An agent low-code arrangement method and system for multi-source data

Abstract

The application provides a multi-source data-oriented agent low-code arrangement method and system, relates to the technical field of data intelligent analysis and low-code development, and comprises the following steps: according to a pre-acquired standardized initial data set, a data processing pipeline is constructed by dragging and dropping operators in a preset low-code arrangement canvas; the standardized initial data set is subjected to data cleaning according to the data processing pipeline, and cleaned data is obtained; according to the cleaned data, feature vectorization processing is performed in the data processing pipeline, and an intermediate analysis set with business semantic labels and feature vectorization operator parameters are obtained; and according to the business semantic labels of the intermediate analysis set, corresponding semantic routing rules are obtained. The application can realize continuous perception and adaptive iterative optimization of an arrangement model to user interaction feedback, and effectively solves the problems of semantic routing threshold rigidity and operator parameter degradation.

Description

Technical Field

[0001] This invention relates to the fields of data intelligent analysis and low-code development technology, and in particular to a method and system for low-code orchestration of intelligent agents for multi-source data. Background Technology

[0002] In the field of data-driven intelligent decision-making, building intelligent agent applications for multi-source data through low-code orchestration canvases has become an important means for enterprise digital transformation. These intelligent agents need to complete the entire analysis chain from data cleaning and feature processing to semantic intent recognition, and then to multi-task routing and result output.

[0003] In existing low-code orchestration methods for intelligent agents, the configuration of semantic routing rules and data processing operators generally relies on offline predefined static parameters. After the orchestration instance is deployed online, the parameters of intent recognition nodes, tool invocation nodes, semantic routing rules, and feature vectorization operators remain fixed. This approach often suffers from the following drawbacks: the orchestration model struggles to perceive actual user interaction feedback and adjust accordingly; when business analysis criteria or data distribution characteristics evolve over time, the preset semantic routing rule matching thresholds cannot automatically adapt to new semantic boundaries, leading to a continuous decline in the matching accuracy between intent parsing results and triggering links; and because the core parameters of feature vectorization processing are difficult to dynamically calibrate based on user adoption or correction of analysis results, the quality of intermediate analysis sets with business semantic labels gradually decreases, thus affecting the output performance of targeted queries and model inference.

[0004] Taking enterprise supply chain management query scenarios as an example, if users repeatedly correct the results of natural language queries related to inventory turnover analysis within a continuous cycle, existing methods are unable to transform such feedback signals into closed-loop optimization of semantic routing strategies and feature processing parameters, resulting in the continuous accumulation of deviations between the agent's response and the actual business needs. Summary of the Invention

[0005] This invention provides a low-code orchestration method and system for intelligent agents with multi-source data. While ensuring the convenience of low-code orchestration, it enables the orchestration model to continuously perceive and adaptively iteratively optimize user interaction feedback, effectively solves the problems of rigid semantic routing thresholds and degenerate operator parameters, and continuously improves the matching accuracy between intelligent agent responses and business requirements.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a low-code orchestration method for intelligent agents oriented towards multi-source data, the method comprising: Step 1: Based on the pre-acquired standardized initial dataset, construct a data processing pipeline in a pre-set low-code orchestration canvas by dragging and dropping operators; perform data cleaning on the standardized initial dataset according to the data processing pipeline to obtain cleaned data; Step 2: Based on the cleaned data, feature vectorization processing is performed in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; based on the business semantic labels of the intermediate analysis set, the corresponding semantic routing rules are obtained. Step 3: Based on the semantic routing rules, configure intent recognition nodes and tool invocation nodes in the preset low-code orchestration canvas to obtain the agent decision graph; perform logical verification on the agent decision graph to obtain an executable agent orchestration instance; Step 4: Based on the executable intelligent agent orchestration instance, perform intent parsing on the received natural language query to obtain the intent parsing result. Activate the corresponding semantic routing rule based on the intent parsing result to trigger the association operator in the data processing pipeline to perform targeted querying and model reasoning on the intermediate analysis set to obtain the structured analysis result set. Step 5: Based on the structured analysis result set, obtain the target display view; collect user interaction feedback data for the target display view; based on the user interaction feedback data, iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold to obtain the optimized agent low-code orchestration model.

[0007] Secondly, a low-code orchestration system for intelligent agents oriented towards multi-source data includes: The data pipeline construction module is used to construct a data processing pipeline by dragging and dropping operators in a preset low-code orchestration canvas based on a pre-acquired standardized initial dataset; and to perform data cleaning on the standardized initial dataset according to the data processing pipeline to obtain cleaned data. The feature semantic mapping module is used to perform feature vectorization processing on the cleaned data in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; and to obtain the corresponding semantic routing rules based on the business semantic labels of the intermediate analysis set. The decision graph orchestration and verification module is used to configure intent recognition nodes and tool invocation nodes in a preset low-code orchestration canvas according to the semantic routing rules to obtain an agent decision graph; and to perform logical verification on the agent decision graph to obtain an executable agent orchestration instance. The intent routing execution module is used to perform intent parsing on the received natural language query based on the executable intelligent agent orchestration instance, obtain the intent parsing result, activate the corresponding semantic routing rule based on the intent parsing result, so as to trigger the association operator in the data processing pipeline to perform targeted query and model inference on the intermediate analysis set, and obtain the structured analysis result set. The feedback iterative optimization module is used to obtain the target display view based on the structured analysis result set; collect user interaction feedback data on the target display view; and iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold based on the user interaction feedback data to obtain the optimized agent low-code orchestration model.

[0008] The above-described solution of the present invention has at least the following beneficial effects: Because it adopts a progressive processing link from the standardized initial dataset to the cleaned data, and then to the intermediate analysis set and semantic routing rules, and introduces a pre-constructed implicit topological metric space in the feature vectorization stage to map the initial feature vectors to a non-Euclidean state potential surface, and uses business semantic labels as prototype attractors to perform adaptive semantic field discretization processing, and constructs semantic phase transition hypersurface boundaries, it overcomes the technical defects of existing methods where semantic routing rules and feature vectorization operator parameters are statically fixed, and the orchestration model cannot perceive user interaction feedback and adjust itself accordingly.

