A blockchain and agent-based on-chain directory retrieval and supply-demand matching method

Abstract

The application relates to the field of blockchains and data circulation, and discloses a chain-based directory retrieval and supply-demand matching method based on a blockchain and an intelligent agent, which is applied to a supply-demand matching intelligent agent, the method calls a large language model to generate semantic enhanced metadata for directory metadata and write the semantic enhanced metadata into a vector index library, converts a natural language retrieval request into a structured hard constraint condition set and a demand semantic feature vector, filters and calculates the similarity according to the hard constraint to obtain a rough screening candidate set, calculates a dynamic authority score in combination with a directed bipartite graph constructed by a blockchain shared approval log, fuses the score to obtain a fine sorting candidate directory list, if a rejection instruction is received, abstracts system flow into a directed acyclic graph state machine, combines a constraint increment to perform pathfinding calculation to output an optimal rollback path for reoperation, and converts an algorithm residual error into a negative constraint edge to be supplemented to the directed bipartite graph, and the application effectively improves the accuracy of supply-demand matching and the autonomous error correction and optimization capability of the system.

Description

Technical Field

[0001] This invention relates to the field of blockchain and data circulation, and in particular to an on-chain directory retrieval and supply-demand matching method based on blockchain and smart agents. Background Technology

[0002] With the advancement of digital transformation across industries, the demand for cross-departmental and cross-organizational data sharing is increasing. Data catalogs are the foundation for data discovery and sharing. Currently, in cross-node data sharing networks, achieving supply and demand matching between data providers and users still faces numerous technical bottlenecks.

[0003] First, existing data catalog retrieval typically relies on exact keyword matching or shallow text analysis. Because data providers often use standard structured fields when registering their catalogs, while data users frequently use natural language tailored to their specific business scenarios when initiating searches, there are objective differences in the textual expressions between the two parties. This literal matching approach struggles to connect simple structured fields with complex business logic, causing the retrieval system to fail to accurately identify the deep semantics of the data, resulting in low accuracy of search results.

[0004] Secondly, in the ranking stage of directory retrieval, most existing supply and demand matching mechanisms use the similarity between the search text and the directory text as the sole evaluation metric. This processing logic ignores the actual flow of data within the shared network. In real-world business scenarios, while some data may have a high degree of textual description relevance, it may suffer from issues such as low historical call frequency and poor data quality feedback. Existing systems, unable to effectively extract and quantify historical call frequency and evaluation data from shared approval logs, cause directory ranking to detach from the objective business context, thus limiting the reliability of the matching results.

[0005] Finally, most existing data matching algorithms are static, one-way retrieval architectures, lacking the ability to autonomously correct errors in the face of abnormal matching situations. In practice, when the matched data directory is rejected by the data user due to insufficient granularity or mismatched source, the existing system usually just terminates the process. The system cannot extract the constraint deviation features hidden behind these rejection instructions, nor can it use these failed matching records to adjust the underlying network topology and weight allocation. This causes the system to repeatedly execute the failed matching path when faced with similar retrieval needs in the future, making it difficult for the matching algorithm to iteratively optimize based on real supply and demand feedback. Summary of the Invention

[0006] To overcome the above shortcomings, this invention provides an on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents. It aims to improve the existing data directory retrieval and supply-demand matching mechanisms, which suffer from low retrieval accuracy due to differences in semantic expression of supply and demand, limited reliability of results due to the single ranking index ignoring historical call performance, and lack of autonomous error correction and iterative optimization capabilities due to the inability to effectively utilize rejection feedback data.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents, applied to a supply-demand matching intelligent agent communicating with a blockchain network, data provider nodes, and data user nodes; the method includes:

[0008] Obtain the directory metadata sent to the blockchain network by the data provider node, call the large language model to perform logical inference based on the directory metadata to generate semantically enhanced metadata, perform vectorization mapping processing on the semantically enhanced metadata to obtain the directory semantic vector, and write the directory semantic vector into the configured vector index library; Receive the natural language retrieval request sent by the data user node, extract the entity parameters inside the natural language retrieval request, generate a set of structured hard constraints based on the entity parameters, perform text cleaning operation on the natural language retrieval request to generate cleaned retrieval text, and convert the cleaned retrieval text into a requirement semantic feature vector. Based on the structured hard constraint set, a data filtering operation is performed on the vector index library to obtain a compliant directory subset. The similarity value between the demand semantic feature vector and the directory semantic vector in the compliant directory subset is calculated to select a coarse candidate set. The shared approval logs stored in the blockchain network are obtained. A directed bipartite graph is constructed based on the shared approval logs. The dynamic authority score of the corresponding data nodes within the coarse candidate set is calculated based on the directed bipartite graph. The similarity value and the dynamic authority score are merged for secondary sorting and a fine-ranked candidate directory list is output. Based on the refined ranking candidate directory list, a matching explanation text is generated, and the refined ranking candidate directory list and the matching explanation text are sent to the data user node. When the refined ranking candidate directory list is generated and sent for the first time, the forward operation execution path record when the constraint increment is not loaded for the first time is cached as the initial optimal solution. If a rejection instruction carrying feedback reason text data is received from the data user node, the system calculation steps and system data flow process within the supply and demand matching agent are mapped to nodes and edges respectively and abstracted into a directed acyclic graph state machine. The rejection instruction is converted into a constraint increment, the feedback reason text data is parsed to obtain constraint conflict deviation data, the topological authority gradient of the data node in the directed bipartite graph is calculated and the topological authority gradient is used as a damping parameter. The damping parameter and the constraint increment are combined to perform pathfinding calculation in the directed acyclic graph state machine to output the optimal rollback path. The calculation is re-executed along the optimal rollback path. At the same time, the algorithm residual generated by the mapping comparison between the initial optimal solution and the constraint conflict deviation data is extracted, and the algorithm residual is converted into a negative constraint edge with numerical weight and added to the directed bipartite graph.

[0009] As a further description of the above technical solution: The directory metadata contains raw structured fields, which contain text character data. The step of calling a large language model to perform logical inference based on the directory metadata to generate semantically enhanced metadata includes: Extract the text character data contained in the original structured fields; Configure a preset text analysis instruction, import the text character data into the preset text analysis instruction, and transmit the preset text analysis instruction to the large language model; The large language model combines an external domain knowledge base to perform natural language reasoning operations and outputs business scenario description text, application purpose description text, related data description text, and data granularity description text. The semantically enhanced metadata is generated by summarizing and concatenating the original structured fields with the business scenario description text, the application purpose description text, the associated data description text, and the data granularity description text, and then generating the semantically enhanced metadata based on the summarized and concatenated overall text data.

[0010] As a further description of the above technical solution: The step of extracting entity parameters from the natural language retrieval request and generating a set of structured hard constraint conditions based on the entity parameters includes: The cleaned search text is input into a pre-configured named entity recognition model, and the named entity recognition model is used to extract the entity parameters contained in the cleaned search text. The cleaned and retrieved text is input into a pre-configured intent classification model, and the intent classification model is used to perform feature mapping and classification probability calculation operations to output intent category labels; The rule mapping process is performed based on the entity parameters to generate explicit constraint rule parameters, and the explicit constraint rule parameters are then concatenated and merged to generate the structured hard constraint condition set.

