Intelligent data element query method and system based on multi-agent cooperation

Abstract

The method and system for intelligent query of data elements based on multi-agent cooperation belong to the technical field of data element circulation. The method comprises the following steps: a demand analysis agent disassembles a natural language query into atomic subtasks and constructs a dependency graph; a collaborative scheduling agent schedules a metadata management agent to pull a metadata subset and construct a vector index according to the graph, schedules a demand reasoning agent to perform semantic matching and link reasoning to generate a candidate asset set, and schedules a cross-domain and cross-source query execution agent to generate a query plan and initiate a query to a data source; a permission and security management agent performs security verification and desensitization processing on the query result; and the collaborative scheduling agent fuses and resolves conflicts of the processed result to generate a final result set. Through multi-agent division of labor and cooperation, the technical scheme improves the query semantic understanding ability, cross-source processing efficiency, and result accuracy and security.

Description

Technical Field

[0001] This invention belongs to the field of data element circulation technology, specifically relating to a data element intelligent query method and system based on multi-agent collaboration. Background Technology

[0002] The marketization of data elements is deepening, and the application scenarios of data element circulation across industries, regions, and levels are accelerating. Data assets are showing typical characteristics of being multi-sourced and heterogeneous, cross-domain and decentralized, and stored independently in nodes.

[0003] Against this industry backdrop, new demands have emerged for data element querying. For example, there is a need to complete collaborative queries of data assets across multiple independent nodes and heterogeneous data sources through a single entry point. Furthermore, there are higher requirements from the demand side for the semantic understanding accuracy of query requests and the business matching degree between query results and recommended content. Technical solutions need to be deeply adaptable to complex business scenarios to achieve accurate decomposition of requirements and compliant, personalized, and accurate output of results.

[0004] Existing data element query technologies struggle to achieve efficient collaborative retrieval of cross-domain, multi-source, heterogeneous data, and lack accuracy in complex semantic understanding and multi-hop reasoning. They fail to meet the core need of users to quickly obtain accurate data. Specifically, existing technologies suffer from the following shortcomings: 1. Existing semantic query solutions based on a single large model achieve demand parsing and semantic matching through a single model, but they have problems such as long inference chains that are prone to illusions, insufficient ability to decompose complex cross-source query tasks, and inability to balance professionalism and execution efficiency. 2. Existing data recommendation technologies based on fixed business rules or traditional collaborative filtering can only achieve shallow behavioral associations and cannot deeply integrate core information such as business scenarios, data lineage, and user permissions. This can easily lead to a disconnect between recommendation results and actual business needs, and even the risk of sensitive or unauthorized data being recommended incorrectly, reducing the efficiency and availability of data queries. 3. Most existing solutions are geared towards building retrieval capabilities for single-source, same-domain data; while existing cross-source data query technologies cannot achieve efficient collaborative retrieval of multi-node, dispersed data from a single entry point.

[0005] Therefore, there is an urgent need for an intelligent query solution that enables cross-domain, multi-source data collaborative retrieval, accurate semantic understanding, and personalized recommendations. Summary of the Invention

[0006] To overcome at least some of the problems existing in related technologies, this invention provides a data element intelligent query method and system based on multi-agent collaboration. It aims to improve the semantic parsing capability, cross-source execution efficiency, and accuracy and security of result output of data element queries through multi-agent collaborative scheduling and hierarchical reasoning.

[0007] First aspect The intelligent data element query method based on multi-agent collaboration proposed in this application is applied to an intelligent data element query system. This system comprises several agents connected to a pre-built standardized data catalog and knowledge base. The agents include at least a demand parsing agent, a demand reasoning agent, a collaborative scheduling and feedback agent, a metadata management agent, multiple cross-domain and cross-source query execution agents, and an authorization and security control agent. The method includes: The demand parsing agent receives natural language query requests from demand users, preprocesses the query requests, breaks them down into multiple atomic subtasks, and constructs a dependency graph of the atomic subtasks. The collaborative scheduling and feedback agent, based on the dependency graph, schedules the metadata management agent to pull an accessible subset of metadata from the standardized data directory based on user permission scope and build a real-time vector index for it. The agent also schedules the demand reasoning agent to perform semantic matching and link reasoning based on the real-time vector index and the knowledge base to generate a candidate data asset set. Finally, the agent schedules the corresponding cross-domain and cross-source query execution agent to generate a query plan based on the candidate data asset set, initiate a query to the supplier data source, and obtain the query results. The permission and security management intelligent agent performs security verification and desensitization processing on the query results; The collaborative scheduling and feedback agent integrates and resolves conflicts in the processed query results, generates a final result set, and outputs it to the demand user.

[0008] Second aspect This application proposes a data element intelligent query system, which includes: The user-side layer is used to provide operation entry points and result display; The service gateway layer communicates with the client layer and is used for access requests, authentication, and encrypted transmission. The intelligent agent cluster layer communicates with the service gateway layer and includes multiple intelligent agent nodes; The global storage layer, which communicates with the intelligent agent cluster layer, is used for distributed storage of standardized data catalogs, knowledge bases, metadata, knowledge graphs, and vector indexes. A heterogeneous data source interface layer, which communicates with the intelligent agent cluster layer, is used to adapt to heterogeneous data sources and complete query integration; wherein, the plurality of intelligent agent nodes include at least: The demand parsing agent is used to receive natural language query requests from demand users, preprocess the query requests, break them down into multiple atomic subtasks, and construct a dependency graph of the atomic subtasks. A metadata management agent is used to pull an accessible subset of metadata from the standardized data directory based on user permission scope and build a real-time vector index for it; A demand reasoning agent is used to perform semantic matching and link reasoning based on the real-time vector index and the knowledge base to generate a candidate data asset set. Multiple cross-domain and cross-source query execution agents are used to generate query plans based on the candidate data asset set and initiate queries to the supplier data source to obtain query results. An intelligent agent for permission and security management is used to perform security verification and desensitization processing on the query results; The collaborative scheduling and feedback agent is connected to the demand parsing agent, the metadata management agent, the demand reasoning agent, the cross-domain and cross-source query execution agents, and the permission and security control agent, respectively. It is used to schedule the collaborative work of each agent according to the dependency graph, and to merge and resolve the query results after security verification and desensitization, and generate a final result set to be output to the demand user.

[0009] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention breaks down complex query tasks into multiple atomic subtasks and constructs a dependency graph through the professional division of labor and collaborative scheduling of multiple agents. Each agent then collaboratively performs metadata retrieval, semantic matching, link reasoning, and cross-source querying based on the dependency graph. This effectively avoids the illusion problem that can easily occur with long inference chains of a single large model, and significantly improves the accuracy and execution efficiency of complex cross-source queries. 2. By using a demand reasoning agent to perform semantic matching and link reasoning based on real-time vector indexing and a pre-built knowledge base, combined with a metadata management agent to perform real-time pre-filtering based on user permission scope, and a permission and security control agent to perform security verification and de-identification of query results, high-precision matching between query results and user business needs is achieved while ensuring data security and compliance. 3. By using a collaborative scheduling and feedback intelligent agent to integrate and resolve conflicts in query results, such as deduplication, format standardization, and comprehensive sorting based on preset weights for multi-source query results, the accuracy, consistency, and readability of the final result set are further improved. Attached Figure Description

[0010] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0011] Figure 1 This is a flowchart illustrating the intelligent data element query method of the present invention. Figure 2 This is a schematic diagram illustrating the pre-process of the present invention. Detailed Implementation

[0012] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0013] In the technical solution proposed in this application, the intelligent data element query method based on multi-agent collaboration is applied to the intelligent data element query system. In actual implementation, the intelligent data element query system can run on the data element circulation platform. In some embodiments, the data element circulation platform takes the trusted data space as its landing form and is deployed on a Linux server cluster. The cluster provides the resource support required for platform construction, business carrying and business operation.

[0014] The intelligent data element query system proposed in this application comprises several intelligent agents and is connected to a pre-built standardized data catalog and knowledge base. Please refer to [link / reference]. Figure 1 and Figure 2 In a real-world implementation scenario, the overall application scenario involves pre-processes (corresponding to...) Figure 2 ) and core processes (corresponding to Figure 1 The pre-process provides a standardized foundation for the core process, namely, the pre-construction of standardized data catalogs and knowledge bases, and the core process realizes end-to-end response to the needs of the client.

[0015] This application's intelligent data element query method based on multi-agent collaboration mainly involves the core process, in which several agents include at least a demand parsing agent, a demand reasoning agent, a collaborative scheduling and feedback agent, a metadata management agent, multiple cross-domain and cross-source query execution agents, and an authorization and security control agent. Each agent operates independently and achieves message communication and data sharing through a unified agent collaboration bus. Similarly, the pre-process ( Figure 2 The standardized data catalog and knowledge base pre-construction process (as shown) can also be implemented in a similar way.