[0009] Based on this, by collecting user interaction feedback data on the target display view, the potential energy gradient offset of the feedback signal feature vector relative to the semantic phase transition hypersurface boundary is calculated, a boundary offset calibration vector is generated, and the feature vectorization operator parameters and the preset semantic routing rule matching threshold are iteratively updated according to the boundary offset calibration vector. This achieves the technical effect of the intelligent agent orchestration model continuously perceiving the evolution of business semantic distribution, dynamically calibrating and analyzing link parameters, and accurately matching the intelligent agent response with the real business needs. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a low-code orchestration method for intelligent agents oriented towards multi-source data, provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of an intelligent agent low-code orchestration system for multi-source data provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a low-code orchestration method for intelligent agents oriented towards multi-source data, the method comprising the following steps: Step 1: Based on the pre-acquired standardized initial dataset, construct a data processing pipeline in a pre-set low-code orchestration canvas by dragging and dropping operators; perform data cleaning on the standardized initial dataset according to the data processing pipeline to obtain cleaned data; Step 2: Based on the cleaned data, feature vectorization processing is performed in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; based on the business semantic labels of the intermediate analysis set, the corresponding semantic routing rules are obtained. Step 3: Based on the semantic routing rules, configure intent recognition nodes and tool invocation nodes in the preset low-code orchestration canvas to obtain the agent decision graph; perform logical verification on the agent decision graph to obtain an executable agent orchestration instance; Step 4: Based on the executable intelligent agent orchestration instance, perform intent parsing on the received natural language query to obtain the intent parsing result. Activate the corresponding semantic routing rule based on the intent parsing result to trigger the association operator in the data processing pipeline to perform targeted querying and model reasoning on the intermediate analysis set to obtain the structured analysis result set. Step 5: Based on the structured analysis result set, obtain the target display view; collect user interaction feedback data for the target display view; based on the user interaction feedback data, iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold to obtain the optimized agent low-code orchestration model.

[0014] In this embodiment of the invention, by constructing a progressive processing chain from data cleaning and feature vectorization to semantic routing generation, the data processing pipeline is built and the agent decision graph is configured in a low-code orchestration canvas, realizing the visual and rapid orchestration of agent question-and-answer applications. By collecting user interaction feedback data on the target display view, the feature vectorization operator parameters and the preset semantic routing rule matching threshold are iteratively updated, enabling the orchestration model to continuously perceive changes in the business semantic distribution and adaptively calibrate and analyze the chain parameters, effectively improving the matching accuracy between agent responses and real business needs.

[0015] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Parse the standardized initial dataset to obtain the field structure and data type of the standardized initial dataset; based on the field structure and data type, match the corresponding target operator from the preset low-code orchestration canvas operator library to obtain the operator matching result. Specifically, this includes: parsing the standardized initial dataset, traversing the column name information and data samples of each data column in the dataset, extracting the field name and storage type of each column, and obtaining the field structure and data type of the standardized initial dataset.

[0016] The pre-built low-code orchestration canvas operator library is a collection of operators deployed along with the pre-built low-code orchestration canvas, used to provide drag-and-drop operator resources for building data processing pipelines. Each registered operator in the operator library carries corresponding operator metadata, which includes operator identifier, operator name, operator function description, operator input field type requirements, operator output field type definition, operator applicable scenario label, and operator default parameter configuration item. Among them, the operator input field type requirements define the storage type range of the input data columns that the operator can process; the operator output field type definition describes the storage type of the output data columns after the operator has processed the data; the operator applicable scenario label is used to mark the data quality scenarios to which the operator is applicable, including integrity applicable label, consistency applicable label, and standardization applicable label; the operator default parameter configuration item is used to store the preset parameter values ​​of the operator in the initialization state.

[0017] Based on the field structure and data type, the registered operator metadata in the operator library is traversed, and the operator input field type requirement of each operator is compared with the storage type of each field in the field structure. If the operator input field type requirement of an operator is consistent with the storage type of at least one field in the field structure, the operator is added to the filtering pass set. After all operators are traversed, all operators in the filtering pass set are combined into an operator candidate set. For each operator in the operator candidate set, its applicable scenario label is obtained. Based on the data samples of the standardized initial dataset, the data quality profile of the standardized initial dataset is calculated, including completeness index, consistency index, and normalization index. The completeness index is used to measure the proportion of non-empty values ​​in each field of the dataset, the consistency index is used to measure the degree of consistency of value ranges between related fields in the dataset, and the normalization index is used to measure the degree of conformity of field values ​​in the dataset to the preset format specifications.

[0018] The operator's applicable scenario label is semantically matched with the data quality profile of the standardized initial dataset. Specifically, the completeness applicable label in the operator's applicable scenario label is matched with the completeness index of the data quality profile. If the completeness index is lower than a preset completeness threshold and the operator's applicable scenario label contains a missing value handling marker, then the operator is determined to be a successful match in this dimension. The preset completeness threshold ranges from 0.65 to 0.75, with a default value of 0.7. The consistency applicable label in the operator's applicable scenario label is matched with the consistency index of the data quality profile. If the consistency index is lower than a preset completeness threshold and the operator's applicable scenario label contains a missing value handling marker, then the operator is determined to be a successful match in this dimension. If the consistency index is lower than the preset consistency threshold and the operator's applicable scenario label contains an outlier handling flag, then the operator is determined to be a successful match in this dimension. The preset consistency threshold ranges from 0.75 to 0.85, with a default value of 0.8. The standardization application label in the operator's applicable scenario label is matched with the standardization index of the data quality profile. If the standardization index is lower than the preset standardization threshold and the operator's applicable scenario label contains a format conversion flag, then the operator is determined to be a successful match in this dimension. The preset standardization threshold ranges from 0.80 to 0.90, with a default value of 0.85.

[0019] The number of successful matches for each operator in the completeness, consistency, and normalization dimensions is counted, and the preset dimension weights for each dimension are obtained. These preset dimension weights are pre-configured weight parameters based on the data quality requirements of the business analysis scenario corresponding to the standardized initial dataset. The preset dimension weight for each dimension ranges from 0 to 1, and the sum of the preset dimension weights for the three dimensions is 1. These preset dimension weights are used to adjust the contribution of the successful match for that dimension to the final fit score. The number of successful matches for each dimension is multiplied by the preset dimension weight for that dimension to obtain the dimension fit score. The dimension fit scores for completeness, consistency, and normalization are added together to obtain the fit score between the operator and the data quality profile. The candidate set of operators is sorted from high to low according to the fit score, and the operator with the highest fit score is selected as the target operator to obtain the operator matching result.

[0020] Step 1.2: Based on the operator matching results, the target operator is added to the canvas editing area by dragging and dropping, and the connection relationship between operators is established to obtain the initial data processing pipeline. Specifically, this includes: highlighting the target operator in the operator selection panel of the preset low-code orchestration canvas based on the operator matching results. This operator selection panel is a functional area in the preset low-code orchestration canvas used to centrally display selectable operators and perform retrieval, preview, and drag-and-drop operations on operators. The operator selection panel includes an operator search box, an operator category directory area, an operator list display area, and an operator details preview area. The operator search box is used to receive keywords input by the user and perform fuzzy matching retrieval of operator names and operator function descriptions. The operator category directory area is used to group and display operators according to data cleaning, feature engineering, and natural language processing categories. The operator list display area is used to display the icon, name, and brief function description of each operator in card form. The operator details preview area is used to display the function description and input / output port information of an operator when the user clicks or hovers over an operator card.