[0011] As a further description of the above technical solution: The vector index library is configured with unique identifiers. The step of performing data filtering operations on the vector index library based on the structured hard constraint condition set to obtain a compliant directory subset, and calculating similarity values ​​to screen out a coarse candidate set includes: The explicit constraint rule parameters contained in the structured hard constraint condition set are parsed, and Boolean condition matching calculations are performed in the vector index library based on the explicit constraint rule parameters to extract the directory semantic vectors that conform to the explicit constraint rule parameters and their corresponding unique identifiers to generate the compliant directory subset. Based on the intent category label, perform a mapping query operation in a pre-configured category threshold mapping table to output the dynamic recall quantity; A cosine similarity calculation operation is performed on the semantic feature vector of the requirement and the semantic vector of each directory in the compliant directory subset to output the similarity value corresponding to each directory semantic vector; The directory semantic vectors are sorted in descending order based on the similarity values, and the directory semantic vectors that are within the range of the dynamic recall are extracted and associated with unique identifiers in combination with the dynamic recall number to generate the coarse candidate set.

[0012] As a further description of the above technical solution: The shared approval log contains historical data supply and demand matching records. These records include data provider node identifiers, data user node identifiers, approval completion timestamp parameters, and feedback evaluation score parameters. The step of constructing a directed bipartite graph based on the shared approval log and calculating a dynamic authority score based on the directed bipartite graph includes: The data provider node identifier is mapped and configured as a supply-side network graph node, and the data user node identifier is mapped and configured as a demand-side network graph node. Directed connection edges are constructed based on the flow and call relationship, and the initial connection weights are calculated and assigned to the directed connection edges in combination with the feedback evaluation score parameters, and then the directed bipartite graph is generated. Based on the unique identifier contained in the coarse candidate set, the corresponding supply-side network graph node is matched and located in the directed bipartite graph. The directed connection edges pointing to the supply-side network graph node are extracted to generate a connection edge set, and the number of edges contained in the connection edge set is counted. For each historical flow call, the sum of the in-degree and out-degree of the corresponding demand-side network graph node in the directed bipartite graph is extracted as the node centrality parameter, and the numerical difference between the current system timestamp and the corresponding approval completion timestamp parameter is calculated as the time difference parameter. Extract the preset basic functional weight parameters, time decay constant, and adjustment coefficient. Multiply the node centrality parameter corresponding to each call by the exponential decay term generated based on the product of the time decay constant and the time difference parameter. Summate the multiple values ​​obtained after multiplication. Multiply the sum by the adjustment coefficient to obtain the adjustment compensation value. Add the basic functional weight parameters and the adjustment compensation value to obtain the final dynamic authority score.

[0013] As a further description of the above technical solution: The step of fusing the similarity value and the dynamic authority score for secondary sorting and outputting a refined candidate directory list includes: The cleaned retrieved text is merged with the semantically enhanced metadata to generate a concatenated text sequence. The concatenated text sequence is then input into a pre-configured cross-encoder model for dimensionality reduction mapping to output a deep semantic matching score. Based on the intent category label, a matching query operation is performed in the preset weight configuration matrix, and the vector recall weight coefficient, semantic ranking weight coefficient, and authority weight coefficient are output. The first product is obtained by multiplying the similarity value by the vector recall weight coefficient, the second product is obtained by multiplying the deep semantic matching score by the semantic ranking weight coefficient, and the third product is obtained by multiplying the dynamic authority score by the authority weight coefficient. Perform a linear weighted summation operation on the first product, the second product, and the third product to generate a comprehensive re-ranking score, and integrate the unique identifiers in the coarse candidate set in descending order based on the comprehensive re-ranking score to generate the fine-ranking candidate directory list.

[0014] As a further description of the above technical solution: The step of converting the rejection instruction into a constraint increment includes: The feedback reason text data is input into a pre-configured feedback reason classification model to perform text classification probability calculation to output reason category labels. The reason category labels include semantic deviation labels, insufficient granularity labels, missing data item labels, and source mismatch labels. Based on the cause category label, a mapping query operation is performed in the pre-configured constraint increment transformation matrix, and the mapping output includes the constraint increment containing the corresponding bias adjustment parameter or filtering parameter; Specifically, when the cause category label is an insufficient granularity label or a missing data item label, the constraint increment includes additional hard boundary filtering parameters to guide the optimal rollback path to trigger data filtering operations; when the cause category label is a source mismatch label, the constraint increment includes vector space bias adjustment parameters for the authority weight coefficient.

[0015] As a further description of the above technical solution: The directed acyclic graph state machine includes state graph nodes and state graph connecting edges. The step of performing pathfinding calculations in the directed acyclic graph state machine by combining the damping parameters and the constraint increments to output the optimal rollback path includes: Based on the estimated computational load of the state diagram nodes, extract the system computing power overhead value corresponding to the system operation, and configure a cost balance factor vector including a first cost balance factor, a second cost balance factor, and a penalty coefficient. The first computational cost is obtained by multiplying the system computing power cost value by the first cost balance factor; the second computational cost is obtained by multiplying the exponential decay term, which has the natural constant as the base and the negative number of the damping parameter as the exponent, by the second cost balance factor; and the third computational cost is obtained by multiplying the additional penalty term, which is converted from the constraint increment, by the penalty coefficient. The pathfinding cost is calculated by summing the first calculated cost value, the second calculated cost value, and the third calculated cost value. Combining the constraint increment with the preset optimization objective solution conditions, a node traversal pathfinding operation is performed inside the directed acyclic graph state machine. Based on the pathfinding substitution value after incorporating the additional penalty term, the flow path with the minimum substitution value is extracted, and the optimal rollback path containing the entry calculation node identifier is generated.

[0016] As a further description of the above technical solution: The step of extracting the algorithm residual and converting it into negative constraint edges includes: The pre-configured text embedding model is invoked to perform feature encoding and dimension interpolation truncation operations on the constraint conflict deviation data to generate deviation feature vectors; The state machine node sequence features corresponding to the initial optimal solution are extracted using a graph attention network model or a random walk algorithm, and then aggregated into a one-dimensional feature vector through a pooling layer. The one-dimensional feature vector and the deviation feature vector are subjected to probability normalization, and the degree of deviation of the probability distribution is calculated using the cross-entropy loss function to output the residual of the algorithm. Configure a reverse penalty mapping rule, extract the data provider node identifier that generated the rejection instruction as the topology connection start point, extract the data provider node identifier corresponding to the rejected directory metadata as the topology connection end point, construct a connection edge between the topology connection start point and the topology connection end point as the negative constraint edge, and use the algorithm residual after numerical normalization as the negative weight parameter of the negative constraint edge.

[0017] As a further description of the above technical solution: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements an on-chain directory retrieval and supply-demand matching method based on blockchain and smart agents.

[0018] The present invention has the following beneficial effects: 1. This invention utilizes a semantic enhancement and indexing module to invoke a large language model to perform logical inference based on directory metadata to generate semantically enhanced metadata. A text embedding model is then invoked to perform vectorization mapping on the semantically enhanced metadata, resulting in a directory semantic vector. Simultaneously, a text cleaning operation is performed on the natural language retrieval request to generate cleaned retrieval text, and the text embedding model is invoked to transform the cleaned retrieval text into a request semantic feature vector. This processing method unifies the mapping of directory information and user needs into a high-dimensional semantic space, resolving the discrepancy in textual expression between data providers and users, and improving the accuracy of retrieval results.