[0016] Based on this, in some embodiments, within the overall technical scenario covered by this application, the plurality of intelligent agents may further include a data catalog building intelligent agent; and to achieve personalized supplementary recommendations for query results, the plurality of intelligent agents may further include a personalized recommendation intelligent agent. Each intelligent agent operates independently and achieves message communication and data sharing through a unified intelligent agent collaboration bus. Table 1 below defines each intelligent agent.

[0017] Table 1 - Definition of Multi-Agent Cluster To facilitate understanding of the technical solution of this application, a brief introduction to the implementation of the pre-process is given below, based on Table 1 and Figure 2 As shown, the specific steps of the pre-process are as follows: Step P1: Supplier data source access and asset scanning; a: The metadata management intelligent agent receives data source authorization, asset listing application, national / industry catalog standard configuration, and compliance rule configuration information from data element suppliers, and completes the initial collection and archiving of supplier access information; b: The collaborative scheduling and feedback intelligent agent schedules the data catalog to build an intelligent agent, which connects to heterogeneous data sources from suppliers (including relational databases, data warehouses, data lakes, API interfaces, etc.), completes the automatic scanning of the full data assets, extracts core information such as original technical metadata, data structure, data samples, and update frequency, and forms the original data asset set; c: Synchronously trigger the permission and security control intelligent agent to perform secure access verification, data source legality audit and initial identification of sensitive data for the supplier data source, block illegal data source access, and ensure the compliance and security of the access data source.

[0018] Step P2: Metadata standardization and compliance verification; a: The data catalog building agent performs standardized cleaning, format conversion, missing field completion, and duplicate data removal on the original technical metadata to generate a standardized technical metadata dataset; b: Collaborative permission and security management intelligent agent, based on preset classification rules, completes automatic hierarchical classification of data assets, comprehensive compliance verification, identifies state secrets, work secrets, sensitive personal information, non-compliant data, generates compliance verification reports and rectification suggestions, and initiates interactive rectification prompts to suppliers; It should be noted here that: The preset classification rules involved here are set in accordance with the requirements of the "Data Security Technology Data Classification and Grading Rules" (Standard No.: GB / T43697-2024).

[0019] c: For compliant assets that have completed rectification, collaborate with the metadata management intelligent agent to complete the initial version of standardized metadata storage and build a basic metadata index.

[0020] Step P3: Construction and semantic association of business knowledge base; a: The intelligent agent for building the data catalog is based on a nationally unified business terminology database and industry knowledge graph. It automatically annotates standardized metadata with business semantics, and completes the accurate association mapping between business terms, industry indicators and technical metadata. b: Based on the business attributes, application scenarios, relationships, and usage specifications of data assets, automatically build corresponding business knowledge bases to supplement semantic information such as asset application instructions, business scenario cases, usage guidelines, related indicators, and upstream and downstream asset relationships; c: Collaborative demand reasoning intelligent agent, completes the association and alignment of the business knowledge base built this time with the global business knowledge graph, verifies the consistency of business logic, completes the cross-asset and cross-domain association link information, and ensures that the catalog content adapts to the semantic retrieval needs of the demand side.

[0021] Step P4: Standardized data catalog generation and closed-loop iteration; a: The intelligent agent for building the data catalog automatically generates a formal data catalog that conforms to national / industry standards based on standardized metadata, business knowledge base, and compliance verification results. This includes basic catalog information, metadata details, business semantic information, compliance classification information, data access rules, and sharing and open scope. It also outputs the final version of the catalog, which can be further received from suppliers for manual correction, supplementary information, and listing confirmation instructions. b: After the supplier confirms the listing, the collaborative metadata management intelligent agent completes the full entry of the formal data catalog and knowledge base, and updates the distributed vector index. If a finely tuned industry-specific BERT embedding model is used, the metadata and business semantic information are converted into multi-dimensional vectors, and a distributed vector index is built based on the HNSW algorithm. This index is then synchronized to the demand reasoning intelligent agent and the personalized recommendation intelligent agent to officially provide query and recommendation services to the demand side. c: The collaborative scheduling and feedback intelligent agent receives real-time query feedback and usage behavior data from the demand side, and synchronizes it back to the data element supply-side catalog building intelligent agent to continuously iterate and optimize catalog content, business semantic annotation, and knowledge base coverage, forming a closed-loop optimization of supply and demand.

[0022] Based on this, the standardized data catalog and knowledge base generated from the above-mentioned pre-process provide services, while incorporating user trust verification, query text matching, and result sorting output operations to realize the corresponding core processes. It should be noted that the standardized data catalog and knowledge base can also be pre-built in other ways, and the construction method does not affect the implementation of the query method of this application.

[0023] In this implementation scenario, the pre-process constructs a standardized data catalog and knowledge base using the methods described above. Based on this, the intelligent data element query method in this application includes: The demand parsing agent receives natural language query requests from demand users, preprocesses the query requests, breaks them down into multiple atomic subtasks, and constructs a dependency graph of the atomic subtasks. The collaborative scheduling and feedback agent, based on the dependency graph, schedules the metadata management agent to pull accessible metadata subsets from the standardized data directory based on user permission scope and constructs a real-time vector index for them. It then schedules the demand reasoning agent to perform semantic matching and link reasoning based on the real-time vector index and knowledge base, generating a candidate data asset set. Finally, it schedules the corresponding cross-domain, cross-source query execution agent to generate a query plan based on the candidate data asset set, initiates a query to the supplier data source, and retrieves the query results. The permission and security control agent performs security verification and anonymization on the query results. The collaborative scheduling and feedback agent then merges and resolves conflicts in the processed query results, generating a final result set and outputting it to the demand user.

[0024] The technical solution of this application will be further described below from a specific implementation perspective, with reference to embodiments. The core process of this application can be divided into 5 core steps, combined with... Figure 1 As shown, specifically: Step S1: User Request Access and System Initialization a) Obtain natural language query requests, user identity information, business context information, and historical interaction records from the demand side user. In implementation, a single, unique query entry point can be constructed through a web portal, mobile APP, or trusted data space connector. The demand side user submits a natural language query request through this entry point, and the system receives and synchronously collects the user identity information, business context information, and historical interaction records. In other words, this method can meet the user's expectation of querying all full content through a unified entry point, solving the problem mentioned in the background technology that existing related technical solutions cannot achieve efficient collaborative retrieval of multi-node scattered data through a single entry point.

[0025] b) The collaborative scheduling and feedback intelligent agent completes the system initialization, triggers the permission and security control intelligent agent based on the user's identity information, completes the verification of the trusted identity in the data element circulation market, and retrieves the user's permission scope, data access compliance rules, and the scope of accessible supplier directories. c) Based on the request type, the collaborative scheduling and feedback agent completes the resource allocation and communication link initialization of the multi-agent cluster.

[0026] In specific step S1, the permission and security control intelligent agent and the collaborative scheduling and feedback intelligent agent are used to collect natural language query requests, user identity information, business context information, and historical interaction records input by the demand user. The permission and security control intelligent agent performs real-time identity verification of the demand user and classifies the demand user into trusted demand users and untrusted demand users based on the verification results, thus obtaining user trust screening data. This step includes: Step S11: Collect data on knowledge base users who are currently logging into the target knowledge base to obtain multiple demand users. It should be noted that, in this application, the demand users are specifically knowledge base users who have logged into the target knowledge base and have data query and retrieval needs.

[0027] Step S12: During the process of monitoring the behavior of the user, the time point when the user completes the login operation is marked as the first monitoring feature time point, the time point corresponding to the current moment is marked as the second monitoring feature time point, and the time period between the first monitoring feature time point and the second monitoring feature time point is marked as the pre-selected behavior monitoring period. Step S13: Randomly select one sample user from the multiple acquired users, and perform user behavior analysis on the sample user in the pre-selected behavior monitoring period. Based on the analysis results, the sample user is screened into trusted users and untrusted users. In actual implementation, step S13 can be specifically as follows: set several different types of abnormal operation behaviors for the user permissions corresponding to the user in the sample demand side, and arbitrarily select a specific abnormal behavior from the multiple abnormal operation behaviors set. It should be noted that, in this application, the abnormal operation behavior referred to here specifically refers to behavior that exceeds the user's operation permissions. The abnormal operation behavior referred to here includes, but is not limited to, attempting to modify the knowledge base query log, attempting to export sensitive data beyond the user's permissions, and high-frequency intensive queries.