[0021] In the operator list display area of ​​the operator selection panel, the operator card corresponding to the target operator is highlighted and rendered, and the functional description and input / output port information of the target operator are presented to the user. The drag event detector is activated to detect the user's drag operation command on the target operator. The drag event detector is an event handling component in the pre-built low-code orchestration canvas responsible for capturing the user's drag interaction behavior. It includes a mouse event detection unit, a touch event detection unit, and a drag state tracking unit. The mouse event detection unit is used to detect the coordinate position and timestamp of mouse press, move, and release events. The touch event detection unit is used to detect the touch point coordinates and timestamps of finger touch start, move, and leave events on the touch screen. The drag state tracking unit is used to determine the start state, in progress state, and completion state of the drag operation based on mouse events or touch events, and updates the following position of the dragged operator card in real time when the drag is in progress state.

[0022] When the drag event detector determines that the release position of the drag operation command is within the legal placement area of ​​the canvas editing area, a canvas node is created at that release position, the target operator is instantiated as the canvas node, and the canvas coordinates of this canvas node are recorded to obtain the instantiated node. According to the data input port definition and data output port definition of the instantiated node, the user's drag operation connecting the output port of the current instantiated node to the input port of another instantiated node is detected, the completion event of the connection drag operation is responded to, a directed connection edge is established between the two instantiated nodes, and this directed connection edge is recorded in the edge set of the canvas editing area. After all target operators have been instantiated and all directed connection edges have been established in the operator matching results, a node set is formed by all instantiated nodes, and an edge set is formed by all directed connection edges. The initial data processing pipeline is generated based on the node set and the edge set.

[0023] Step 1.3: Based on the initial data processing pipeline, obtain the parameter configuration items of each operator in the initial data processing pipeline, and set data cleaning rules according to the parameter configuration items to obtain the configured data processing pipeline. Specifically, this includes: based on the initial data processing pipeline, traversing each operator node in the initial data processing pipeline according to the node topology order, and obtaining the parameter configuration items of each operator node; the parameter configuration items include field mapping rules, missing value imputation strategy, outlier truncation threshold, and data type conversion format; wherein, the field mapping rules are used to define the correspondence between the source dataset fields and the operator output fields, and the field mapping rules include source field name, target field name, and field mapping method; the source field name is used to specify the data column selected from the input data stream, and the target field name is used to specify the name of the data column output after mapping; the field mapping method includes direct mapping and computational mapping. Direct mapping means passing the source field value to the target field as is, and computational mapping means passing two or more source field values ​​to the target field after arithmetic operations or string concatenation.

[0024] The missing value imputation strategy defines how to handle null fields in the input data stream. It includes the target field, imputation method, and imputation value. The target field specifies the name of the field for which missing value imputation needs to be performed. Imputation methods include four types: fixed value imputation, mean imputation, median imputation, and mode imputation. The imputation value specifies the exact value to be imputed when using the fixed value imputation method. When using the mean imputation method, the system automatically calculates the average of the existing non-null values ​​in the field as the imputation value. When using the median imputation method, the system automatically calculates the median of the existing non-null values ​​in the field as the imputation value. When using the mode imputation method, the system automatically calculates the most frequent value among the existing non-null values ​​in the field as the imputation value.

[0025] Outlier truncation thresholds define the upper and lower bounds for detecting and truncating outliers in the input data stream. The outlier truncation threshold includes an upper threshold and a lower threshold. The upper threshold specifies the upper limit of a field's value; values ​​exceeding the upper threshold will be truncated to the upper limit. The lower threshold specifies the lower limit of a field's value; values ​​below the lower threshold will be truncated to the lower limit. In one example, the upper threshold is the 95th percentile of the field's percentile ranking, and the lower threshold is the 5th percentile of the field's percentile ranking. In another example, the upper threshold is the mean of the field plus three standard deviations, and the lower threshold is the mean of the field minus three standard deviations.

[0026] The data type conversion format defines the target storage type of the operator's output fields. It includes the source field name, source data type, and target data type. Source data types include integer, floating-point, string, and date / time types. The target data type is selected by the user from these three types based on subsequent analysis needs. According to the field mapping rules, the mapping relationship between the source and target fields is set in the operator node's attribute configuration panel, resulting in the field mapping configuration. Based on the missing value imputation strategy, the imputation target and method for missing value columns are specified, resulting in the missing value imputation configuration. Based on the outlier truncation threshold, the upper and lower bound thresholds for outlier detection are set, resulting in the outlier truncation configuration. Based on the data type conversion format, the data type of the target output field is selected, resulting in the type conversion configuration. These four configurations—field mapping, missing value imputation, outlier truncation, and type conversion—are merged into a data cleaning rule that the operator can parse. This data cleaning rule is then bound to the corresponding operator node, resulting in the configured data processing pipeline.

[0027] Step 1.4: Based on the configured data processing pipeline, perform data cleaning on the standardized initial dataset to obtain cleaned data. Specifically, this includes: calling the pipeline execution engine for execution according to the configured data processing pipeline. This pipeline execution engine is the core execution component in the pre-built low-code orchestration canvas responsible for parsing the data processing pipeline structure and driving each operator node to run sequentially. The pipeline execution engine internally includes a topology parser, a node scheduler, and a data flow manager. The topology parser is used to read the node set and edge set of the configured data processing pipeline, calculate the dependencies between nodes, generate the node topology order, and output the node execution sequence. This node execution sequence is arranged according to the topological sorting result of the directed acyclic graph from the first node to the last node. The node scheduler is used to send start signals to each operator node sequentially according to the node execution sequence, and receive a completion receipt after the current node finishes execution before triggering the start of the next node. The data flow manager is used to pass intermediate data streams between adjacent nodes and allocate temporary storage space for each intermediate data stream.

[0028] The pipeline execution engine calls the topology parser to parse the configured data processing pipeline, obtaining the node execution sequence. Based on this sequence, the node scheduler uses the standardized initial dataset as the input data stream for the first node and sends a start signal to initiate its execution. Upon receiving the start signal, the first node reads the field mapping configuration from its bound data cleaning rules. According to this configuration, it transfers or calculates the source field values ​​from the input data stream to the target field, resulting in a field-mapped data stream. The first node then reads the missing value filling configuration from the data cleaning rules and fills in the null values ​​in the target field of the field-mapped data stream, generating the first intermediate data stream. The data stream manager passes the first intermediate data stream to the first downstream node in the node execution sequence. Upon receiving the first intermediate data stream from the upstream node, this downstream node reads the outlier truncation configuration from its bound data cleaning rules. Based on upper and lower thresholds, it performs boundary checks on each numerical field in the received data stream, replacing values ​​exceeding the upper threshold with the upper threshold value and values ​​below the lower threshold value with the lower threshold value, resulting in a truncated data stream.