[0019] 2. This invention constructs a directed bipartite graph based on the node call relationships recorded in the shared approval log through a dynamic topology calculation and hierarchical retrieval module. It then calculates the dynamic authority score of corresponding data nodes within the coarsely screened candidate set based on the directed bipartite graph, and performs a secondary sorting by fusing similarity scores with the dynamic authority scores, outputting a finely ranked candidate directory list. This mechanism quantifies objective indicators such as the historical call frequency and evaluation data of both supply and demand sides into feature scores, making directory sorting no longer solely dependent on text similarity, effectively improving the reliability of supply and demand matching results in real-world business scenarios.

[0020] 3. Upon receiving a rejection instruction from a data user node, this invention transforms the rejection instruction into a constraint increment. Combining the damping parameters and the constraint increment, pathfinding calculations are performed in a directed acyclic graph state machine, outputting the optimal rollback path. Subsequently, the system extracts the algorithm residual generated by the mapping comparison between the initial optimal solution and the constraint conflict deviation data. This residual is transformed into negative constraint edges with numerical weights, and these negative constraint edges are added to the directed bipartite graph. This mechanism enables the system to automatically find the computation node with the lowest cost for recalculation when encountering demand mismatches, and adjusts the network topology environment for subsequent searches by recording negative weights, achieving autonomous error correction and iterative optimization of the matching algorithm. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the internal modules and model configuration of the supply and demand matching intelligent agent of the present invention; Figure 2 This is a schematic diagram illustrating the specific process of data matching processing performed by the semantic enhancement and indexing module of the present invention; Figure 3 This is a flowchart of the dimensionality reduction and feature extraction module required by the present invention; Figure 4 This is a flowchart of the dynamic topology calculation, hierarchical retrieval, and reordering process of this invention; Figure 5 This is a flowchart illustrating the interactive explanation and confirmation / rejection branch processing of this invention; Figure 6 This is a schematic diagram illustrating the directed acyclic graph state machine, optimal rollback path finding, and penalty calculation of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1 - Figure 6 This invention provides an on-chain directory retrieval and supply-demand matching system based on blockchain and intelligent agents, including a blockchain network, data provider nodes, data user nodes, and supply-demand matching intelligent agents. The data user nodes are communicatively connected to input devices.

[0024] The blockchain network communicates with data provider nodes and data user nodes. Data provider nodes send directory metadata to the blockchain network. The blockchain network stores the directory metadata and sharing approval logs. Data user nodes connect to the supply and demand matching intelligent agent. Data user nodes send natural language retrieval requests, confirmation feedback signals, or rejection instructions with text data carrying feedback reasons to the supply and demand matching intelligent agent. Data user nodes submit sharing application forms to the blockchain network.

[0025] The supply-demand matching agent includes a semantic enhancement and indexing module, a demand dimensionality reduction and feature extraction module, a dynamic topology calculation and hierarchical retrieval module, and an interactive interpretation and rollback scheduling module. The semantic enhancement and indexing module has embedded or communicatively connected large language models (e.g., pre-trained language models based on the Transformer architecture) and text embedding models. The demand dimensionality reduction and feature extraction module is equipped with named entity recognition models and intent classification models (e.g., using classifier architectures such as TextCNN or Support Vector Machines). The dynamic topology calculation and hierarchical retrieval module is equipped with a cross-encoder model. The interactive interpretation and rollback scheduling module has a built-in graph attention network model, a random walk algorithm module, a feedback cause classification model, and a collaborative cost function, and is equipped with a rejection dialogue counter.

[0026] The semantic enhancement and indexing module reads the directory metadata, calls the large language model to perform logical inference based on the directory metadata to generate semantically enhanced metadata, calls the text embedding model to perform vectorization mapping processing on the semantically enhanced metadata to obtain the directory semantic vector, constructs a vector index library, and writes the directory semantic vector into the vector index library.

[0027] The requirement dimensionality reduction and feature extraction module obtains the natural language retrieval request. The requirement dimensionality reduction and feature extraction module uses the named entity recognition model to extract the entity parameters inside the natural language retrieval request. The requirement dimensionality reduction and feature extraction module performs rule mapping processing based on the entity parameters to generate explicit constraint rule parameters and generate a set of structured hard constraint conditions. The module performs text cleaning operation on the natural language retrieval request to generate cleaned retrieval text, and calls the text embedding model to convert the cleaned retrieval text into a requirement semantic feature vector.

[0028] The dynamic topology calculation and hierarchical retrieval module performs data filtering operations on the vector index library based on the structured hard constraint condition set to obtain a compliant directory subset. The module then calculates the similarity value between the required semantic feature vector and the corresponding directory semantic vector in the compliant directory subset. Based on the similarity value, the module sorts the results to select a coarse candidate set. Next, the module extracts shared approval logs from the blockchain network and constructs a directed bipartite graph based on the node call relationships recorded in the shared approval logs. Based on the directed bipartite graph, the module calculates the dynamic authority score of the corresponding data nodes within the coarse candidate set. Finally, the module combines the similarity value and the dynamic authority score for a second sort, outputting a refined candidate directory list.

[0029] The interactive explanation and rollback scheduling module generates matching explanation text based on the fine-ranking candidate directory list. The interactive explanation and rollback scheduling module sends the fine-ranking candidate directory list and the matching explanation text to the data user node. When the interactive explanation and rollback scheduling module generates and distributes the fine-ranking candidate directory list for the first time, it simultaneously caches the supply and demand matching agent's forward operation execution path record when the constraint increment is not loaded for the first time as the initial optimal solution.

[0030] The interactive interpretation and rollback scheduling module receives rejection commands from data-using nodes. It maps the system computation steps and data flow processes within the supply-demand matching agent to nodes and edges, abstracting them into a directed acyclic graph state machine. The module transforms the rejection commands into constraint increments and calculates the topological authority gradients of nodes within the directed bipartite graph. Using these topological authority gradients as damping parameters, the module combines the damping parameters and constraint increments, employing Dijkstra's algorithm and A... The search algorithm or topology sorting dynamic programming algorithm performs pathfinding calculations in the directed acyclic graph state machine, outputting the optimal rollback path. The interactive interpretation and rollback scheduling module controls the supply and demand matching agent to jump to the corresponding module along the optimal rollback path to re-execute the calculation. The interactive interpretation and rollback scheduling module calls the large language model or named entity recognition model to extract the logical conflict parameters obtained by parsing the feedback reason text data carried by the rejection instruction, and configures the logical conflict parameters as constraint conflict deviation data. The interactive interpretation and rollback scheduling module extracts the algorithm residual generated by the mapping comparison between the initial optimal solution and the constraint conflict deviation data. According to the reverse penalty mapping rule, the interactive interpretation and rollback scheduling module takes the data user node as the topology starting point and the data provider node as the topology ending point, and transforms the algorithm residual into negative constraint edges with numerical weights. The interactive interpretation and rollback scheduling module adds the negative constraint edges to the directed bipartite graph.