[0028] The operation time points of specific abnormal behaviors of sample demand users are marked to obtain multiple abnormal behavior marked time points. The time interval of any two abnormal behavior marked time points in an adjacent state is collected to obtain multiple abnormal behavior interval durations. The average value of the multiple abnormal behavior intervals is calculated to obtain the first abnormal behavior interval duration feature value. The standard deviation of the multiple abnormal behavior intervals is calculated to obtain the second abnormal behavior interval duration feature value. The sum of the first and second abnormal behavior interval duration feature values ​​is calculated to obtain the upper limit of the abnormal behavior interval. The difference between the first and second abnormal behavior interval duration feature values ​​is calculated to obtain the lower limit of the abnormal behavior interval. The numerical interval formed by the upper and lower limits of the abnormal behavior interval is set as the real-time interval interval of the abnormal behavior. Furthermore, historical users marked as untrusted demand users due to specific abnormal behaviors are obtained, and the real-time interval of abnormal behavior corresponding to each historical user is obtained. The union of the obtained real-time intervals of multiple abnormal behaviors is then used to obtain the preset merge interval for specific behaviors. The overlap between the real-time interval of abnormal behavior and the preset merging interval of specific behavior is collected to obtain the unreliable overlap of abnormal behavior. If an abnormal behavior is set to be untrustworthy, and the overlap of the untrustworthy abnormal behavior is greater than or equal to the untrustworthy abnormal behavior judgment value, then the sample demand user is directly marked as an untrustworthy user. If the overlap of the untrustworthy abnormal behavior is less than the untrustworthy abnormal behavior judgment value, then any abnormal operation behavior other than the specific abnormal behavior is used to replace the feature abnormal behavior, until each abnormal operation behavior replaces the feature abnormal behavior. If the overlap of the untrustworthy abnormal behavior corresponding to the replaced abnormal operation behavior is less than the untrustworthy abnormal behavior judgment value, then the corresponding sample demand user is marked as a trustworthy user, thus obtaining user trustworthy screening data.

[0029] Repeat the process of trust screening for sample demand users, and perform trust screening for each demand user separately to obtain user trust screening data.

[0030] It is understandable that the above screening logic is not limited to the specific path mentioned above in actual implementation. In the early stage of system operation or when the number of historical untrusted users is insufficient, the preset merging interval for specific behaviors may be temporarily empty. At this time, an initial attack interval based on security experience can be introduced as a temporary benchmark, or the judgment can be made by indicators such as behavior frequency threshold and permission overstepping level. After accumulating enough data, it can be smoothly transitioned to a judgment mode based on interval overlap. In addition, if the standard deviation of abnormal behavior intervals is extremely small such that the interval length approaches zero, the interval may be subjected to a predetermined micro-expansion process, or such highly regular abnormal interval patterns may be directly regarded as high-risk features to output an untrustworthy determination; if the occurrence count of a specific abnormal behavior is too small to construct a valid interval, the corresponding untrustworthy overlap degree may be regarded as zero, or other preset rules may be combined with the context such as the risk level of the behavior and user permissions for comprehensive judgment, and all these implementation variants do not depart from the technical concept of the above screening logic of the present application.

[0031] In the overall process of step S1, the cooperative scheduling and feedback agent is responsible for completing system initialization, further triggering the permission and security control agent based on user identity information and user trusted screening data, completing secondary verification of trusted identities in the data element circulation market, and pulling the permission scope of trusted users, data access compliance rules, and the accessible supplier directory scope; the cooperative scheduling and feedback agent completes resource allocation and communication link initialization of the multi-agent cluster based on the request type.

[0032] Step S2: Requirement parsing and atomic task decomposition In terms of requirement parsing, the cooperative scheduling and feedback agent schedules the requirement parsing agent (corresponding to Figure 1 the Agent requirement parsing agent in the drawing) to preprocess the user's natural language request, including text error correction, stop word filtering, and business entity extraction, that is, the requirement parsing agent receives the natural language query request from the demander user and preprocesses the query request, where the preprocessing at least includes text error correction, stop word filtering, and business entity extraction; Specifically, the requirement parsing agent may invoke multiple text processing tools, and the specific implementation manners of preprocessing are as follows: For text error correction, after receiving the user's natural language query, a spelling error correction tool (such as a confusion dictionary based on edit distance or prompt word error correction based on a large language model) may be invoked; for stop word filtering, a stop word list (including words and punctuation marks with no actual semantic meaning such as "de", "le", "ma", and "ne") may be maintained, the corrected text is segmented, and words in the stop word list are filtered out; for business entity extraction, a pre-built business term library (such as that built in the pre-process) and the named entity recognition capability of the large language model may be utilized, or relying on the business term library in combination with the NER named entity recognition model, key entities are extracted from the filtered text, and these key entities include core entities such as business indicators, data subjects, time ranges, and statistical dimensions.

[0033] In this way, through the above manner, the requirement parsing agent converts the original natural language query into a structured and standardized intermediate representation for subsequent intent recognition and task decomposition.

[0034] Based on this, the demand parsing intelligent agent, using the business terminology library, business knowledge base, and large language model built from the pre-process, completes the standardized mapping and alignment of user intent recognition, business terminology to technical metadata, distinguishes between query-type demands, recommendation-type demands, and hybrid demands, and solves the semantic gap between users' colloquial expressions and technical metadata. The large language model here is an industry-specific large language model, which is based on a general basic large model and trained by domain fine-tuning and adaptation using an industry-annotated corpus dataset.

[0035] In step S2, regarding the decomposition of atomic tasks, as shown in Figure 1, the user's complex needs are decomposed into dependency-free atomic subtasks using the demand parsing agent. Specifically, the atomic subtasks here include at least metadata retrieval subtasks, semantic reasoning subtasks, cross-source data query subtasks, personalized recommendation subtasks, and compliance verification subtasks, and a dependency graph of atomic subtasks is constructed. In real-world scenarios, user-inputted queries may contain vague requirements or missing information, hindering the query process. Therefore, as a specific implementation method, the technical solution of this application, for query requests with vague requirements or missing information, involves a requirement parsing agent generating interactive clarification prompts and initiating supplementary inquiries to the user until the query request can be completely broken down into atomic subtasks.

[0036] In practice, the requirement parsing agent, combined with the business entity extraction results, calculates the degree of ambiguity of the extracted non-standardized terminology set. Using the AHP-fuzzy comprehensive evaluation model, it quantifies the fuzziness level of the requirements from four dimensions: terminology ambiguity, semantic completeness, entity coverage, and business matching. The requirements are divided into three fuzziness levels: low, medium, and high, and corresponding processing strategies are adopted. For low ambiguity levels, proceed directly to the atomic task decomposition process; for medium ambiguity levels, supplement default parameters based on user history and business context, generate alternative requirement solutions, and simultaneously send a one-click confirmation prompt to the user; for high ambiguity levels, automatically generate interactive clarification scripts, send supplementary questions to the user until complete requirement information is obtained, ensuring that the query request can be completely decomposed.

[0037] The expression form of the AHP-fuzzy comprehensive evaluation model is as follows: S=W1·R1+W2·R2+W3·R3+W4·R4 (1) In expression (1), S represents the fuzzy comprehensive score of the demand (0~1). In an implementation scenario, the fuzzy comprehensive score S is used to classify and grade the requirements. The specific grading rules are: S≤0.3 corresponds to low fuzziness, 0.3<S≤0.6 corresponds to medium fuzziness, and S>0.6 corresponds to high fuzziness.

[0038] In expression (1), W1 to W4 represent the AHP weights of the four dimensions. The weights can be calculated by inviting experts in the relevant fields to conduct pairwise importance comparisons, constructing a hierarchical analysis judgment matrix, and then performing a consistency test.

[0039] In expression (1), R1 to R4 represent the fuzzy membership scores (0 to 1) of the four dimensions, specifically; R1 corresponds to the degree of ambiguity of terms, which is used to determine the degree of ambiguity caused by non-standard business terms, polysemous words, and colloquial expressions in user queries. The calculation method is R1 = number of ambiguous terms / total number of extracted terms. No ambiguity R1→0, all ambiguous R1→1.

[0040] R2 corresponds to semantic completeness, which is used to determine whether the query contains necessary semantic elements such as time, region, indicator, object, scope, and statistical caliber. The calculation method is R2=1-(number of filled components / number of standard necessary components). Information completeness R2→0, information severely missing R2→1.

[0041] R3 corresponds to entity coverage, which is the proportion of business entities (metrics, subjects, dimensions, tags, etc.) successfully extracted from the query out of the entities that should be extracted. The calculation method is R3=1-(actual number of extracted entities / expected number of extracted entities). If all entities are extracted, R3→0; if very few are extracted, R3→1.

[0042] R4 corresponds to the business matching degree, which is used to determine the degree of matching between the query intent and the standard business scenario, data directory, and knowledge base. R4 = 1 - (demand - scenario semantic similarity). Complete match R4 → 0, complete mismatch R4 → 1.