[0029] The downstream node continues to read the type conversion configuration from the data cleaning rules, converting the source data type of each field in the truncated data stream to the target data type, thus obtaining the converted data stream. The data stream manager passes the converted data stream to the next downstream node in the node execution sequence. When the last node in the node execution sequence completes all the operations defined by its bound data cleaning rules, the final data stream is generated. The data stream manager then writes the final data stream back to the target storage location corresponding to the standardized initial dataset in the data storage layer, thus obtaining the cleaned data.

[0030] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Based on the cleaned data, the feature vectorization operator in the data processing pipeline is invoked to perform vector mapping on the cleaned data to obtain an initial feature vector set. Specifically, the feature vectorization operator is a processing node bound to the data processing pipeline for converting structured data into numerical vector representations. The feature vectorization operator first reads the field names and field values ​​of each field in the cleaned data and executes different vectorization strategies according to the field type. For numerical fields, the original value is retained as the vector component of the field, and the value of the component is the field value itself. For text fields, the text content is segmented into words by a preset word segmenter to obtain a word sequence containing multiple words. Then, each word in the word sequence is mapped to a word vector of fixed dimension. The arithmetic mean of each dimension of all word vectors of the field is taken to obtain the vector component of the text field.

[0031] For categorical fields, obtain the set of all possible values ​​for the field, assign one binary bit to each category in the set, set the binary bit corresponding to the category of the current data row's field value to 1, and set the remaining binary bits to 0, forming the vector component of the categorical field; concatenate the vector components of each field in each data row according to the field order to form the original feature vector of the data row; let the original feature vector be represented as... ,in Indicates the sequence number of the data row. The values ​​are positive integers from 1 to the total number of data rows M. The original feature vectors of all data rows form the initial feature vector set. , recorded as .

[0032] Step 2.2: Based on the initial feature vector set, the pre-constructed implicit topological metric space is invoked to map the initial feature vector set to a non-Euclidean state potential energy surface within the implicit topological metric space, obtaining the state potential energy surface mapping result. Specifically, the pre-constructed implicit topological metric space is a computational framework built based on a priori general data distribution during the system initialization phase. It internally defines the topological structure of the non-Euclidean state potential energy surface, used for potential field modeling of feature vectors in a non-Euclidean geometric space; this non-Euclidean state potential energy surface is pre-arranged with... There are 32 to 128 reference points (typically ranging from 32 to 128, determined based on the number of business semantic tags N; it is recommended that K ≥ 2N), and each reference point is denoted as . ,in Indicates the reference point number. The value ranges from 1 to Positive integers, for each reference point It has surface coordinates that fit the non-Euclidean potential energy surface; and the initial eigenvector set. Each original feature vector in As input, a pre-defined distance metric function within the implicit topological metric space is invoked to compute the original feature vector. With each reference point Geodetic distance between This geodesic distance simulates the shortest path length characteristics along the surface morphology of a non-Euclidean potential energy surface. In flat regions, it is equivalent to the Euclidean distance, while in curved regions, it reflects the amplification effect of surface curvature on the distance.

[0033] In a preferred embodiment, the distance metric function calculates the geodesic distance using the following expression:

[0034] This distance metric function has the following property: a low curvature limit, when the curvature at the reference point... When it approaches 0, Approaching The distance is precisely degenerated into Euclidean distance; the high curvature characteristic means that for the same Euclidean distance, the greater the curvature, the greater the geodesic distance, reflecting the influence of the curvature of non-Euclidean surfaces on the shortest path length; monotonicity means that the geodesic distance increases monotonically with the Euclidean distance, ensuring the consistency of nearest neighbor selection; among them, Represents the original feature vector With reference point The Euclidean distance between them Indicates reference point The local curvature value of the surface at that point. The preset curvature modulation coefficient, The value ranges from 0 to 1; The method for determining is as follows: During the system initialization phase, calculate the arithmetic mean of the local curvature values ​​of the surface at all K reference points on the non-Euclidean potential energy surface, denoted as . To obtain the maximum value of the local curvature of the surface at all reference points, denoted as . , curvature modulation coefficient The initial value is set to This allows the geodesic distance to be moderately amplified in high curvature regions and to approach the Euclidean distance in low curvature regions. The expression nonlinearly adjusts the Euclidean distance through a curvature correction term in multiplicative form, so that the geodesic distance is moderately amplified in high curvature regions and to accurately approach the Euclidean distance in low curvature regions.

[0035] Based on the calculated total geodesic distances, select the original feature vector. The surface coordinates corresponding to the reference point with the smallest geodesic distance are taken as the potential energy projection position of the original eigenvector on the non-Euclidean potential energy surface, denoted as . ; the original feature vector Mapped to potential energy projection position This yields the potential energy surface mapping points of the original eigenvectors; after all original eigenvectors have been mapped, a set of potential energy distribution points of the eigenvectors is formed on the non-Euclidean potential energy surface. , recorded as , The state potential energy surface mapping result is obtained.

[0036] Step 2.3: Based on the state potential energy surface mapping results, using business semantic tags as prototype attractors, perform adaptive semantic field discretization processing to construct a semantic phase transition hypersurface boundary, obtaining an intermediate analysis set with business semantic tags and feature vectorization operator parameters. Specifically, this includes: obtaining the business semantic tag set based on the state potential energy surface mapping results. , recorded as ,in Indicates the first Individual business semantic tags, The value ranges from 1 to the total number of business semantic tags. Positive integers; each business semantic tag As prototype attractors, an initial potential well center position is assigned to each prototype attractor on the non-Euclidean potential energy surface, denoted as . Perform adaptive semantic field discretization processing for the potential energy distribution point set. Each potential energy distribution point in Calculate its position relative to the center of each prototype attractor potential well. The potential energy function values ​​between the points of potential energy distribution. With the center position of the potential well The geodesic distance between them is a variable.

[0037] In a preferred embodiment, the potential energy function Calculate using the following expression:

[0038] In the formula, Represents the potential energy distribution point With the center position of the potential well The geodesic distance between them is calculated in the same way as the expression of the distance metric function in step 2.2; It is an exponential function; For the prototype attractor The corresponding potential well width parameter, Greater than 0, with the dimension being the reciprocal of the square of the distance, used to control the attraction range of the prototype attractor. This dimension is defined to match the distance dimension of the geodesic distance, ensuring that the exponential term of the potential energy function is a dimensionless pure numerical value. The preset potential energy amplitude coefficient, A value greater than 0 and a dimensionless constant is determined by calculating the average geodesic distance from all potential energy distribution points to their nearest potential well center during the system initialization phase, denoted as . The potential energy amplitude coefficient The initial value is set to (Or, if necessary, take other positive dimensionless constants) such that the potential energy function value is a dimensionless relative potential energy measure; the expression makes the potential energy function value reach its minimum at the center of the potential well. As the geodetic distance increases, it approaches 0, forming a potential energy trap with the center of the potential trap as the lowest point.