[0031] The specific process by which the supply and demand matching agent performs data matching processing based on directory metadata and natural language retrieval requests is as follows.

[0032] The semantic enhancement and indexing module reads the directory metadata stored in the blockchain network through an interface. The directory metadata includes raw structured fields, data sharing strategy fields, and unique identifiers. The raw structured fields include data form name field, data item name field, node department affiliation field, and registration time field. The semantic enhancement and indexing module extracts the text character data contained in the raw structured fields.

[0033] The semantic enhancement and indexing module calls the large language model to expand the dimensions of the original structured fields contained in the directory metadata. The semantic enhancement and indexing module is configured with preset text analysis instructions, which embed a prior mapping dictionary of departmental functions and business scenarios, as well as a calling interface of the external domain knowledge base. For example, the prior mapping dictionary is configured with the correspondence between "agriculture and rural departments" and "greenhouse microclimate control data". The large language model can infer the business scenario of the data based on this. The preset text analysis instructions contain reasoning logic for business scenarios, application purposes, related data, and data granularity dimensions. The semantic enhancement and indexing module imports the text character data contained in the original structured fields into the preset text analysis instructions. The semantic enhancement and indexing module transmits the preset text analysis instructions to the large language model. The large language model combines the external domain knowledge base to perform natural language reasoning operations to output business scenario description text, application purpose description text, related data description text, and data granularity description text.

[0034] The semantic enhancement and indexing module obtains the business scenario description text, application purpose description text, related data description text, and data granularity description text output by the large language model. The semantic enhancement and indexing module summarizes and concatenates the original structured fields with the business scenario description text, application purpose description text, related data description text, and data granularity description text. Based on the summarized and concatenated overall text data, the semantic enhancement and indexing module generates semantically enhanced metadata.

[0035] The semantic enhancement and indexing module calls the text embedding model to perform vectorized mapping processing on the semantic enhancement metadata. The semantic enhancement and indexing module inputs the semantic enhancement metadata into the text embedding model. The text embedding model extracts the text data contained in the semantic enhancement metadata. The text embedding model performs high-dimensional space mapping on the text data. The text embedding model outputs a fixed-length dense feature vector based on the high-dimensional space mapping result. The semantic enhancement and indexing module obtains the fixed-length dense feature vector and configures the fixed-length dense feature vector as the directory semantic vector.

[0036] The semantic enhancement and indexing module initializes the storage space structure to build a vector index library. The semantic enhancement and indexing module extracts the unique identifiers configured in the directory metadata. The semantic enhancement and indexing module writes the directory semantic vectors into the vector index library. The semantic enhancement and indexing module establishes a mapping and binding relationship between the unique identifiers and the directory semantic vectors in the vector index library.

[0037] The requirement dimensionality reduction and feature extraction module obtains the natural language search request, performs text cleaning operation on the natural language search request, removes redundant characters contained in the natural language search request, and outputs the cleaned search text.

[0038] The requirement dimensionality reduction and feature extraction module configures a named entity recognition model. The module inputs the cleaned search text into the named entity recognition model. The named entity recognition model extracts the entity parameters contained in the cleaned search text. The entity parameters include data topic entity parameters, time range entity parameters, spatial range entity parameters, data granularity entity parameters, and application scenario entity parameters.

[0039] The requirement dimensionality reduction and feature extraction module configures the intent classification model. The module inputs the cleaned search text into the intent classification model. The intent classification model performs feature mapping and classification probability calculation operations to output intent category labels. Intent category labels include exploratory search labels, exact match search labels, and condition-oriented search labels. The requirement dimensionality reduction and feature extraction module stores the intent category labels.

[0040] The requirement dimensionality reduction and feature extraction module performs rule mapping processing based on entity parameters to generate explicit constraint rule parameters. The explicit constraint rule parameters include department scope constraint parameters, timeliness constraint parameters, and data item inclusion constraint parameters. The requirement dimensionality reduction and feature extraction module combines the explicit constraint rule parameters to generate a structured hard constraint condition set.

[0041] The requirement dimensionality reduction and feature extraction module calls the text embedding model to perform vectorization mapping processing on the cleaned retrieved text. The requirement dimensionality reduction and feature extraction module inputs the cleaned retrieved text into the text embedding model. The text embedding model performs high-dimensional space mapping on the cleaned retrieved text. The text embedding model outputs the requirement semantic feature vector based on the high-dimensional space mapping result. The requirement dimensionality reduction and feature extraction module records the requirement semantic feature vector and the set of structured hard constraints.

[0042] The dynamic topology calculation and hierarchical retrieval module retrieves the structured hard constraint condition set and the vector index library. It then parses the explicit constraint rule parameters contained in the structured hard constraint condition set. Based on these explicit constraint rule parameters, the module performs Boolean condition matching calculations within the vector index library. Finally, it removes directory semantic vectors and their corresponding unique identifiers from the vector index library that do not conform to the explicit constraint rule parameters. Finally, it extracts directory semantic vectors and their corresponding unique identifiers from the vector index library that conform to the explicit constraint rule parameters to generate a compliant directory subset.

[0043] The dynamic topology calculation and hierarchical retrieval module reads the intent category tags stored in the requirement dimensionality reduction and feature extraction module. The dynamic topology calculation and hierarchical retrieval module configures a category threshold mapping table, which contains a preset mapping relationship between intent category tags and recall threshold values. The dynamic topology calculation and hierarchical retrieval module performs a mapping query operation in the category threshold mapping table based on the intent category tags to output the dynamic recall number.

[0044] The dynamic topology calculation and hierarchical retrieval module obtains the semantic feature vector of the requirement. The dynamic topology calculation and hierarchical retrieval module extracts the semantic vector of the directory contained in the compliant directory subset. The dynamic topology calculation and hierarchical retrieval module performs a cosine similarity calculation operation on the semantic feature vector of the requirement and each semantic vector of the directory in the compliant directory subset. The dynamic topology calculation and hierarchical retrieval module outputs the similarity value corresponding to each semantic vector of the directory based on the result of the cosine similarity calculation operation.

[0045] The dynamic topology calculation and hierarchical retrieval module performs a descending sort operation on the directory semantic vectors within the compliant directory subset based on similarity values. The dynamic topology calculation and hierarchical retrieval module truncates the results of the descending sort operation in conjunction with the dynamic recall quantity. The dynamic topology calculation and hierarchical retrieval module extracts the directory semantic vectors and associated unique identifiers that are ranked within the dynamic recall quantity range. The dynamic topology calculation and hierarchical retrieval module summarizes the directory semantic vectors and associated unique identifiers that are ranked within the dynamic recall quantity range to generate a coarse-screened candidate set.

[0046] The dynamic topology calculation and hierarchical retrieval module retrieves the shared approval logs stored in the blockchain network through the communication interface. The shared approval logs contain historical data supply and demand matching records, which include the data provider node identifier, the data user node identifier, the approval completion timestamp parameter, and the feedback evaluation score parameter.