[0043] In the process of determining R4, the semantic similarity between the demand and the scenario refers to the degree of semantic closeness between the business intent vector of the user's natural language query and the standard business scenario vector built into the system. It is used to measure the degree of matching between the user's demand and the existing data assets, business knowledge base, and data catalog. The value range is [0,1], and the closer it is to 1, the higher the degree of matching. The semantic similarity between demand and scenario is determined by calculating the vector cosine similarity. For the business intent vector, the user query text is input into a pre-trained industry semantic embedding model (such as a fine-tuned BERT or an industry-specific embedding model), and the vector is obtained through model encoding. For the standard business scenario vector, the semantic vectors corresponding to each standard business scenario (such as "industrial revenue statistics" and "new energy vehicle sales analysis") are extracted from the business knowledge base as the corresponding vectors.

[0044] In the above process, the non-standardized term set is obtained in the following way: After the demand parsing agent completes the extraction of business entities, it further performs standardization verification on the extracted business entities, comparing each extracted business entity with the standardized terms in the pre-built business term library and cross-industry data element knowledge graph: if the entity completely matches the standardized term, it is marked as a "standardized term"; if the entity has a semantic relationship with the standardized term but is not completely consistent, or cannot match any standardized term, it is classified into the non-standardized term set.

[0045] To facilitate understanding of this technical solution, the following examples illustrate scenarios where ambiguous requirements or missing information arise in practice, and the corresponding coping methods.

[0046] In one implementation scenario, if a user inputs: "Query the revenue data of industrial enterprises above designated size in Beijing in the first quarter of 2025", after evaluation across various dimensions, the terminology ambiguity is 0.1 (unambiguous, industry standard terminology), the semantic completeness is 0.95 (time, region, indicator, and scope are all clear), the entity coverage is 0.9 (all core business entities are extracted), and the business matching degree is 0.95 (completely matches the data query scenario). After comprehensive evaluation, it is determined to be of low fuzziness level (clear requirements, unambiguous), and then directly enters the atomic task decomposition process without interactive clarification. Atomic tasks are broken down as follows: Metadata retrieval: Metadata related to industrial enterprise revenue; Semantic reasoning: Matching Beijing, Q1 2025, and enterprises above designated size; Cross-source data query: Invokes the corresponding data source to execute the query; Compliance verification: Permission and data security verification.

[0047] For example, if a user inputs "Query the sales data of new energy vehicles in the past six months", after evaluation across various dimensions, the terminology ambiguity is 0.3 ("new energy vehicles" is unambiguous), semantic completeness is 0.6 (the region, vehicle type, and statistical scope are not clearly defined), entity coverage is 0.7 (core indicators are covered, but dimensions are missing), and business matching is 0.8 (matches the sales query requirement). After comprehensive evaluation, it is determined to be of medium fuzziness level (the semantics are basically complete, but a few parameters are missing). The processing method involves supplementing default parameters based on user history / business context, generating alternative solutions, and initiating a one-click confirmation prompt. The interactive clarification script could be: "We have queried the sales data of new energy vehicles nationwide, including pure electric and plug-in hybrid models, according to official statistics for the most recent six months. Do you confirm? [Confirm] [Modify Parameters]."

[0048] For example, if a user inputs "Check car-related data", after evaluation across various dimensions, the terminology ambiguity is 0.8 ("car-related" is extremely broad and has no clear target), semantic completeness is 0.2 (no time, region, indicator, or dimension), entity coverage is 0.3 (only the vague entity "car" is extracted), and business matching is 0.4 (cannot match specific business needs). After comprehensive evaluation, it is determined to be a high level of ambiguity (semantic incompleteness, ambiguity, and cannot be directly broken down). An interactive clarification script is automatically generated, and multiple rounds of questions are asked to supplement key information until the requirement is complete. The interactive clarification process is as follows: First round: Would you like to inquire about car sales, production, ownership, or import / export data? Second round: Which province / national data do we need to query? Third round: Do you want to query data from the past year, the past three years, or a specific time period? Fourth round: Is it necessary to differentiate between gasoline vehicles, new energy vehicles, and commercial vehicles? After clarification and definition of the needs, the user finally specified that they wanted to query the sales data of new energy passenger vehicles in Guangdong Province in 2024, and then the system would perform atomic task breakdown.

[0049] This application achieves hierarchical identification and interactive clarification of fuzzy requirements through the aforementioned methods. Compared with traditional fixed rules and single-dimensional threshold judgments, which rely on hard-coded conditional branching, this approach effectively overcomes the problems of low accuracy, poor adaptability, lack of intermediate states, weak interpretability, and difficulty in iteration associated with traditional methods. The model constructs an evaluation system based on four dimensions: terminology ambiguity, semantic completeness, entity coverage, and business matching. This effectively avoids bias and misjudgment in evaluation. By using continuous scores of 0 to 1 to finely measure and classify low, medium, and high fuzziness levels, it can effectively handle semi-clarity and semi-fuzzy intermediate states. Furthermore, relying on AHP, it can flexibly adjust the dimension weights according to industry and business scenarios to achieve scenario-based customization and adaptation. It can also quantify and parse non-standardized query expressions such as colloquialisms, ambiguous terms, and fuzzy references, accurately distinguishing between valid business entities and invalid fuzzy content, and significantly improving the understanding and adaptation capabilities for complex natural language requirements.

[0050] Based on step S2, proceed to step S3: multi-agent hierarchical collaborative reasoning and task execution. In this step, based on Figure 1, the collaborative scheduling and feedback agents schedule corresponding agents to execute sub-tasks based on the sub-task dependency graph. Each agent synchronizes its status and data in real time through a unified message bus. The specific execution flow is as follows: a. Pre-existing metadata and permission filtering: This involves scheduling the metadata management agent, that is, scheduling the metadata management agent to pull the full set of compliant and standardized catalog metadata (technical metadata, business metadata, data lineage, data quality, hierarchical classification information) that the supplier has put on the platform based on the user's permission scope, and build a real-time vector index. That is, based on the user's permission scope, pull the accessible metadata subset from the standardized data catalog and build a real-time vector index for it, and synchronize it to the demand reasoning agent. Furthermore, during this process, the permission and security control intelligent agent simultaneously performs permission pre-verification of metadata, intercepting supplier directory metadata that users do not have permission to access from entering subsequent processes.

[0051] b. Semantic Matching and Link Reasoning: Based on the obtained real-time vector index, the demand reasoning agent is invoked. The demand reasoning agent calculates semantic similarity to the pre-filtered metadata index (i.e., real-time vector index) based on the decomposed atomic tasks, standardized business terms, the business knowledge base built by the supplier, and the global business knowledge graph. At the same time, combined with the query text matching data, it completes link reasoning based on data lineage and business domain association, and mines upstream and downstream supplier data assets related to user needs to generate a preliminary set of candidate data assets. At the same time, the candidate set is checked for business logic consistency, and assets that conflict with the demand business logic are eliminated.

[0052] c. Cross-domain and cross-source query plan generation and execution: This involves calling the cross-domain and cross-source query execution agent. Based on the user's precise data query needs, the cross-domain and cross-source query execution agent generates a distributed query execution plan that adapts to the heterogeneous data sources of cross-domain providers, and executes it, based on the semantic matching results (i.e., the candidate data asset set) and atomic task requirements. In this process, the collaborative scheduling and feedback agent selects an agent that matches the type of the target data source from multiple cross-domain and cross-source query execution agents based on the type of the target data source in the candidate data asset set. The scheduled cross-domain and cross-source query execution agent generates a distributed query execution plan adapted to the heterogeneous data source of the supplier it is responsible for, based on the semantic matching result (i.e., the candidate data asset set) and the atomic task requirements. It also performs syntax conversion and execution path optimization according to the database type of the supplier data source. Then, it initiates a query to one or more corresponding supplier data sources to obtain the original query results.

[0053] For each cross-domain, cross-source query execution agent, it first formats the acquired raw query results into a standardized query result set. If the agent is responsible for multiple source data providers, it further aggregates the multiple formatted query result sets internally (including merging, deduplication, and format unification) to generate the agent's standardized intermediate result set; if it is responsible for only one source data provider, the formatted query result set is directly used as the standardized intermediate result set. Afterward, each cross-domain, cross-source query execution agent sends its generated standardized intermediate result set to the collaborative scheduling and feedback agent.

[0054] When the candidate data asset set involves multiple source data providers, the collaborative scheduling and feedback agent schedules multiple corresponding cross-domain, cross-source query execution agents in parallel (each agent executes independently according to the above process), and receives standardized intermediate result sets sent by all agents, summarizing them to obtain a summarized query result set. Specifically, each cross-domain, cross-source query execution agent internally aggregates its own query results and sends them to the collaborative scheduling and feedback agent, which then summarizes the received aggregated results to obtain the summarized query result set.