[0039] For each potential energy distribution point Perform the following operations: Obtain the surface coordinates of the potential energy distribution point on the non-Euclidean potential energy surface, and iterate sequentially from 1 to N according to the index of the prototype attractor. For the first... A prototype attractor is used to obtain the center position of its potential well. The surface coordinates are used; the distance metric function described in step 2.2 is called to calculate the potential energy distribution point. The surface coordinates and the center position of the potential well The geodesic distance between the surface coordinates is used to calculate the potential energy distribution point by substituting this geodesic distance into the potential energy function expression. Relative to the prototype attractor The potential energy function value; after traversing all N prototype attractors, obtain N potential energy function values ​​corresponding to the potential energy distribution point. Sort these N potential energy function values ​​in ascending order of numerical value, find the potential energy function value with the smallest value, and determine the prototype attractor corresponding to the smallest potential energy function value; the potential energy distribution point The point is assigned to the semantic field corresponding to the prototype attractor, and is also the potential energy distribution point. Add a field affiliation identifier, which records the business semantic label name corresponding to the prototype attractor; after completing the above allocation operation for each potential energy distribution point in the potential energy distribution point set P, each potential energy distribution point is assigned to a unique semantic field, and the preliminary semantic field partitioning result is obtained.

[0040] After all potential energy distribution points are divided, a semantic phase transition hypersurface boundary is constructed. First, the adjacency relationships between all semantic fields are identified. The identification method is as follows: traverse all pairs of semantic fields. For any pair of semantic fields, obtain the set of surface coordinates of all potential energy distribution points contained in the first semantic field and the set of surface coordinates of all potential energy distribution points contained in the second semantic field. Extract surface coordinates one by one from the set of surface coordinates of the first semantic field and calculate the geodesic distance from this surface coordinate to each surface coordinate in the set of surface coordinates of the second semantic field. If the geodesic distance between at least one pair of surface coordinates is less than a preset neighborhood determination radius (typically 0.1 to 0.3 times the average potential well width), then these two semantic fields are determined to be adjacent semantic fields. For all adjacent semantic field pairs, perform the following processing: obtain the center position of the prototype attractor potential well corresponding to the first semantic field in the adjacent semantic field pair. The center position of the prototype attractor potential well corresponding to the second semantic field ; in connection and On the geodesic segment, uniform sampling is performed with a set spatial step size to generate a densely arranged sequence of sampling points. For each sampling point in the sampling point sequence, the distance from that sampling point to the center of the first prototype attractor potential well is calculated. The potential energy function value, and the distance from the sampling point to the center of the second prototype attractor potential well. The potential energy function value.

[0041] The two potential energy function values ​​are further compared. If the absolute value of the difference between them is less than a preset equivalence threshold (typically 0.01 to 0.05β, where β is the potential energy amplitude coefficient), the sampling point is recorded as a candidate point for the potential energy isosurface between adjacent semantic field pairs. After processing all sampling points in the sampling point sequence, all candidate points for potential energy isosurfaces are connected in chronological order according to their spatial location along the geodesic direction to form a continuous potential energy isosurface curve. This potential energy isosurface curve is then extended along a surface direction perpendicular to the geodesic by a set extension. The step size is expanded to form a surface patch covering the boundary region between two fields. This surface patch is the potential energy isosurface between adjacent semantic field pairs. The potential energy isosurfaces between all adjacent semantic field pairs are summarized to form a set of surfaces covering the boundary region of all semantic fields. This set of surfaces constitutes the semantic phase transition hypersurface boundary. Each surface patch in the semantic phase transition hypersurface boundary precisely defines the boundary position of the corresponding two business semantic tags on the state potential energy surface. Any point falling on this surface patch is equidistant from the two adjacent prototype attractors in the sense of potential energy function.

[0042] The potential energy distribution points are clustered according to their semantic field affiliation. All semantic fields are traversed, and for each semantic field, an empty cluster set corresponding to that semantic field is created. Each potential energy distribution point in the potential energy distribution point set P is traversed, and its field affiliation identifier is read. Based on the business semantic label name recorded in the field affiliation identifier, the potential energy distribution point is added to the cluster set corresponding to that business semantic label. After all potential energy distribution points are classified, each cluster set contains the surface coordinates and original feature vector indices of all potential energy distribution points belonging to the same semantic field. Simultaneously, the operator configuration parameters used in this feature vectorization process are recorded, including the final potential well center position coordinates after iterative updates of each prototype attractor, the potential well width parameter value corresponding to each prototype attractor, the potential energy amplitude coefficient value, the iterative convergence threshold, and the actual number of iterations. Each cluster set, its associated business semantic labels, and the recorded operator configuration parameters are merged and encapsulated to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters.

[0043] Step 2.4: Based on the intermediate analysis set, extract the state transition topological associations and information entropy gradient distribution of the intermediate analysis set to obtain a structured feature description of the intermediate analysis set. Specifically, this includes: extracting the state transition topological associations and information entropy gradient distribution of the intermediate analysis set with business semantic labels; wherein, the state transition topological association refers to the migration path and adjacency relationship of potential energy distribution points between different semantic fields within the intermediate analysis set. When extracting the state transition topological association, traverse the adjacent semantic field pairs on both sides of the semantic phase transition hypersurface boundary. For adjacent semantic fields... and ,in and For the sequence number identifier of the semantic field and ; Count the potential energy distribution point pairs that appear within the preset neighborhood on both sides of the boundary. If one of the distribution points in the potential energy distribution point pair belongs to the semantic field... Furthermore, the other distribution point belongs to the semantic field. If the distribution point pair is marked as a set of migration paths, the migration direction is from the field with higher potential energy value to the field with lower potential energy value, and the migration frequency is the number of distribution point pairs in the migration direction; by summing up the migration paths, migration directions and migration frequencies of all adjacent semantic field pairs, the state transition topology association is obtained.

[0044] Information entropy gradient distribution refers to the rate of change of the uncertainty measure of potential energy distribution points in each semantic field within the intermediate analysis set along the spatial direction. When calculating the information entropy gradient distribution, for each semantic field... Statistically analyze the frequency distribution of potential energy distribution points within the field belonging to each original business semantic tag; let the semantic field be... The total number of potential energy distribution points included is Among them, the business semantic tags The number of distribution points is Then business semantic tags In the semantic field The frequency percentage in Calculate the information entropy value of this semantic field. It is expressed by the following formula:

[0045] In the formula, Represents the semantic field All business semantic tags appearing within Summation, This represents a logarithmic function with base 2, where... This item is defined when it is 0. The value is 0; for adjacent semantic fields and Calculate the information entropy difference between the two fields. The information entropy difference is then divided by the geodesic distance between the centers of the two fields. The information entropy gradient value is obtained; the information entropy gradient values ​​of all adjacent semantic field pairs are summarized to obtain the information entropy gradient distribution; the extracted state transition topological associations and information entropy gradient distributions are combined to form a structured feature description of the intermediate analysis set.