[0047] The dynamic topology calculation and hierarchical retrieval module parses the historical data supply and demand matching records contained in the shared approval log. It extracts the data provider node identifiers and data user node identifiers from these records. The module then partitions and maps the data provider node identifiers as supply-side network graph nodes and the data user node identifiers as demand-side network graph nodes. Based on the flow and call relationships contained in the historical data supply and demand matching records, the module constructs directed connections between demand-side and supply-side network graph nodes. These directed connections include forward connections representing data calls and reverse connections representing data feedback or evaluation. The module statistically calculates the call frequency based on the flow and call relationships in the historical data supply and demand matching records and, combined with the feedback and evaluation score parameters, calculates and assigns initial connection weights to the directed connections, ensuring that both types of nodes possess computable in-degree and out-degree characteristics. Finally, the module merges the supply-side network graph nodes, demand-side network graph nodes, and directed connections to generate a directed bipartite graph.

[0048] The dynamic topology calculation and hierarchical retrieval module reads the unique identifiers contained in the coarse-screened candidate set. Based on these unique identifiers, it queries the corresponding directory metadata, extracts the node department affiliation field from the directory metadata, and uses this field as the data provider node identifier. This allows it to match and locate the corresponding supply-side network graph node in the directed bipartite graph. The module then extracts the set of directed edges pointing to this supply-side network graph node within the directed bipartite graph. It counts the number of edges in this set and calculates the node centrality parameter based on the connectivity value of the demand-side network graph node in the directed bipartite graph. Specifically, the dynamic topology calculation... When a historical flow call occurs, the hierarchical retrieval module extracts the sum of the in-degree and out-degree of the corresponding demand-side network graph node in the directed bipartite graph at that historical moment, and uses this as the node centrality parameter corresponding to the current historical call. The dynamic topology calculation and hierarchical retrieval module extracts the approval completion timestamp parameter contained in the historical data supply and demand matching record. The dynamic topology calculation and hierarchical retrieval module calculates the numerical difference between the current system timestamp and the approval completion timestamp parameter, and converts the numerical difference into a scalar value in days or hours to avoid function underflow in subsequent exponential decay calculations. Finally, it outputs the normalized time difference parameter corresponding to each call. It should be noted that the node centrality parameter and the time difference parameter are parameter sets corresponding to each historical flow call relationship.

[0049] The dynamic topology calculation and hierarchical retrieval module configures basic functional weight parameters, time decay constants, and adjustment coefficients. It extracts edge quantity parameters, node centrality parameters, and time difference parameters. Finally, it inputs these parameters into a preset dynamic authority calculation formula, which is as follows:

[0050] The dynamic topology calculation and hierarchical retrieval module defines the formula variable mapping relationship, which includes: Indicates dynamic authority score. This represents the basic functional weight parameter. This represents the number of edges. Indicates the first The corresponding node centrality parameter is called each time. Represents the time decay constant. Indicates the current system timestamp. Indicates the first The corresponding approval completion timestamp parameter is called each time. Indicates the first The normalized time difference parameter corresponding to each call. The adjustment coefficient is represented by the dynamic topology calculation and hierarchical retrieval module, which performs numerical substitution and feature calculation operations included in the dynamic authority calculation formula. Based on the results of the feature calculation operation, the dynamic topology calculation and hierarchical retrieval module outputs the dynamic authority score corresponding to the data node in the coarsely screened candidate set.

[0051] The Dynamic Topology Calculation and Hierarchical Retrieval module reads the cleaned retrieval text stored by the Dimensionality Reduction and Feature Extraction module. The Dynamic Topology Calculation and Hierarchical Retrieval module extracts the unique identifiers contained in the coarse candidate set. Based on the unique identifiers, the Dynamic Topology Calculation and Hierarchical Retrieval module extracts the corresponding semantically enhanced metadata.

[0052] The dynamic topology computation and hierarchical retrieval module extracts and cleans the retrieved text and semantically enhanced metadata. The dynamic topology computation and hierarchical retrieval module merges the cleaned retrieved text and semantically enhanced metadata to generate a cascaded text sequence. The dynamic topology computation and hierarchical retrieval module inputs the cascaded text sequence into the cross-encoder model. The cross-encoder model extracts the contextual attention features contained in the cascaded text sequence. The cross-encoder model uses its internally configured fully connected layer to reduce the dimensionality of the contextual attention features to a one-dimensional scalar value, which serves as the deep semantic matching score.

[0053] The dynamic topology calculation and hierarchical retrieval module extracts intent category tags. It configures a weight configuration matrix, which contains mapping association rules with intent category tags as index keys and sets containing three weight coefficient values ​​as key values. The dynamic topology calculation and hierarchical retrieval module performs matching query operations in the weight configuration matrix based on intent category tags. Based on the results of the matching query operations, the dynamic topology calculation and hierarchical retrieval module outputs vector recall weight coefficients, semantic ranking weight coefficients, and authority weight coefficients.

[0054] The dynamic topology calculation and hierarchical retrieval module extracts similarity values, deep semantic matching scores, and dynamic authority scores. The dynamic topology calculation and hierarchical retrieval module then substitutes the similarity values, deep semantic matching scores, dynamic authority scores, vector recall weight coefficients, semantic ranking weight coefficients, and authority weight coefficients into a preset comprehensive re-ranking formula, which is as follows.

[0055]

[0056] The dynamic topology calculation and hierarchical retrieval module sets the variable mapping relationship for the comprehensive reordering formula. The variable mapping relationship for the comprehensive reordering formula includes: This represents the overall reordering score. This represents the vector recall weight coefficient. Represents the similarity score. This represents the semantic ranking weight coefficient. Represents the deep semantic matching score. This represents the authority weighting coefficient. This indicates a dynamic authority score.

[0057] The dynamic topology calculation and hierarchical retrieval module performs numerical substitution and linear weighted calculation operations included in the comprehensive re-ranking formula. Based on the results of the linear weighted calculation operation, the dynamic topology calculation and hierarchical retrieval module generates a comprehensive re-ranking score corresponding to each unique identifier in the coarse candidate set. Based on the comprehensive re-ranking score, the dynamic topology calculation and hierarchical retrieval module performs a numerical descending order sorting operation on the unique identifiers included in the coarse candidate set. The dynamic topology calculation and hierarchical retrieval module integrates the unique identifiers output by the descending order sorting operation to generate a refined candidate directory list.

[0058] The interactive interpretation and rollback scheduling module obtains the candidate directory list for fine ranking, extracts the unique identifier contained in the candidate directory list, extracts the associated directory metadata, semantic enhancement metadata and dynamic authority score from the local storage based on the unique identifier, and retrieves the natural language retrieval request.

[0059] The interactive interpretation and rollback scheduling module configures interpretation generation instructions, which include natural language generation prompt rules for matching criteria, data coverage, and source authority endorsement. The interactive interpretation and rollback scheduling module configures the natural language retrieval request, directory metadata, semantic enhancement metadata, and dynamic authority score as input parameters for the large language model. The interactive interpretation and rollback scheduling module merges the input parameters and interpretation generation instructions and transmits them to the large language model. The large language model performs semantic reasoning operations to extract entity mapping overlap data between the natural language retrieval request and directory metadata. Based on the prompt rules, the large language model converts the entity mapping overlap data and dynamic authority score into matching interpretation text in natural language form.

[0060] The interactive explanation and rollback scheduling module obtains the matching explanation text output by the large language model. The interactive explanation and rollback scheduling module establishes a data binding mapping relationship between the matching explanation text and the corresponding items in the fine ranking candidate list. The interactive explanation and rollback scheduling module sends the fine ranking candidate list and the matching explanation text bound to it to the data user node.