[0055] It should be noted that the "formatting", "internal aggregation" and "summarization" in this step are all data-level structured processing operations to form a single result set, which do not contradict the subsequent integration and conflict resolution that are more business-related.

[0056] As an optional implementation, the agents also include a personalized recommendation agent (as defined in Table 1). The collaborative scheduling and feedback agent further schedules the personalized recommendation agent to generate a supplementary recommendation asset set based on user profiles, historical interaction behaviors, business context, and candidate data asset sets. Subsequently, the collaborative scheduling and feedback agent merges and resolves conflicts between the processed supplementary recommendation asset set and the query results (as mentioned above, the query result set) to generate the final result set. Correspondingly, step S3 may also include sub-step d. d. Personalized Recommendation Generation: Based on user profiles, historical interaction behavior, business context, semantic reasoning results, data quality scores, and asset popularity, the personalized recommendation agent generates a supplementary set of recommended assets related to the main query requirement, within the user's permission scope. This includes supplier-related data tables, metrics, fields, data APIs, analysis reports, and business application cases. For new users with no historical behavior, cold start recommendations are completed based on the demand user's industry, profile, and the logged-in user's department, combined with a business knowledge graph.

[0057] Step S3 also includes sub-step e, full-process security and compliance verification. This involves an intelligent agent for permission and security management. Specifically, the intelligent agent performs security verification and anonymization on the query results. This includes: secondary permission review to confirm that each piece of data in the query results is within the user's permission scope to prevent unauthorized access; sensitive data identification to detect whether the results contain sensitive information such as personal privacy or trade secrets; and data anonymization to blur, replace, or mask identified sensitive fields to ensure that the output results meet data security regulations. The query results processed in this way will then be sent to step S4 for fusion and conflict resolution.

[0058] When a personalized recommendation agent is also configured, the permission and security control agent performs the same security verification and de-identification processing on the supplementary recommendation asset set it generates, and generates a full-process operation audit log (including user identity, query time, accessed data source, de-identification records, etc.), which is synchronized to the data security control platform. For unauthorized operations or non-compliant requests, the permission and security control agent intercepts them and generates compliance prompts; In other words, during the full-process security and compliance verification process in sub-step e, the permission and security control intelligent agent performs a full secondary verification of the query result set and the supplementary recommended asset set, including secondary permission review, sensitive data identification, and data anonymization. At the same time, it generates a full-process operation audit log and synchronizes it to the data security control platform; it intercepts unauthorized operations and non-compliant requests and generates compliance prompts.

[0059] It is easy to understand that in step S3, the query result is generated by aggregating the query results of multiple agents. To ensure that the query results are unified, standardized, and consistent, the multi-source heterogeneous results need to be integrated and processed for consistency before output. The core process of the technical solution of this application also includes multi-agent result fusion and conflict resolution, such as... Figure 1 As shown, the collaborative scheduling and feedback agent fuses and resolves conflicts in the processed query results, generating a final result set which is then output to the user. That is, based on step S3, step S4 is performed: multi-agent result fusion and conflict resolution.

[0060] Step S4 includes: a. The collaborative scheduling and feedback agent obtains the aggregated query result set (which already includes the merging and formatting of cross-source query results) generated in sub-step c of step S3, and the supplementary recommended asset set (if personalized recommendations are enabled) generated in sub-step d of step S3, and merges the two to form the result set to be processed.

[0061] b: Deduplication and format standardization are performed on the merged result set (if not fully processed in the aggregation stage). For conflicting content, conflict resolution is performed based on preset weight rules (such as semantic matching degree 40%, business relevance 20%, data quality 15%, permission matching degree 15%, user preference 10%), and low-weight conflicting results are removed. c: Based on the above weighting rules, the final query results and recommendation results are comprehensively scored and ranked to generate the final result set.

[0062] For example, firstly, unique identifiers of data assets (such as data table IDs or field paths) can be used as primary keys to delete duplicate records. For different aliases or descriptions of the same asset returned by different data sources, asset mapping relationships (such as a table mapping business terms to standard metadata) provided by the metadata management agent are used for identification and merging. All results are then uniformly converted into a preset standard format (such as a JSON object containing fields such as asset name, metadata details, supplier information, and score).

[0063] Next, scores are calculated for each candidate asset across five dimensions, with each dimension ranging from 0 to 1. A comprehensive score is then calculated according to a preset weighting rule: Comprehensive Score = Semantic Matching Degree × 0.4 + Business Relevance × 0.2 + Data Quality Score × 0.15 + Permission Matching Degree × 0.15 + User Preference Score × 0.1. If an asset's permission matching degree is 0, that asset is directly eliminated. For assets with the same comprehensive score, semantic matching degree, business relevance, data quality score, and user preference score are compared sequentially until the order is determined.

[0064] It should be noted that the scores for each dimension in the above process can be determined in the following way: Semantic matching score, which can be obtained by calculating the cosine similarity between the user demand vector and the asset metadata vector; Business relevance is scored based on the length of the path between assets and needs in the business knowledge graph. The shorter the path, the higher the relevance score. Data quality: During implementation, based on system configuration, the system can automatically scan for missing values ​​in required fields of data assets, monitor the difference between the latest update time of data assets and the current time, and compare whether the field values ​​of data assets are consistent in different business systems. The data quality score is obtained by weighting the results of the three tests. In addition, external systems involved in data asset management (such as enterprise data platform / data asset system) will have data quality evaluation reports, which have corresponding scoring items for data tables. The scoring data can also be used to process the scores and normalize them to the [0,1] value range as the score for that dimension. Permission matching score: 1 for a complete match with user permissions, and 0 for no permissions; User preference score: A user preference recommendation score calculated using a collaborative filtering algorithm based on users' historical behavior data.

[0065] During the conflict resolution process, when content conflicts exist among multiple candidate assets, only the asset with the highest overall score is retained, while other low-scoring conflicting results are eliminated. Furthermore, assets with an overall score below a preset threshold (e.g., 0.2) are eliminated to reduce the interference of low-quality results on users.

[0066] This process sorts all remaining candidate assets after conflict resolution in descending order of their overall scores. If multiple assets have the same overall score, the final order is determined by a comparison rule (semantic matching priority). The final result set is a sorted list of assets, with each entry containing the asset name, metadata details, supplier information, overall score, and optional sub-scores.

[0067] In addition, to complete the final output and drive continuous system optimization, the five core steps of the core process also include step S5: result output and closed-loop feedback iteration. Step S5 includes: a: Output the final query result set, recommendation result set, and updated data storage and sorting results to the user through a visual interface, and simultaneously output the data asset provider information, metadata details, data lineage, permission description, application and usage process, and business usage guide; b: The collaborative scheduling and feedback agent receives user feedback information in real time, including result clicks, favorites, downloads, application for use, feedback, query corrections, and supplementary requests, while recording all user behavior data; c: Based on user feedback, the collaborative scheduling and feedback agent completes iterative optimization across the entire process. The iterative optimization here includes demand-side optimization, such as synchronizing demand correction data with the demand parsing agent to optimize the demand decomposition model and query text matching algorithm; synchronizing positive and negative samples with the demand reasoning agent to optimize the semantic matching model and knowledge graph; synchronizing behavioral data with the personalized recommendation agent to optimize the recommendation strategy, etc.; that is, the technical solution of this application also includes a closed-loop feedback iterative step: the collaborative scheduling and feedback agent receives feedback information from the demand side users on the final result set, and updates the demand decomposition model of the demand parsing agent, the semantic matching model and knowledge graph of the demand reasoning agent, and the recommendation strategy of the personalized recommendation agent according to the feedback information.

[0068] For example, in the above process, positive and negative samples can be obtained through two modes: automatic collection based on user behavior and periodic manual annotation collection. Periodic manual annotation collection involves business personnel manually labeling positive and negative samples based on historical user access logs and query interaction logs, generating standard positive and negative example datasets in batches, and periodically pushing them to the demand inference agent.

[0069] The automatic collection of user behavior data involves the system backend automatically collecting all user interactions, covering four categories: explicit positive, implicit positive, explicit negative, and implicit negative behaviors. Specifically: Explicit positive behaviors include clicking to view details of query results, saving data assets, applying to use or purchase data, and submitting a satisfaction rating; implicit positive behaviors include spending more time on the results page than a set threshold, repeatedly reviewing the same data asset, correcting the query based on the recommended results and completing a valid query, and finally completing a data transaction or business use. Explicit negative behaviors include clicking "unsatisfactory" and leaving negative reviews, providing inaccurate or irrelevant feedback results, manually correcting search criteria and re-searching, explicitly stating that the results do not meet the needs, and reporting data with errors, omissions, or sensitive issues; implicit negative behaviors include quickly exiting the results page, staying on the page for less than the threshold, repeatedly re-entering the query statement, and not clicking or interacting with the recommended results.