[0046] Step 2.5: Based on the structured feature description of the intermediate analysis set, obtain the corresponding semantic routing rules. Specifically, this includes: traversing the state transition topology associations in the structured feature description, sequentially reading the starting semantic field, ending semantic field, and migration frequency of each migration path; using the business semantic label corresponding to the starting semantic field as the source semantic label in the routing conditions, using the business semantic label corresponding to the ending semantic field as the target semantic label in the routing conditions, and normalizing the migration frequency as the routing priority weight. The routing priority weight ranges from 0 to 1, with a larger value indicating a higher priority for the migration path in the routing decision; combining the source semantic label, target semantic label, and routing priority weight into a single routing rule entry. All such routing rule entries constitute the first type of semantic routing rules.

[0047] The information entropy gradient distribution in the structured feature description is traversed, and the information entropy gradient value of each pair of adjacent semantic fields is read. This information entropy gradient value is compared with a preset gradient threshold, which is used to determine whether the difference in information uncertainty between adjacent semantic fields reaches the limit to trigger a routing decision. The preset gradient threshold is set based on the following: during the system initialization phase, the information entropy gradient values ​​of all adjacent semantic field pairs in the information entropy gradient distribution are statistically analyzed, and the arithmetic mean and standard deviation of all information entropy gradient values ​​are calculated. The arithmetic mean plus one standard deviation is used as the initial value of the preset gradient threshold. This value ensures that the field boundary located in the upper-middle position of the overall gradient distribution is included in the triggering condition. In one example, the preset gradient threshold is set to 0.15. If the information entropy gradient value exceeds the preset gradient threshold, the boundary identifier of the adjacent semantic field pair is used as the semantic boundary triggering condition in the routing condition. This semantic boundary triggering condition is used to trigger a routing decision when the semantic label of the intent parsing result crosses the boundary. All such semantic boundary triggering conditions constitute the second type of semantic routing rule. The first type of semantic routing rule and the second type of semantic routing rule are merged to obtain the corresponding semantic routing rule.

[0048] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Based on the semantic routing rules, parse the routing conditions and target actions in the semantic routing rules to obtain the routing rule parsing results. Specifically, this includes: parsing the routing conditions and target actions contained in each rule entry of the semantic routing rules; for the first type of semantic routing rules, extracting the source semantic tags in each rule entry as intent triggering conditions in the routing conditions, extracting the target semantic tags as intent matching targets in the routing conditions, and extracting the routing priority weights as priority sorting criteria in the routing conditions; for the second type of semantic routing rules, extracting the semantic boundary triggering conditions in each rule entry as boundary activation conditions in the routing conditions; during the parsing process, grouping and organizing various types of routing conditions according to categories, including intent matching and boundary activation categories, registering the execution actions corresponding to the target semantic tags and the execution actions corresponding to the semantic boundary triggering conditions as target actions, including switching analysis topics, calling data query interfaces, triggering model inference calculations, and refreshing visualization views; combining the grouped and organized routing conditions with the corresponding target actions to obtain the routing rule parsing results.

[0049] Step 3.2: Based on the routing rule parsing results, configure the intent recognition node in the preset low-code orchestration canvas, and simultaneously set the intent category and triggering conditions corresponding to the intent recognition node to obtain the intent recognition node configuration. Specifically, this includes: This preset low-code orchestration canvas is the core operation interface for the system to provide visual node orchestration capabilities, including a canvas editing area, a node configuration panel, and a node library panel; In the node configuration panel of the preset low-code orchestration canvas, select an intent recognition node template from the node library panel, create a new intent recognition node, and add the intent recognition node to the canvas editing area; This intent recognition node is the processing node in the agent decision graph responsible for receiving natural language queries and determining the query intent type; Extract all routing conditions for intent matching classes from the routing rule parsing results, register the source semantic label in each routing condition as an intent category, and after all intent categories are registered, add an additional fallback intent category to handle situations where the user intent cannot be clearly identified.

[0050] All registered intent categories and fallback intent categories are aggregated to form an intent category list. This intent category list is a structured data container within the intent recognition node used to store all identifiable intent types. Each entry in the list contains an intent category name and a description. The constructed intent category list is then bound to the attribute configuration of the intent recognition node, setting corresponding trigger conditions for each intent category. These trigger conditions define the semantic similarity requirements that must be met when the semantic parsing result of a natural language query matches that intent category. Specifically, the trigger conditions are set by establishing a minimum semantic similarity threshold; when the natural language query... When the semantic similarity of a query to a certain intent category reaches a minimum threshold after semantic parsing, the intent category is determined to be triggered. The minimum threshold for semantic similarity ranges from 0.65 to 0.75, with a default value of 0.7. This range is determined based on the semantic matching benchmark commonly used in the field of natural language processing, combined with the error tolerance requirements of actual question-answering scenarios. It controls the false trigger rate while ensuring the recall rate of intent recognition. Users can adjust the value within the range according to the accuracy requirements of the actual application scenario. After configuring all intent categories and their corresponding triggering conditions, the configuration parameters are saved to the attribute set of the intent recognition node to obtain the intent recognition node configuration.

[0051] Step 3.3: Based on the routing rule parsing results, configure tool invocation nodes in the preset low-code orchestration canvas, and simultaneously establish the association between the tool invocation nodes and the target actions in the semantic routing rules to obtain the tool invocation node configuration. Specifically, this includes: extracting all target actions from the routing rule parsing results; traversing each target action and reading its action type identifier; merging and classifying target actions with the same action type identifier or equivalent functions, based on the consistency of the system capability types invoked by the target actions; after classification, obtaining a list of tool invocation node types to be configured, including data query tool nodes, etc. The system includes a model inference tool node, a view refresh tool node, a theme switching tool node, and a fallback tool node. The fallback tool node handles user queries that do not match any predefined intent, performing query guidance, transferring to human customer service, or providing fallback results. The data query tool node calls query operators in the data processing pipeline to perform structured queries on the data storage layer. The model inference tool node calls model inference operators in the data processing pipeline to perform predictive calculations. The view refresh tool node triggers the visualization rendering engine to re-render the target display view. The theme switching tool node switches the context of the current analysis theme based on the routing target.