[0061] After receiving the list of candidate directories for fine ranking and the matching explanation text, the data user node executes the branch processing flow in response to different operation instructions sent by the input device.

[0062] As part of the confirmation branch process, the data user node receives a confirmation operation command sent by the input device. The data user node converts the confirmation operation command into a confirmation feedback signal containing a unique identifier. The data user node sends the confirmation feedback signal to the interaction interpretation and rollback scheduling module. The interaction interpretation and rollback scheduling module receives the confirmation feedback signal and, based on the unique identifier carried by the confirmation feedback signal, extracts the directory metadata corresponding to the target directory from the blockchain network. It then parses the data sharing policy fields contained in the directory metadata. The interaction interpretation and rollback scheduling module configures the application form file according to the data sharing policy fields and sends the application form file to the data user node. The data user node obtains the data to be filled in according to the field requirements contained in the application form file, merges them to generate a sharing application form, and uploads and submits the sharing application form to the blockchain network. The blockchain network receives the sharing application form and triggers its internal smart contract approval process.

[0063] As part of the rejection branch process, the data user node receives the rejection operation instruction sent by the input device, extracts the feedback reason text data contained in the rejection operation instruction, encapsulates the feedback reason text data into a rejection instruction, and sends the rejection instruction to the interactive interpretation and rollback scheduling module.

[0064] The interactive interpretation and rollback scheduling module receives rejection instructions sent by data user nodes. It extracts the system computation steps and system data flow processes within the supply and demand matching agent. The module maps the system computation steps as state graph nodes and the system data flow processes as state graph connecting edges. It merges the state graph nodes and state graph connecting edges to construct a directed acyclic graph state machine. The directed acyclic graph state machine presents the forward computation execution path and logical dependency topology of the supply and demand matching agent.

[0065] The interactive interpretation and rollback scheduling module calls a large language model or named entity recognition model to parse the feedback reason text data contained in the rejection instruction, extracts the entity deviation or dimension missing fields contained therein as logical conflict parameters, and maps the logical conflict parameters to constraint conflict deviation data. The interactive interpretation and rollback scheduling module inputs the feedback reason text data into the feedback reason classification model, which performs a text classification probability calculation operation to output reason category labels. Reason category labels include semantic deviation labels, insufficient granularity labels, missing data item labels, and source mismatch labels. The interactive interpretation and rollback scheduling module extracts the reason category labels output by the feedback reason classification model.

[0066] The interactive interpretation and rollback scheduling module configures a constraint increment transformation matrix. This matrix contains a mapping relationship between cause category labels and preset mathematical correction dimension parameters. The module performs a mapping query operation on the constraint increment transformation matrix based on the cause category labels, outputting constraint increments containing the corresponding parameters. Specifically, the mapping relationship includes: when the cause category label is a semantic deviation label, the constraint increment includes a bias adjustment parameter pointing to the requirement semantic feature vector, guiding the optimal rollback path to the requirement dimensionality reduction and feature extraction module; when the cause category label is an insufficient granularity label or a missing data item label, the constraint increment includes additional hard boundary filtering parameters, guiding the optimal rollback path to the dynamic... The dynamic topology calculation and hierarchical retrieval module triggers data filtering operations. When the cause category label is a source mismatch label, the constraint increment includes a vector space bias adjustment parameter for the authority weight coefficient to guide the optimal rollback path to locally enter the reordering calculation node of the dynamic topology calculation and hierarchical retrieval module. When the rejection dialog counter detects that the cumulative number of rejection instructions for the same directory entry has reached a preset threshold, the constraint increment includes a global reset signal to guide the optimal rollback path to directionally enter the semantic enhancement and indexing module, triggering the regeneration of semantic enhancement metadata and vector index update operations for the target directory. The constraint increment includes corresponding additional hard boundary filtering parameters or vector space bias adjustment parameters, depending on the cause category label.

[0067] The interactive interpretation and rollback scheduling module loads constraint increments into the directed acyclic graph (DAG) state machine. It configures corresponding optimization objective solution conditions within the DAG state machine and sets a constraint satisfaction evaluation function. This function embeds a Markov decision process model. Specifically, the Markov decision process model treats the current node of the DAG state machine as a state, the pathfinding jump operation of loading constraint increments as an action, and the constraint satisfaction as a reward function to calculate the probability distribution of node state transitions. A preset tolerance threshold is also configured. The interactive interpretation and rollback scheduling module uses the constraint satisfaction evaluation function to calculate the output result distribution of the state graph nodes after loading constraint increments, outputting confidence probability values. The optimization objective solution conditions stipulate that the confidence probability value of loading constraint increments from state graph nodes and executing to the terminal output node is greater than the preset tolerance threshold value. The interactive interpretation and rollback scheduling module records the coupled computational structure of the DAG state machine, constraint increments, and optimization objective solution conditions.

[0068] The interactive interpretation and rollback scheduling module retrieves the directed bipartite graph constructed by the dynamic topology calculation and hierarchical retrieval module. Based on the dynamic authority score of each data node in the directed bipartite graph and its gradient distribution as it changes with topological connections, the interactive interpretation and rollback scheduling module calculates the topological authority gradient of each data node. Specifically, the interactive interpretation and rollback scheduling module calculates the difference in dynamic authority scores between the target data node and its first-order adjacent nodes in the directed bipartite graph, and performs a weighted average based on the weights of the corresponding directed connection edges to obtain the topological authority gradient under the discrete graph structure. The interactive interpretation and rollback scheduling module configures the collaborative cost function and transforms the topological authority gradient into the damping parameters configured in the collaborative cost function.

[0069] The interactive interpretation and rollback scheduling module extracts the system computing power overhead corresponding to the execution of operations by the directed acyclic graph state machine (DAG) based on the estimated computational load of the state graph nodes contained in the DAG. The estimated computational load is the average number of floating-point operations (FLOPs) or the average peak memory usage recorded by each module within the supply-demand matching agent during historical execution, which is normalized to obtain the system computing power overhead value. The interactive interpretation and rollback scheduling module configures a cost balancing factor vector, which includes a first cost balancing factor, a second cost balancing factor, and a penalty coefficient. The interactive interpretation and rollback scheduling module combines the system computing power overhead value, the cost balancing factor vector, the damping parameter, and the additional penalty term derived from the constraint increment (…). Substituting into the collaborative cost function, the specific calculation formula for the collaborative cost function is as follows:

[0070] The interactive interpretation and rollback scheduling module sets the mapping relationship of the cooperative cost function variables. The mapping relationship of the cooperative cost function variables includes: Indicates the value of pathfinding. This represents the system's computing power overhead. This represents the topological authority gradient as a damping parameter. This indicates additional penalties. This represents the first cost balancing factor. This represents the second cost balancing factor. This represents the penalty coefficient.

[0071] The interactive interpretation and rollback scheduling module executes node traversal and pathfinding operations within the directed acyclic graph state machine. Combining constraint increments and optimization objective solution conditions, the module calculates the pathfinding cost corresponding to the node traversal and pathfinding operations using a collaborative cost function. Specifically, during pathfinding, the module transforms the constraint increments into dynamic penalty weights for the corresponding state graph connection edges and integrates these dynamic penalty weights as additional penalty terms into the pathfinding cost. The integrated cost is then used to update Dijkstra's algorithm and A*. The edge weights or heuristic function estimates in pathfinding algorithms such as search algorithms are used by the interactive interpretation and rollback scheduling module to extract the flow path with the minimum pathfinding cost based on the integrated cost value after fusion, and generate the optimal rollback path.

