: parse the query task to obtain multiple query targets and the data type, data range, and time constraints corresponding to each query target; for each query target, construct sub-tasks based on the query target, data type, data range, and time constraints, wherein, based on the data type, a preset multi-dimensional evaluation weight vector is configured for the sub-task, and the query conditions for the sub-task are constructed based on the query target, data type, data range, and time constraints; calculate the complexity of each sub-task; if the complexity is greater than a preset complexity threshold, re-split the query task until the complexity of each sub-task obtained by splitting is less than or equal to the preset complexity threshold, thus obtaining multiple sub-tasks.
[0193] Optionally, the first determining unit is further configured to: construct a data semantic vector for each data object in a preset pragmatic relation graph, and construct a query semantic vector for each subtask; for each query semantic vector, calculate the matching score between the query semantic vector and each keyword in each data object; for each subtask, traverse the preset pragmatic relation graph and characterize the traversed data objects as the current data object; calculate the initial matching degree between the subtask and the current data object based on the query semantic vector, the data semantic vector of the current data object, the matching score of each keyword in the current data object, and the number of keywords; determine the data distribution information entropy of the current data object and its operational activity at each time point, and determine the historical average activity of the current data object; calculate the matching score between the subtask and the current data object based on the initial matching degree, the data distribution information entropy, all operational activities, and the historical average activity; and determine the data object indicated by the maximum matching score as the target data object.
[0194] Optionally, the first determining unit is further configured to: traverse a preset pragmatic relation graph using the target data object, filter out all data objects whose edge weights are greater than a first preset weight threshold, and obtain an initial candidate data object set; determine whether the number of retrieved content based on the initial candidate data object set meets a preset content quantity; if the number of retrieved content does not meet the preset content quantity, continue to filter out data objects whose edge weights are greater than a second preset weight threshold, and add the data objects to the initial candidate data object set, until the number of retrieved content meets the preset content quantity, and obtain a candidate data object set, wherein the first preset weight threshold is greater than the second preset weight threshold.
[0195] Optionally, the judgment unit is further configured to: determine the object score returned by the distributed nodes that have completed the task, wherein the object score is calculated by the distributed nodes based on the individual score of the data object in each dimension and the preset multi-dimensional evaluation weight vector; determine the maximum score returned, and determine the historical average score and historical score standard deviation of all distributed nodes that have not completed the task; determine the global termination score based on the historical average score and historical score standard deviation; stop the execution of all distributed nodes that have not completed the task if the maximum score is greater than the global termination score; determine the average deviation based on all returned data objects; determine the task score of the subtask based on the object scores of all returned data objects, the number of all returned data objects, and the average deviation; determine that the subtask has been completed if the task score is greater than the preset task threshold; and reconfigure the query conditions and preset multi-dimensional evaluation weight vector of the subtask if the task score is less than or equal to the preset task threshold, and re-execute the subtask.
[0196] Optionally, the task retrieval device is further configured to: after determining the global termination score, if the maximum score is less than or equal to the global termination score, determine the average object score of all object scores returned by each currently running distributed node; determine the average of the lowest object scores returned by all distributed nodes that have completed tasks, and determine the standard deviation of all distributed nodes that have completed tasks; determine the abnormal score based on the average and standard deviation; and if the average object score is less than the abnormal score, identify the distributed node as an abnormal node and stop the execution of the abnormal node.
[0197] The task retrieval device described above may also include a processor and a memory. The splitting unit 70, the first determining unit 71, the second determining unit 72, the judgment unit 73, etc., are all stored in the memory as program units, and the processor executes the program units stored in the memory to realize the corresponding functions.
[0198] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, it can determine whether a subtask has been completed based on multiple search results. If all subtasks have been completed, it determines the target search result corresponding to the query task based on all search results.
[0199] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.
[0200] This invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: receiving a query task initiated by a user, splitting the query task into multiple subtasks, determining the target data object corresponding to each subtask based on a preset pragmatic relational graph, determining a set of candidate data objects based on the target data objects, determining a set of target distributed nodes for executing the subtasks based on the set of candidate data objects, determining whether the subtasks have been completed based on multiple search results, and determining the target search result corresponding to the query task based on all search results if all subtasks have been completed.