[0052] In the node configuration panel of the pre-built low-code orchestration canvas, create a corresponding tool call node instance for each type in the tool call node type list, and assign a unique node identifier to each tool call node instance; configure input parameters and output format for each tool call node instance. The input parameters define the data source identifier, query field list, and parameter value constraints that the tool call node needs to receive when it executes; the output format defines the field names, field types, and number of fields of the result data structure produced after the tool call node completes its execution; based on the correspondence between target actions and routing conditions in the routing rule parsing results, establish an association between each tool call node instance and the corresponding target action in the semantic routing rules. This association is achieved by recording the corresponding target action identifier and the semantic routing rule identifier in the attributes of the tool call node instance, which is used to guide the execution signal to the correct tool call node after the intent recognition node completes the intent judgment; save all configured tool call node instances and their associations to obtain the tool call node configuration.

[0053] Step 3.4: Based on the intent recognition node configuration and tool invocation node configuration, construct the execution dependency path between nodes to obtain the initial agent decision graph. Specifically, this includes: in the pre-set low-code orchestration canvas editing area, spatially arranging the intent recognition node and all tool invocation node instances according to functional logic, placing the intent recognition node in the left area of ​​the canvas editing area as the flow starting node, and arranging the tool invocation node instances in the right area of ​​the canvas editing area from top to bottom in the order of data query tool node, model inference tool node, view refresh tool node, and theme switching tool node; based on the correspondence between routing conditions and target actions in the routing rule parsing results, establishing directed connections from the intent recognition node to each tool invocation node instance one by one. An additional directed connection is established from the intent recognition node to the fallback tool call node instance, and the transmission condition of this directed connection is set to the intent category name of the fallback intent category. This fallback intent category is a general intent category predefined by the system, which is not bound to any business semantic tags and has a lower priority than all predefined business intents. The starting end of each directed connection is connected to the specified output port of the intent recognition node, and the ending end is connected to the specified input port of the corresponding tool call node instance.

[0054] For each directed connection edge, a transmission condition is set. The content of the transmission condition is the name of the intent category associated with the target action. The transmission condition is read by the execution engine at runtime. The execution engine compares the intention category name matched in the output of the intent recognition node with each transmission condition to determine which execution dependency path to activate. If multiple target actions are associated with the same intent category, the same transmission condition is set for each of the directed connections corresponding to that intent category, and a priority sorting number is attached. The priority sorting number is sorted according to the routing priority weight of each routing rule entry in the first type of semantic routing rules. The higher the weight, the earlier the sorting number. After all directed connections are established and the transmission conditions are configured, the initial agent decision graph is formed by the intent recognition node, all tool invocation node instances, and all directed connections.

[0055] Step 3.5: Based on the initial agent decision graph, perform logical verification on the execution dependency paths. After successful verification, an executable agent orchestration instance is obtained. Specifically, the logical verification process includes the following three verification items, which are executed sequentially: The first verification item is connectivity verification. Traverse all directed edges in the initial agent decision graph, starting from the intent recognition node, and perform a depth-first search along the direction of the directed edges to check if there is at least one directed path reaching each tool invocation node instance. If a tool invocation node instance has no valid path reachable after the search, then the tool invocation node instance is terminated. The first check is to mark a node instance as a connectivity defect and output a defect message to the user. The second check is to check for conflicts. The process involves traversing the transmission conditions of all directed edges in the initial agent decision graph and registering the content of each transmission condition in the transmission condition registration table. If, during the registration process, two or more directed edges are found to have the same transmission conditions and these directed edges are connected to different tool call node instances, the situation is marked as a transmission condition conflict and a conflict message is output to the user. The user can then choose to keep one of the directed edges or reassign different intent categories to each conflicting edge.

[0056] The third verification item is the completeness verification. Each target action in the routing rule parsing result is compared one by one with the tool call node instance and directed connection edge in the initial agent decision graph. It is checked whether each target action has a corresponding tool call node instance and whether the tool call node instance has a valid connection path. If there is an uncovered target action, the target action is marked as a completeness defect and a defect prompt message is output to the user. After all three verification items pass, the initial agent decision graph is marked as verified. The verified initial agent decision graph is merged and packaged with the routing rule parsing result, intent recognition node configuration and tool call node configuration into an orchestration instance description file. After the orchestration instance description file is parsed and loaded, an executable agent orchestration instance is formed.

[0057] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Based on the executable agent orchestration instance, receive natural language queries sent by external terminals to obtain query statements to be processed. Specifically, this includes: activating a pre-built query monitoring service within the executable agent orchestration instance, which continuously monitors natural language query requests sent by external terminals through a preset communication interface; when an external terminal submits a text-based natural language query to the system via an HTTP interface or a WebSocket interface, the query monitoring service captures the request message and extracts the query content field and request source identifier from the request message; preprocessing the extracted query content field, including removing leading and trailing whitespace characters, compressing multiple consecutive spaces into a single space, and converting full-width characters in the text to half-width characters; after preprocessing, a standardized query text is obtained, which is then used as the query statement to be processed.

[0058] Step 4.2: Based on the query statement to be processed, the intent recognition node in the executable agent orchestration instance is invoked to perform semantic parsing to obtain the intent parsing result. Specifically, after receiving the query statement to be processed, the intent recognition node first invokes the pre-set semantic encoder to perform semantic encoding on the query statement, converting the query statement into a fixed-dimensional semantic vector representation; further, the obtained semantic vector is compared with the reference semantic vector of each intent category in the intent category list configured by the intent recognition node, and the similarity calculation adopts the cosine similarity measure; all the calculated similarity values ​​are sorted in descending order, and the intent category with the highest similarity value and not lower than the minimum semantic similarity threshold is selected as the matching intent; the minimum semantic similarity threshold is used to limit the minimum confidence requirement for intent matching, and its setting is based on balancing the recall and precision of intent recognition. In one example, the minimum semantic similarity threshold is set to 0.7, and this value has been tested in multiple rounds of actual... After testing and verification in the question-and-answer scenario, it was determined that the system effectively filters out mismatches of semantically ambiguous or irrelevant queries while ensuring that most correct intents can be identified. If a matching intent exists, the intent category name and corresponding similarity value of the matching intent are output as the intent parsing result. If no matching intent exists, the intent category name of the fallback intent category is output as the intent parsing result. This fallback intent category is used to handle situations where the user's intent cannot be clearly identified. This intent parsing result will trigger the fallback tool call node to execute the preset fallback processing flow: return guidance information to the user that they could not understand their query and ask them to rephrase it or select one of the following common questions, and provide quick access to 3 to 5 high-frequency business questions; at the same time, the unidentified query statement, request source identifier, and timestamp are recorded in the unidentified intent dataset for subsequent iterative optimization of the intent recognition model and semantic routing rules; the output intent parsing result is recorded in the execution log to obtain the intent parsing result.