[0070] Based on this, the specific generation and synchronization mechanism for positive and negative samples can be as follows: For positive sample synchronization, when a user saves, downloads, or rates a query result higher than a preset threshold (e.g., 4 stars or above), or explicitly clicks "helpful" or "related" buttons on the interface, the collaborative scheduling and feedback agent sends the positive example pair consisting of the query request and the query result as a positive sample to the demand reasoning agent. The demand reasoning agent uses this positive example pair to fine-tune the model and enhance the semantic similarity between similar queries and similar assets. Conversely, for negative sample synchronization, when a user clicks "not interested" or "result is irrelevant" on a query result, or quickly skips it without any interaction (such as staying for less than a preset threshold), the collaborative scheduling and feedback agent will send the negative sample pair consisting of the query request and the corresponding query result to the demand reasoning agent. The demand reasoning agent will then use this negative sample pair to adjust the model parameters and reduce the matching weight of such irrelevant assets in future queries.

[0071] In addition, the pre-construction stage of the front-end process can also achieve supply-side optimization. That is, the intelligent agent built by the data element supply-side catalog can synchronize the demand side's query hotspots, demand pain points, and feedback suggestions, driving the supplier to optimize the data catalog content, supplement the business knowledge base, and improve metadata information, forming a two-way closed-loop optimization between supply and demand.

[0072] In some embodiments, this application also proposes a data element intelligent query system, which adopts a five-layer hierarchical architecture, with communication connections and collaborative linkage between each layer, including: The user-side layer provides the operation entry and result display. In other words, it provides a single query entry for users, enabling core functions such as natural language query request submission, query result visualization, and user feedback interaction. It is the only interface for users to interact with the system, eliminating the need for users to connect to multiple nodes and platforms of scattered query entry.

[0073] The service gateway layer communicates with the client layer and is used for requesting access, authentication, and encrypted transmission. As the unified access boundary of the system, the service gateway layer ensures end-to-end security for user access and data transmission.

[0074] The intelligent agent cluster layer, which communicates with the service gateway layer, is the core execution layer of the system. It includes multiple intelligent agent nodes, each of which operates independently and performs its own duties. Relying on a unified collaborative bus, it achieves cross-node collaborative linkage to jointly complete the entire process of user query requests.

[0075] The global storage layer communicates with the intelligent agent cluster layer and is used for distributed storage of standardized data catalogs, knowledge bases, metadata, knowledge graphs, and vector indexes, providing a standardized and highly reliable data foundation for system operation.

[0076] The heterogeneous data source interface layer communicates with the intelligent agent cluster layer to adapt to cross-node, multi-source heterogeneous data sources and complete query interface, realizing the standardized distribution of distributed query requests and the return processing of query results. These data sources include, but are not limited to, relational databases, data warehouses, data lakes, API interfaces, etc.

[0077] Among them, multiple intelligent agent nodes include at least: The demand parsing agent is used to receive natural language query requests from demand users, preprocess the query requests, break them down into multiple atomic subtasks, and construct a dependency graph of the atomic subtasks. Metadata management intelligent agent, used to pull accessible subsets of metadata from a standardized data catalog based on user permission scope and build real-time vector indexes for them; A demand reasoning agent is used to perform semantic matching and link reasoning based on real-time vector index and the knowledge base to generate a candidate data asset set. Multiple cross-domain and cross-source query execution agents are used to generate query plans based on candidate data asset sets, initiate queries to supplier data sources to obtain query results, or to adapt to different types of heterogeneous data sources, generate and execute distributed query plans, and complete query execution and result standardization processing across node data sources. The permission and security management intelligent agent is used to perform security verification and de-identification processing on query results. In some implementation scenarios, it is also used to verify permissions, identify sensitive data, de-identify data and audit operations in other processes, so as to achieve full-chain security and compliance management. The collaborative scheduling and feedback agent, as the core scheduling hub of the cluster, communicates with the demand parsing agent, metadata management agent, demand reasoning agent, cross-domain and cross-source query execution agents, and permission and security control agent. It is used to schedule the collaborative work of each agent according to the dependency graph, and to merge and resolve conflicts of the query results after security verification and desensitization, and generate the final result set to be output to the demand user.

[0078] Furthermore, as an optional configuration, the multiple intelligent agent nodes also include a directory building intelligent agent and a personalized recommendation intelligent agent, which respectively complete the automated construction of standardized data directories and business knowledge bases, as well as personalized and accurate recommendations of data assets.

[0079] Regarding the system product implementation examples, the specific functions, interaction logic, and workflow of each of the above-mentioned intelligent agents have been described in detail in the embodiments of the relevant methods, and will not be elaborated further here.

[0080] The following is an overall introduction to the technical solution of this application, using a typical application scenario as an example. This typical scenario is a cross-domain data element collaborative query scenario based on a trusted data space.

[0081] As a core infrastructure for the market-oriented circulation of data elements, the Trusted Data Space operates on the core principles of "data usable but invisible, controllable and measurable, and cross-domain mutual trust and traceability." It achieves secure sharing and collaborative application of data from multiple entities through a distributed architecture. Currently, the implementation of the Trusted Data Space faces two major bottlenecks: First, data assets are physically dispersed and logically isolated. Each data provider's core data assets are stored on independent nodes within its local domain, achieving limited connectivity with the Trusted Data Space only through standard interfaces. Demanders cannot complete one-stop collaborative queries across multiple domain nodes and heterogeneous data sources through a single trusted entry point; they must connect to multiple providers' independent systems, resulting in cumbersome query processes and low execution efficiency. Second, existing query solutions cannot adapt to the end-to-end trust requirements of the Trusted Data Space. Single large-model query solutions are prone to inference illusions and deviations in demand analysis. Traditional rule-based retrieval solutions cannot achieve deep and accurate matching between demands and data assets, and struggle to balance query accuracy, execution efficiency, and end-to-end compliance control. These solutions fail to meet the core demands of demanders for accurate, efficient, and secure cross-domain data queries within the Trusted Data Space scenario.

[0082] In this embodiment, the multi-agent collaborative data element intelligent query system of the present invention is used as the core query service hub of the trusted data space. Both the supply and demand sides complete the access and use through the trusted data space standard connector to adapt to the technical architecture and compliance requirements of the trusted data space and solve the above-mentioned industry pain points.

[0083] In this embodiment, all participating entities complete trusted access through the standard connector of the consumer industry trusted data space. The connector serves as the sole trusted interaction channel between each entity's local system and the trusted data space, realizing a closed-loop end-to-end system encompassing distributed trusted identity (DID) authentication, standardized and secure reporting of metadata, encrypted transmission of query commands, execution of local privacy-preserving computations, de-identified return of compliance results, and audit and evidence storage of the entire operation process. The core participating entities and access methods are as follows: Trusted Data Space Operator: Deploys the multi-agent collaborative data element intelligent query system of this invention as the core query service hub of the trusted data space in the consumer industry. It is responsible for the operation and maintenance of the multi-agent cluster, the formulation of industry standard rules, the full-link trusted audit and operation control, and ensures that the entire process complies with the compliance requirements of data circulation in the consumer industry.

[0084] Data providers are all compliant data holders in the consumer industry, including: leading comprehensive e-commerce platforms (public domain consumer transaction data), brand offline chain stores (in-store consumption and membership data), compliant third-party payment institutions (user payment behavior data), leading logistics service providers (order fulfillment and delivery data), and compliant third-party consumer insight agencies (consumer category preference and industry tag data). Each provider completes a trusted connection between its local system and the trusted data space through a trusted data space standard connector. User raw data is stored entirely on independent nodes within the provider's local domain. The connector is used only for standardized metadata reporting, trusted identity authentication, secure reception of query commands, local compliant query execution, and anonymized result feedback, strictly adhering to the core principle of "data usable but not visible."

[0085] Data demander: A leading domestic trendy beauty and new consumer brand connects to the system's single-entity trusted query portal through the Trusted Data Space Standard Connector. This allows for distributed trusted identity authentication, natural language query request submission, query result acquisition, and business feedback. There is no need to connect to the independent business systems of various data providers. The system can complete full-domain compliant data queries and user insights across multiple entities and nodes through this single portal.

[0086] The core business requirement of the data client is as follows: The beauty brand plans to conduct targeted marketing campaigns for the launch of new products. For example, it needs to obtain reliable data on "female consumers aged 18-35 in this province who have beauty product consumption records in the past 6 months, have an average monthly beauty product consumption of 800-3000 yuan, prefer domestic trendy ingredients and functional products, have offline beauty store experience records, and are active on short video content platforms". At the same time, it needs to obtain comprehensive insight data on the consumption preferences, channel reach habits, repurchase cycle, and price sensitivity of this group for precise audience selection and customized marketing content for new products.