[0201] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the task retrieval method described above.
[0202] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the task retrieval method described above.
[0203] Figure 8 This is a hardware structure block diagram of an electronic device (or mobile device) for a task retrieval method according to an embodiment of the present invention. Figure 8 As shown, an electronic device may include one or more processors (e.g., Figure 8 The processors 802a, 802b, ..., 802n, etc., may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), and a memory 804 for storing data. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports in the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 8 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, the electronic device may also include components that are more... Figure 8 The more or fewer components shown, or having the same Figure 8 The different configurations shown.
[0204] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0205] The embodiments or examples disclosed herein are not exhaustive, but merely illustrative of some embodiments or examples, and are not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment or example can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment or example can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment or example can be arbitrarily interchanged. Furthermore, optional methods or examples in a particular embodiment or example can be arbitrarily combined; moreover, embodiments or examples can be arbitrarily combined. For example, some or all steps of different embodiments or examples can be arbitrarily combined, and a particular embodiment or example can be arbitrarily combined with optional methods or examples of other embodiments or examples.
[0206] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0207] In the several embodiments provided by this invention, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces; the indirect coupling or communication connection of units or modules can be electrical or other forms.
[0208] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0209] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0210] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0211] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A task retrieval method, characterized in that, Applied to intelligent agents, including: The system receives a query task initiated by a user and breaks it down into multiple subtasks. Each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector. Based on a preset pragmatic relational graph, a target data object corresponding to each subtask is determined, and a candidate data object set is determined based on the target data object. The preset pragmatic relational graph includes multiple data nodes, each data node representing a data object, and each data object corresponding to a distributed node. Based on the candidate data object set, a target distributed node set for executing the sub-task is determined, wherein each target distributed node in the target distributed node set is used to execute the sub-task and return multiple search results, the search results including at least: the retrieved data object and the object score of the retrieved data object; Based on multiple search results, it is determined whether the subtask has been completed. If all the subtasks have been completed, the target search result corresponding to the query task is determined based on all the search results.
2. The task retrieval method according to claim 1, characterized in that, Before receiving user-initiated query tasks, it also includes: Collect the data objects stored in each enterprise department system, and represent each enterprise department system as a distributed node; Each of the data objects is represented as one of the data nodes; For every two data nodes, if there is an association between the data objects corresponding to the two data nodes, then an edge is constructed between the two data nodes; Determine the edge weight for each edge; The preset pragmatic relation graph is constructed based on all the data nodes, all the edges, and the edge weight of each edge.
3. The task retrieval method according to claim 2, characterized in that, The steps to determine the edge weight of each edge include: The number of associations between any two data objects is collected, wherein the number of associations is divided into: reference frequency, co-occurrence frequency, and derivation frequency; Determine the maximum citation frequency, maximum co-occurrence frequency, and maximum derivative frequency from all the aforementioned association counts; Determine the time interval at which the latest association is established between any two of the data objects; For any two data objects, the degree of association between the two data objects is determined based on the reference frequency, the co-occurrence frequency, the derived frequency, the maximum reference frequency, the maximum co-occurrence frequency, the maximum derived frequency, and the time interval. The correlation is determined as the edge weight between the two data nodes indicated by the two data objects.
4. The task retrieval method according to claim 1, characterized in that, The steps of splitting the query task into multiple subtasks include: The query task is parsed to obtain multiple query targets and the data type, data range, and time constraints corresponding to each query target; For each query target, a subtask is constructed based on the query target, the data type, the data range, and the time constraint. The subtask is configured with a preset multi-dimensional evaluation weight vector based on the data type, and the query conditions for the subtask are constructed based on the query target, the data type, the data range, and the time constraint. Calculate the complexity of each of the subtasks; If the complexity is greater than a preset complexity threshold, the query task is split again until the complexity of each subtask is less than or equal to the preset complexity threshold, resulting in multiple subtasks.