[0072] The optimal rollback path includes the entry computing node identifier. The interactive interpretation and rollback scheduling module extracts the entry computing node identifier contained in the optimal rollback path. The interactive interpretation and rollback scheduling module generates a recalculation trigger instruction based on the entry computing node identifier. The interactive interpretation and rollback scheduling module sends the recalculation trigger instruction to the computing module corresponding to the entry computing node identifier within the supply and demand matching agent. The supply and demand matching agent loads the constraint increment based on the recalculation trigger instruction and controls the corresponding computing module to re-execute the data processing operation.

[0073] The interactive interpretation and rollback scheduling module transforms the initial optimal solution and constraint conflict deviation data into feature vectors of the same dimension. Specifically, the module uses a graph attention network model or a random walk algorithm to extract the node sequence features of the directed acyclic graph state machine corresponding to the initial optimal solution, and aggregates them into a fixed-length one-dimensional feature vector through a pooling layer. Simultaneously, the module calls a text embedding model to encode the constraint conflict deviation data and performs dimensional interpolation and truncation operations through its built-in fully connected layer, thereby generating a deviation feature vector of the same dimension as the one-dimensional feature vector. Subsequently, the module performs numerical calculations on the degree of deviation of the initial optimal solution to output the algorithm residual. As a preferred implementation, the cross-entropy loss function can be used to calculate this degree of deviation. Before substituting the cross-entropy loss function, the module uses a Softmax function or a Sigmoid activation function to perform probability normalization on the one-dimensional feature vector corresponding to the initial optimal solution and the deviation feature vector, respectively, so that the feature values ​​of each dimension are mapped to... Within the interval, to meet the comparison requirements of the probability distribution, the specific calculation formula is as follows:

[0074] In the formula, Represents the algorithm residual. The dimension of the feature vector. The first eigenvector representing the one-dimensional eigenvector corresponding to the initial optimal solution. The target feature values ​​or true distribution in each dimension. The first eigenvector representing the deviation from the constraint conflict data. The current feature values ​​or predicted probability distribution of each dimension.

[0075] The interactive interpretation and rollback scheduling module configures a reverse penalty mapping rule. Specifically, the reverse penalty mapping rule stipulates that: the identifier of the data user node that generated the rejection instruction is extracted as the starting point of the topology connection, the identifier of the data provider node corresponding to the rejected directory metadata is extracted as the ending point of the topology connection, a connection edge is constructed between the starting point and the ending point as the negative constraint edge, and the algorithm residual is processed by numerical normalization as the negative weight parameter of the negative constraint edge. The interactive interpretation and rollback scheduling module converts the algorithm residual into a negative constraint edge according to the reverse penalty mapping rule, and writes the negative constraint edge into the directed bipartite graph. The dynamic topology calculation and hierarchical retrieval module updates the dynamic authority score of each data node according to the directed bipartite graph after loading the negative constraint edge.

[0076] It should be noted that the relevant hyperparameters and weight coefficients involved in the embodiments of the present invention (such as basic functional weight parameters, adjustment coefficients, time decay constants, preset tolerance threshold values, etc.) can be preset by those skilled in the art based on actual business scenarios (such as historical data characteristics of cross-departmental data sharing) through empirical values, or dynamically trained using historical log sample data through hyperparameter optimization algorithms such as grid search and random search, as well as machine learning algorithms such as gradient descent. The present invention does not impose absolute limitations on this.

[0077] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for on-chain directory retrieval and supply-demand matching based on blockchain and intelligent agents, characterized in that, An intelligent agent for supply and demand matching, applied to communication connections in a blockchain network, between data provider nodes and data user nodes; the method includes: Obtain the directory metadata sent to the blockchain network by the data provider node, call the large language model to perform logical inference based on the directory metadata to generate semantically enhanced metadata, perform vectorization mapping processing on the semantically enhanced metadata to obtain the directory semantic vector, and write the directory semantic vector into the configured vector index library; Receive the natural language retrieval request sent by the data user node, extract the entity parameters inside the natural language retrieval request, generate a set of structured hard constraints based on the entity parameters, perform text cleaning operation on the natural language retrieval request to generate cleaned retrieval text, and convert the cleaned retrieval text into a requirement semantic feature vector. Based on the structured hard constraint set, a data filtering operation is performed on the vector index library to obtain a compliant directory subset. The similarity value between the demand semantic feature vector and the directory semantic vector in the compliant directory subset is calculated to select a coarse candidate set. The shared approval logs stored in the blockchain network are obtained. A directed bipartite graph is constructed based on the shared approval logs. The dynamic authority score of the corresponding data nodes within the coarse candidate set is calculated based on the directed bipartite graph. The similarity value and the dynamic authority score are merged for secondary sorting and a fine-ranked candidate directory list is output. Based on the refined ranking candidate directory list, a matching explanation text is generated, and the refined ranking candidate directory list and the matching explanation text are sent to the data user node. When the refined ranking candidate directory list is generated and sent for the first time, the forward operation execution path record when the constraint increment is not loaded for the first time is cached as the initial optimal solution. If a rejection instruction carrying feedback reason text data is received from the data user node, the system calculation steps and system data flow process within the supply and demand matching agent are mapped to nodes and edges respectively and abstracted into a directed acyclic graph state machine. The rejection instruction is converted into a constraint increment, the feedback reason text data is parsed to obtain constraint conflict deviation data, the topological authority gradient of the data node in the directed bipartite graph is calculated and the topological authority gradient is used as a damping parameter. The damping parameter and the constraint increment are combined to perform pathfinding calculation in the directed acyclic graph state machine to output the optimal rollback path. The calculation is re-executed along the optimal rollback path. At the same time, the algorithm residual generated by the mapping comparison between the initial optimal solution and the constraint conflict deviation data is extracted, and the algorithm residual is converted into a negative constraint edge with numerical weight and added to the directed bipartite graph.

2. The method for on-chain directory retrieval and supply-demand matching based on blockchain and intelligent agents according to claim 1, characterized in that, The directory metadata contains raw structured fields, which contain text character data. The step of calling a large language model to perform logical inference based on the directory metadata to generate semantically enhanced metadata includes: Extract the text character data contained in the original structured fields; Configure a preset text analysis instruction, import the text character data into the preset text analysis instruction, and transmit the preset text analysis instruction to the large language model; The large language model combines an external domain knowledge base to perform natural language reasoning operations and outputs business scenario description text, application purpose description text, related data description text, and data granularity description text. The semantically enhanced metadata is generated by summarizing and concatenating the original structured fields with the business scenario description text, the application purpose description text, the associated data description text, and the data granularity description text, and then generating the semantically enhanced metadata based on the summarized and concatenated overall text data.

3. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 1, characterized in that, The step of extracting entity parameters from the natural language retrieval request and generating a set of structured hard constraint conditions based on the entity parameters includes: The cleaned search text is input into a pre-configured named entity recognition model, and the named entity recognition model is used to extract the entity parameters contained in the cleaned search text. The cleaned and retrieved text is input into a pre-configured intent classification model, and the intent classification model is used to perform feature mapping and classification probability calculation operations to output intent category labels; The rule mapping process is performed based on the entity parameters to generate explicit constraint rule parameters, and the explicit constraint rule parameters are then concatenated and merged to generate the structured hard constraint condition set.

4. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 3, characterized in that, The vector index library is configured with unique identifiers. The step of performing data filtering operations on the vector index library based on the structured hard constraint condition set to obtain a compliant directory subset, and calculating similarity values ​​to screen out a coarse candidate set includes: The explicit constraint rule parameters contained in the structured hard constraint condition set are parsed, and Boolean condition matching calculations are performed in the vector index library based on the explicit constraint rule parameters to extract the directory semantic vectors that conform to the explicit constraint rule parameters and their corresponding unique identifiers to generate the compliant directory subset. Based on the intent category label, perform a mapping query operation in a pre-configured category threshold mapping table to output the dynamic recall quantity; A cosine similarity calculation operation is performed on the semantic feature vector of the requirement and the semantic vector of each directory in the compliant directory subset to output the similarity value corresponding to each directory semantic vector; The directory semantic vectors are sorted in descending order based on the similarity values, and the directory semantic vectors that are within the range of the dynamic recall are extracted and associated with unique identifiers in combination with the dynamic recall number to generate the coarse candidate set.

5. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 4, characterized in that, The shared approval log contains historical data supply and demand matching records. These records include data provider node identifiers, data user node identifiers, approval completion timestamp parameters, and feedback evaluation score parameters. The step of constructing a directed bipartite graph based on the shared approval log and calculating a dynamic authority score based on the directed bipartite graph includes: The data provider node identifier is mapped and configured as a supply-side network graph node, and the data user node identifier is mapped and configured as a demand-side network graph node. Directed connection edges are constructed based on the flow and call relationship, and the initial connection weights are calculated and assigned to the directed connection edges in combination with the feedback evaluation score parameters, and then the directed bipartite graph is generated. Based on the unique identifier contained in the coarse candidate set, the corresponding supply-side network graph node is matched and located in the directed bipartite graph. The directed connection edges pointing to the supply-side network graph node are extracted to generate a connection edge set, and the number of edges contained in the connection edge set is counted. For each historical flow call, the sum of the in-degree and out-degree of the corresponding demand-side network graph node in the directed bipartite graph is extracted as the node centrality parameter, and the numerical difference between the current system timestamp and the corresponding approval completion timestamp parameter is calculated as the time difference parameter. Extract the preset basic functional weight parameters, time decay constant, and adjustment coefficient. Multiply the node centrality parameter corresponding to each call by the exponential decay term generated based on the product of the time decay constant and the time difference parameter. Summate the multiple values ​​obtained after multiplication. Multiply the sum by the adjustment coefficient to obtain the adjustment compensation value. Add the basic functional weight parameters and the adjustment compensation value to obtain the final dynamic authority score.

6. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 5, characterized in that, The step of fusing the similarity value and the dynamic authority score for secondary sorting and outputting a refined candidate directory list includes: The cleaned retrieved text is merged with the semantically enhanced metadata to generate a concatenated text sequence. The concatenated text sequence is then input into a pre-configured cross-encoder model for dimensionality reduction mapping to output a deep semantic matching score. Based on the intent category label, a matching query operation is performed in the preset weight configuration matrix, and the vector recall weight coefficient, semantic ranking weight coefficient, and authority weight coefficient are output. The first product is obtained by multiplying the similarity value by the vector recall weight coefficient, the second product is obtained by multiplying the deep semantic matching score by the semantic ranking weight coefficient, and the third product is obtained by multiplying the dynamic authority score by the authority weight coefficient. Perform a linear weighted summation operation on the first product, the second product, and the third product to generate a comprehensive re-ranking score, and integrate the unique identifiers in the coarse candidate set in descending order based on the comprehensive re-ranking score to generate the fine-ranking candidate directory list.

7. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 1, characterized in that, The step of converting the rejection instruction into a constraint increment includes: The feedback reason text data is input into a pre-configured feedback reason classification model to perform text classification probability calculation to output reason category labels. The reason category labels include semantic deviation labels, insufficient granularity labels, missing data item labels, and source mismatch labels. Based on the cause category label, a mapping query operation is performed in the pre-configured constraint increment transformation matrix, and the mapping output includes the constraint increment containing the corresponding bias adjustment parameter or filtering parameter; Specifically, when the cause category label is an insufficient granularity label or a missing data item label, the constraint increment includes additional hard boundary filtering parameters to guide the optimal rollback path to trigger data filtering operations; when the cause category label is a source mismatch label, the constraint increment includes vector space bias adjustment parameters for the authority weight coefficient.

8. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 1, characterized in that, The directed acyclic graph state machine includes state graph nodes and state graph connecting edges. The step of performing pathfinding calculations in the directed acyclic graph state machine by combining the damping parameters and the constraint increments to output the optimal rollback path includes: Based on the estimated computational load of the state diagram nodes, extract the system computing power overhead value corresponding to the system operation, and configure a cost balance factor vector including a first cost balance factor, a second cost balance factor, and a penalty coefficient. The first computational cost is obtained by multiplying the system computing power cost value by the first cost balance factor; the second computational cost is obtained by multiplying the exponential decay term, which has the natural constant as the base and the negative number of the damping parameter as the exponent, by the second cost balance factor; and the third computational cost is obtained by multiplying the additional penalty term, which is converted from the constraint increment, by the penalty coefficient. The pathfinding cost is calculated by summing the first calculated cost value, the second calculated cost value, and the third calculated cost value. Combining the constraint increment with the preset optimization objective solution conditions, a node traversal pathfinding operation is performed inside the directed acyclic graph state machine. Based on the pathfinding substitution value after incorporating the additional penalty term, the flow path with the minimum substitution value is extracted, and the optimal rollback path containing the entry calculation node identifier is generated.

9. The on-chain directory retrieval and supply-demand matching method based on blockchain and intelligent agents according to claim 1, characterized in that, The step of extracting the algorithm residual and converting it into negative constraint edges includes: The pre-configured text embedding model is invoked to perform feature encoding and dimension interpolation truncation operations on the constraint conflict deviation data to generate deviation feature vectors; The state machine node sequence features corresponding to the initial optimal solution are extracted using a graph attention network model or a random walk algorithm, and then aggregated into a one-dimensional feature vector through a pooling layer. The one-dimensional feature vector and the deviation feature vector are subjected to probability normalization, and the degree of deviation of the probability distribution is calculated using the cross-entropy loss function to output the residual of the algorithm. Configure a reverse penalty mapping rule, extract the data provider node identifier that generated the rejection instruction as the topology connection start point, extract the data provider node identifier corresponding to the rejected directory metadata as the topology connection end point, construct a connection edge between the topology connection start point and the topology connection end point as the negative constraint edge, and use the algorithm residual after numerical normalization as the negative weight parameter of the negative constraint edge.

10. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the on-chain directory retrieval and supply-demand matching method based on blockchain and smart agents as described in any one of claims 1 to 9.

Images