[0087] The solution implemented using the technical solution of this invention achieves trusted interaction between the supply and demand parties throughout the entire process through the Trusted Data Space Standard Connector. The specific implementation steps are as follows: Step 1: The supplier completes the construction of trusted access and standardized data infrastructure through connectors. This step corresponds to the preliminary implementation process of this invention, building a compliant and standardized basic data foundation for the core query service. The entire process utilizes a trusted data space connector to achieve trusted interaction between the supplier's domain system and the query system. The specific implementation is as follows: Each data provider submits data source authorization documents, asset listing applications, and data security compliance rule configurations to the system through the Trusted Data Space Standard Connector, completing distributed trusted identity (DID) authentication and access permission verification, and simultaneously completing the compliance classification and filing of data assets.

[0088] The collaborative scheduling and feedback intelligent agent constructs the data catalog and connects to heterogeneous data sources within various supplier domains through connectors. It completes an automatic scan of all compliant data assets and securely transmits standardized technical metadata, data update frequency, asset ownership information, and compliant usage scope back through connectors to form the original data asset set. Simultaneously, it triggers the permission and security control intelligent agent to complete the legality audit of data sources, compliance verification of personal information, and initial identification of sensitive data, blocking access from illegal and non-compliant data sources.

[0089] The data catalog building agent standardizes and cleans the original metadata, converts it into a unified format, and completes the precise association mapping between business semantic annotation, user tags and technical metadata based on the unified business terminology library of the consumer industry and the knowledge graph of the beauty industry. It automatically builds a business knowledge base exclusive to the beauty industry and completes the semantic alignment of cross-supplier data.

[0090] After each supplier confirms the listing of assets, the metadata management intelligent agent completes the full import of standardized data catalogs and business knowledge bases. Based on the finely tuned embedding model of the beauty industry, a distributed vector index is built and synchronized to the demand reasoning intelligent agent to officially provide query services to demanders. Throughout the process, the original user data is stored in the local nodes of the supplier, and only the metadata is reliably reported through the connector, without any original personal information being transferred across domains.

[0091] Step 2: The customer connects to a single entry point via a connector to complete the requirement submission and system initialization. The demand-side beauty brand accesses the single entity trusted query portal of this system through the Trusted Data Space Standard Connector, submits the above natural language query request, and simultaneously completes the distributed trusted identity (DID) secondary authentication. The system collects the demand-side business context, historical interaction records, and data access compliance permission scope through the connector (only aggregated and de-identified data for marketing insights can be obtained, and user-identifiable personal information cannot be obtained).

[0092] The collaborative scheduling and feedback agent completes system initialization, triggers the permission and security control agent to complete demander identity verification and compliance permission verification, pulls the range of supplier directories accessible to the demander and data usage compliance rules, completes the allocation of computing resources and communication link initialization of the multi-agent cluster, and intercepts untrusted and unauthorized access requests.

[0093] Step 3: Requirements Analysis and Atomic Task Decomposition The collaborative scheduling and feedback agent schedules the request parsing agent, completing the full parsing of query requests and standardized task decomposition. The specific implementation is as follows: The demand analysis intelligent agent performs text correction, stop word filtering, and business entity extraction on natural language query requests. It accurately identifies core business entities and user tags such as "women aged 18-35, local province, beauty consumption in the past 6 months, average monthly consumption of 800-3000 yuan, preference for domestic trend ingredients, offline store visit records, and activity on short video platforms". It completes user intent recognition and standardized mapping of consumer industry terminology, and clarifies that the core demand of the demand side is a hybrid demand of "precise audience selection + full-domain consumption insight".

[0094] After completing a fuzzy quantitative assessment of the requirements and confirming that the requirements are semantically complete and unambiguous, the complex full-domain query requirements are broken down into six independent atomic subtasks: basic demographic attribute filtering, consumer behavior dimension filtering, product preference feature matching, channel reach habit analysis, user insight data aggregation, and compliance verification. At the same time, a dependency graph of the atomic subtasks is constructed to clarify the execution order and parallel execution rules of each subtask, and the results are reported to the collaborative scheduling and feedback intelligent agent.

[0095] Step 4: Multi-agent collaboration, completing cross-node collaborative reasoning and query execution through connectors. The collaborative scheduling and feedback agents are based on the dependency graph of atomic subtasks. The agents are scheduled to complete cross-domain collaborative reasoning and distributed query execution through the trusted data space standard connector. The user's original data does not leave the local node of the supplier. The interaction of instructions and compliance results is only completed through the connector. The specific process is as follows: Pre-access filtering and real-time vector indexing of metadata: The metadata management agent is scheduled to pull a subset of accessible metadata from five types of cross-domain nodes of the supplier based on the compliant permission scope of the demander. User detail data metadata that the demander does not have permission to access is filtered out, and a real-time vector index is built for the subset of accessible metadata, which is synchronized to the demand inference agent. The permission and security control agent simultaneously completes the permission pre-verification of metadata, and intercepts metadata with excessive permissions from entering the subsequent process, so as to achieve pre-access control of security and compliance.

[0096] Deep semantic matching and link reasoning: The scheduling demand reasoning intelligent agent, based on real-time vector index, beauty industry business knowledge base and global consumer knowledge graph, completes the semantic similarity calculation of user demand and metadata, and completes multi-hop link reasoning by combining user behavior links, tag association relationships and upstream and downstream of consumption scenarios. It mines compliant data assets that are deeply matched with the demand side's marketing needs, eliminates invalid assets that conflict with the demand business logic, and finally generates a candidate data asset set corresponding to 5 types of suppliers, which is reported to the collaborative scheduling and feedback intelligent agent.

[0097] Parallel execution of cross-node distributed queries: The collaborative scheduling and feedback agent matches the corresponding cross-domain and cross-source query execution agent according to the supplier node and data source type corresponding to the candidate data asset set, and initiates parallel scheduling; each query execution agent issues a standardized and compliant distributed query execution plan to the corresponding supplier domain node through the trusted data space standard connector, specifying that only aggregated and de-identified statistical results and population tag distribution can be returned, and user-identifiable original personal information cannot be returned; after receiving the instruction, the supplier connector completes privacy calculation and query execution in its local domain, and only sends back the compliant and non-identifiable standardized query results to the corresponding query execution agent through the connector; each query execution agent completes the format standardization and internal aggregation of the returned results, generates a standardized intermediate result set, and reports it to the collaborative scheduling and feedback agent.

[0098] End-to-end security and compliance verification: The scheduling authority and security management intelligent agent performs full compliance verification on all intermediate result sets, including secondary review of permissions, secondary detection of user identifiable information, verification of data anonymization effectiveness, and verification of compliance rules for data circulation in the consumer industry. The output results have privacy protection capabilities to ensure personal information security. At the same time, it generates a full-process operation audit log and synchronizes it to the blockchain trusted evidence storage system in the trusted data space to achieve end-to-end traceability and auditability, and intercepts unauthorized operations and non-compliant requests in real time.

[0099] Step 5: Multi-source result fusion, conflict resolution, and reliable output The collaborative scheduling and feedback intelligent agent performs merging, deduplication, and format standardization on the multi-supplier intermediate result set after compliance processing. Based on preset weight rules (semantic matching degree 40%, business relevance 20%, data quality 15%, permission matching degree 15%, user preference 10%), it completes comprehensive scoring calculation and conflict resolution, removes low-quality and low-matching result data, and completes compliance aggregation of cross-supplier data based on unified user identifiers to generate the final audience selection results and full-domain insight dataset.

[0100] The system provides a single, trusted query portal to the client through a trusted data space connector, which then visualizes the final result set. It simultaneously displays the data source provider, data update time, scope of compliant data use, authorized usage period, and business application guidelines. It also outputs a user tagging system and insightful conclusions adapted to marketing campaigns, and supports clients in saving results, applying for compliant use, and providing feedback on results.

[0101] Step 6: Closed-loop feedback and bidirectional iterative optimization The collaborative scheduling and feedback agent receives marketing campaign performance feedback and full behavioral data from demanders through connectors, completing a two-way closed-loop iterative optimization of supply and demand: it synchronizes demand correction and performance data with the demand parsing agent and demand reasoning agent, optimizing the demand decomposition model, semantic matching model, and beauty industry knowledge graph; it synchronizes query hotspots, demand pain points, and business feedback with the data catalog building agent, driving each supplier to optimize data catalog content, business semantic annotation, and industry knowledge base coverage through connectors, forming a full-link closed loop of "supply-demand-feedback-optimization", continuously improving the system's query accuracy and business adaptability in consumer industry scenarios.