5. The task retrieval method according to claim 1, characterized in that, The step of determining the target data object corresponding to each subtask based on a preset pragmatic relational graph includes: Construct a data semantic vector for each data object in the preset pragmatic relation graph, and construct a query semantic vector for each subtask; For each of the query semantic vectors, calculate the matching score between the query semantic vector and each keyword in each of the data objects; For each subtask, the preset pragmatic relation graph is traversed, and the traversed data object is represented as the current data object; Based on the query semantic vector, the data semantic vector of the current data object, the matching score of each keyword in the current data object, and the number of keywords, the initial matching degree between the subtask and the current data object is calculated; Determine the data distribution information entropy of the current data object and its operational activity at each time point, and determine the historical average activity of the current data object; Based on the initial matching degree, the data distribution information entropy, all the running activity levels, and the historical average activity level, calculate the matching score between the subtask and the current data object; The data object indicated by the maximum matching score is identified as the target data object.
6. The task retrieval method according to claim 1, characterized in that, The step of determining a set of candidate data objects based on the target data object includes: Using the target data object, traverse the preset pragmatic relation graph, filter out all data objects whose edge weights are greater than the first preset weight threshold, and obtain an initial candidate data object set; Determine whether the number of retrieved contents based on the initial candidate data object set meets the preset content quantity; If the number of searched contents does not meet the preset number of contents, the data objects with edge weights greater than the second preset weight threshold are further filtered out and added to the initial candidate data object set until the number of searched contents meets the preset number of contents, thus obtaining the candidate data object set, wherein the first preset weight threshold is greater than the second preset weight threshold.
7. The task retrieval method according to claim 1, characterized in that, The step of determining whether the subtask has been completed based on multiple search results includes: The distributed node that has completed the task returns the object score, wherein the object score is calculated by the distributed node based on the individual score of the data object in each dimension and the preset multi-dimensional evaluation weight vector; Determine the maximum score returned, and determine the historical average score and historical score standard deviation of all the distributed nodes that have not completed the task; The global termination score is determined based on the historical average score and the historical score standard deviation. If the maximum score exceeds the global termination score, the execution of all distributed nodes with incomplete tasks is stopped. Based on all the returned data objects, determine the average degree of deviation; The task score of the subtask is determined based on the object score of all returned data objects, the number of all returned data objects, and the average deviation. If the task score is greater than a preset task threshold, the subtask is determined to be completed. If the task score is less than or equal to the preset task threshold, the query conditions and the preset multi-dimensional evaluation weight vector of the subtask are reconfigured, and the subtask is re-executed.
8. The task retrieval method according to claim 7, characterized in that, After determining the global termination score, the following is included: If the maximum score is less than or equal to the global termination score, determine the average object score of all object scores returned by each currently running distributed node. Determine the average of the lowest object scores returned by all the distributed nodes that have completed the tasks, and determine the standard deviation of all the distributed nodes that have completed the tasks. Anomaly scores are determined based on the mean and the standard deviation. If the average object score is less than the abnormal score, the distributed node is identified as an abnormal node, and the execution of the abnormal node is stopped.
9. A task retrieval device, characterized in that, Applied to intelligent agents, including: The splitting unit is used to receive a query task initiated by a user and split the query task into multiple subtasks, wherein each subtask includes at least: query conditions and a preset multi-dimensional evaluation weight vector; The first determining unit is configured to determine the target data object corresponding to each subtask based on a preset pragmatic relational graph, and to determine a set of candidate data objects based on the target data objects, wherein the preset pragmatic relational graph includes: multiple data nodes, each data node representing a data object, and each data object corresponding to a distributed node; The second determining unit is used to determine a set of target distributed nodes for executing the sub-task based on the set of candidate data objects, wherein each target distributed node in the set of target distributed nodes is used to execute the sub-task and return multiple search results, the search results including at least: the retrieved data objects and the object rating of the retrieved data objects; The judgment unit is used to determine whether the subtask has been completed based on multiple search results, and, if all the subtasks have been completed, to determine the target search result corresponding to the query task based on all the search results.
10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the task retrieval method according to any one of claims 1 to 8.