[0102] In this embodiment, the deep integration of this invention with the trusted data space of the consumer industry achieves a triple realization of technological value, business value, and compliance value, with the following specific effects: It enables one-stop collaborative querying of cross-domain and distributed node data through a single entry point. The demand side can complete cross-source data query and full-domain insight of 5 types of suppliers and multiple independent nodes by accessing a single query entry point through the trusted data space connector. There is no need to connect to the independent business systems of each supplier, and no need for multi-platform and multi-round operations. The query execution efficiency is significantly improved compared with traditional distributed queries, which greatly reduces the threshold for consumer brands to obtain full-domain data.

[0103] Through specialized collaborative scheduling of multiple agents, a deep understanding of needs and precise matching of results are achieved. The complex process of marketing audience selection and needs insight is broken down into standardized atomic tasks. Through the specialized execution of each agent, the illusion problem of long-chain reasoning of a single large model is significantly reduced, the accuracy of needs analysis is significantly improved, and the matching degree between audience selection and insights and brand marketing needs is greatly improved compared with traditional solutions, effectively helping to improve the ROI of brand new product marketing campaigns.

[0104] A full-chain compliance management system adapted to the consumer industry has been built. Through the trusted data space connector, the system achieves full-process trustworthiness in identity, transmission, execution, and auditing. It strictly adheres to the core principle of "data is usable but not visible". User's original data never leaves the supplier's local domain, and there is no risk of cross-domain transfer of personal information. The entire data processing process has comprehensive security management and privacy protection capabilities, meeting the management requirements for compliant data flow and secure use, and achieving a balance between query efficiency, accuracy, and privacy compliance.

[0105] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A data element intelligent query method based on multi-agent collaboration, characterized in that, This system is applied to an intelligent data element query system. The system comprises several intelligent agents connected to a pre-built standardized data catalog and knowledge base. These intelligent agents include at least a demand analysis agent, a demand reasoning agent, a collaborative scheduling and feedback agent, a metadata management agent, multiple cross-domain and cross-source query execution agents, and an authorization and security control agent. Each supplier completes a trusted connection between its local system and the trusted data space through a trusted data space standard connector. User's original data is stored entirely on independent nodes within the supplier's local domain. Metadata standardization reporting, trusted identity authentication, secure reception of query commands, local compliant query execution, and anonymized result feedback are only completed through the connector. The method includes: The demand parsing agent receives natural language query requests from demand users, preprocesses the query requests, breaks them down into multiple atomic subtasks, and constructs a dependency graph of the atomic subtasks. The collaborative scheduling and feedback agent, based on the dependency graph, schedules the metadata management agent to pull an accessible subset of metadata from the standardized data directory based on user permission scope and build a real-time vector index for it. The agent also schedules the demand reasoning agent to perform semantic matching and link reasoning based on the real-time vector index and the knowledge base to generate a candidate data asset set. Finally, the agent schedules the corresponding cross-domain and cross-source query execution agent to generate a query plan based on the candidate data asset set, initiate a query to the supplier data source, and obtain the query results. The permission and security management intelligent agent performs security verification and desensitization processing on the query results; The collaborative scheduling and feedback agent integrates and resolves conflicts in the processed query results, generates a final result set, and outputs it to the demand user. Specifically, for query requests with ambiguous needs or missing information, the demand parsing agent generates interactive clarification scripts and initiates supplementary inquiries to the user until the query request can be completely decomposed into the atomic subtasks. This process includes: The requirement parsing agent, combined with the business entity extraction results, calculates the degree of ambiguity of the extracted non-standardized terminology set. Using the AHP-fuzzy comprehensive evaluation model, it quantifies the fuzziness level of the requirement from four dimensions: terminology ambiguity, semantic completeness, entity coverage, and business matching. The requirement is divided into three fuzziness levels: low, medium, and high, and corresponding processing strategies are adopted: For low fuzziness, it directly enters the atomic task decomposition process; for medium fuzziness, it supplements default parameters based on user history and business context, generates alternative requirement solutions, and simultaneously sends a one-click confirmation prompt to the user; for high fuzziness, it automatically generates interactive clarification scripts, initiates supplementary inquiries to the user, until complete requirement information is obtained, ensuring that the query request can be completely decomposed. The collaborative scheduling and feedback agent schedules the demand reasoning agent to perform semantic matching and link reasoning based on the real-time vector index and the knowledge base, generating a candidate data asset set, specifically as follows: Based on the obtained real-time vector index, the demand reasoning agent is invoked. The demand reasoning agent performs semantic similarity calculation on the real-time vector index based on the decomposed atomic tasks, standardized business terms, business knowledge bases built by suppliers, and global business knowledge graphs. At the same time, combined with query text matching data, it completes link reasoning based on data lineage and business domain associations, mines upstream and downstream supplier data assets related to user needs, generates a preliminary candidate data asset set, and performs business logic consistency verification on the candidate set to eliminate assets that conflict with the demand business logic.

2. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, The preprocessing includes at least text correction, stop word filtering, and business entity extraction. The atomic subtasks include at least the metadata retrieval subtask, the semantic reasoning subtask, the cross-source data query subtask, the personalized recommendation subtask, and the compliance verification subtask.

3. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, The scope of user permissions is obtained through the following steps: The permission and security control intelligent agent verifies the identity of the demand user, determines the accessible directory range of the demand user based on the verification result, and synchronizes the accessible directory range as the user permission range to the metadata management intelligent agent.

4. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, When the cross-domain, cross-source query execution agent generates a query plan, it performs syntax conversion and execution path optimization based on the database type of the supplier data source; after initiating a query to the supplier data source, it receives the returned original query results and formats the original query results into a standardized query result set.

5. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, When the candidate data asset set involves multiple supplier data sources, the collaborative scheduling and feedback agent schedules multiple corresponding cross-domain and cross-source query execution agents to initiate queries to the multiple supplier data sources in parallel. Each cross-domain and cross-source query execution agent internally aggregates its own query results and sends them to the collaborative scheduling and feedback agent. The collaborative scheduling and feedback agent summarizes the received multiple aggregated results to obtain a summarized query result set.

6. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, The collaborative scheduling and feedback agent integrates and resolves conflicts in the query results. The integration includes deduplication and format standardization of multiple query results. The conflict resolution is performed based on preset weight rules, which include semantic matching degree, business relevance, data quality, permission matching degree, and / or user preference.

7. The intelligent data element query method based on multi-agent collaboration according to claim 1, characterized in that, The aforementioned intelligent agents also include personalized recommendation intelligent agents; The collaborative scheduling and feedback agent also schedules the personalized recommendation agent to generate a supplementary recommendation asset set based on user profiles, historical interaction behaviors, business context, and the candidate data asset set. The permission and security control intelligent agent performs security verification and desensitization processing on the supplementary recommended asset set; The collaborative scheduling and feedback agent integrates the processed supplementary recommended asset set with the query results and resolves conflicts to generate the final result set.

8. The intelligent data element query method based on multi-agent collaboration according to claim 7, characterized in that, It also includes a closed-loop feedback iteration step; the closed-loop feedback iteration step includes: The collaborative scheduling and feedback agent receives feedback information from the demand user regarding the final result set, and updates the demand decomposition model of the demand parsing agent, the semantic matching model and knowledge graph of the demand reasoning agent, and the recommendation strategy of the personalized recommendation agent based on the feedback information.

9. A data element intelligent query system, used to implement the data element intelligent query method of claim 1, comprising: The user-side layer is used to provide operation entry points and result display; The service gateway layer communicates with the client layer and is used for access requests, authentication, and encrypted transmission. The intelligent agent cluster layer communicates with the service gateway layer and includes multiple intelligent agent nodes; The global storage layer, which communicates with the intelligent agent cluster layer, is used for distributed storage of standardized data catalogs, knowledge bases, metadata, knowledge graphs, and vector indexes. A heterogeneous data source interface layer, communicatively connected to the intelligent agent cluster layer, is used to adapt to heterogeneous data sources and complete query integration; characterized in that the plurality of intelligent agent nodes include at least: The demand parsing agent is used to receive natural language query requests from demand users, preprocess the query requests, break them down into multiple atomic subtasks, and construct a dependency graph of the atomic subtasks. A metadata management agent is used to pull an accessible subset of metadata from the standardized data directory based on user permission scope and build a real-time vector index for it; A demand reasoning agent is used to perform semantic matching and link reasoning based on the real-time vector index and the knowledge base to generate a candidate data asset set. Multiple cross-domain and cross-source query execution agents are used to generate query plans based on the candidate data asset set and initiate queries to the supplier data source to obtain query results. An intelligent agent for permission and security management is used to perform security verification and desensitization processing on the query results; The collaborative scheduling and feedback agent is connected to the demand parsing agent, the metadata management agent, the demand reasoning agent, the cross-domain and cross-source query execution agents, and the permission and security control agent, respectively. It is used to schedule the collaborative work of each agent according to the dependency graph, and to merge and resolve the query results after security verification and desensitization, and generate a final result set to be output to the demand user.

Images