[0059] Step 4.3: Based on the intent parsing result, traverse the semantic routing rules in the executable agent orchestration instance, match the target semantic routing rule corresponding to the intent parsing result, and obtain the routing matching result. Specifically, this includes: extracting the intent category name of the matching intent from the intent parsing result, using the intent category name as the matching keyword, and sequentially traversing the semantic routing rules stored in the executable agent orchestration instance. For each routing rule entry in the first type of semantic routing rules, compare the matching keyword with the source semantic label in the routing conditions of the routing rule entry. If the matching keyword is completely consistent with the source semantic label of a routing rule entry, then mark the routing rule entry as the target semantic routing rule. If no matching rule entry is found in the first type of semantic routing rules, continue traversing the second type of semantic routing rules. For each rule entry, obtain the two adjacent semantic fields involved in the semantic boundary triggering condition of the rule entry, and determine whether the semantic field corresponding to the matching keyword is one of these two adjacent semantic fields. If the semantic field corresponding to the matching keyword is the same as either of these two adjacent semantic fields, obtain the semantic vector of the query statement to be processed from step 4.2, and map the semantic vector to the non-Euclidean potential energy surface through the mapping method in step 2.2. Calculate the geodesic distance from the mapping point to the center of the potential well of the prototype attractor of each of the two adjacent semantic fields. If the geodesic distance from the mapping point to the center of the potential well of the semantic field on the other side of the boundary is less than the geodesic distance to the center of the potential well of the current semantic field, it is determined that the semantic label crosses the boundary, and the routing rule entry is marked as the target semantic routing rule. Summarize all marked target semantic routing rules to obtain the routing matching result.

[0060] Step 4.4: Based on the routing matching results, activate the target semantic routing rule, determine the associated operators to be triggered in the data processing pipeline according to the target semantic routing rule, and obtain the associated operator trigger list. Specifically, this includes: reading all target semantic routing rule entries from the routing matching results, sorting them from high to low according to the routing priority weight of each entry, and selecting the one with the highest priority as the currently activated rule; after activating the target semantic routing rule, reading the target action identifier from the rule entry, and searching for the tool call node instance associated with the target action in the agent decision graph according to the target action identifier; reading the input parameters from the configuration parameters of the tool call node instance, including the name of the dataset to be queried, the list of query fields, and the query condition template; locating the operator node chain associated with the dataset in the node topology of the data processing pipeline according to the dataset name specified in the input parameters, extracting the node identifiers of all operators in the operator node chain, arranging all extracted operator node identifiers in the node topology order, and generating the associated operator trigger list.

[0061] Step 4.5: Based on the association operator trigger list, sequentially trigger the corresponding association operators to perform targeted queries and model inference on the intermediate analysis set to obtain a structured analysis result set. Specifically, this includes: based on the association operator trigger list, calling the pipeline execution engine to sequentially trigger the corresponding association operators according to the order of the operator node identifiers in the list; the pipeline execution engine assigns an independent execution thread to each operator node in the trigger list and starts each execution thread sequentially according to the list order; for each triggered association operator, inputting the potential energy distribution point data in the semantic field associated with the operator in the intermediate analysis set, as well as the query parameters parsed from the query condition template. After receiving input, the association operator performs a targeted query operation, filtering out data records that meet the query conditions from the data storage layer. After the targeted query is completed, the association operator continues to perform model inference operation, calling the pre-trained model bound to the operator to perform inference calculations on the filtered data records and obtain the inference result of the operator. After all association operators have been executed, the inference results of all association operators are merged in the order of triggering. During the merging, the result data with the same field name are aligned and concatenated. The merged data is organized according to a preset structured format to form a structured analysis result set containing three components: a query result data table, inference conclusion text, and confidence score.

[0062] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the structured analysis result set, call the visualization rendering engine to generate the target display view, collect user interaction feedback data for the target display view, and obtain the user interaction feedback dataset. Specifically, this visualization rendering engine is the core rendering module responsible for converting structured data into graphical view components. It contains a chart template matcher, a component renderer, and a layout assembler. The chart template matcher is used to automatically select the most suitable chart type based on the type and number of data fields. The component renderer is used to bind data to the selected chart template and generate corresponding visualization component instances. The layout assembler is used to arrange multiple visualization components in space according to the layout template.

[0063] The visualization rendering engine reads the three components of the structured analysis result set: the query result data table, the inference conclusion text, and the confidence score. Based on the field structure of the query result data table, it calls the chart template matcher to automatically match the corresponding chart type template, mapping numerical fields to the vertical axis data of the chart and categorical fields to the horizontal axis labels, generating visualization chart components. The inference conclusion text is rendered as a text interpretation component, and the confidence score is rendered as an indicator card component. The layout assembler is then called to combine and arrange the visualization chart components, text interpretation components, and indicator card components according to a preset dashboard layout template. This preset dashboard layout template is a pre-defined view layout framework used to define the spatial position and size ratio of different types of components on the display page. The layout template divides the display page into a top indicator card area, a middle chart area, and a bottom interpretation area. The top indicator card area is used to place indicator card components such as the confidence score, the middle chart area is used to place visualization chart components, and the bottom interpretation area is used to place text interpretation components such as the inference conclusion text. After the combination and arrangement are completed, a target display view is generated and pushed to the user terminal interface for presentation.

[0064] During the user's browsing of the target display view, user interaction feedback data is collected. The interaction feedback data includes user click confirmation actions on chart components, user acceptance marks for inference conclusions, user switching of analysis dimensions, and user manually entered corrections. All collected interaction feedback data is recorded in chronological order by timestamp to form a user interaction feedback dataset.

[0065] Step 5.2: Based on the user interaction feedback dataset, extract the user's adoption and correction operation records for the target display view to obtain the feedback signal feature vector. Specifically, this includes: reading the interaction type field and operation content field of each interaction feedback record one by one according to the user interaction feedback dataset; marking records with the interaction type of click confirmation as positive adoption behavior, marking records with the interaction type of adoption mark as positive adoption behavior, marking records with the interaction type of analysis dimension switching as correction operation behavior, and extracting the target dimension name after the user's switch from the operation content field; marking records with the interaction type of manual input of correction content as correction operation behavior, and extracting the correction text input by the user from the operation content field; for all records of positive adoption behavior, statistically analyzing the distribution frequency of the corresponding business semantic tags in the intermediate analysis set involved, forming an adoption tag frequency vector; for all records of correction operation behavior, statistically analyzing the distribution frequency of the business semantic tags corresponding to the correction targets involved, forming a correction tag frequency vector; concatenating the adoption tag frequency vector and the correction tag frequency vector according to the set tag order to generate the feedback signal feature vector.

[0066] Step 5.3: Based on the feedback signal feature vector, calculate the potential energy gradient offset of the feedback signal feature vector relative to the semantic phase transition hypersurface boundary to obtain the boundary offset calibration vector. Specifically, this includes: denoting the feedback signal feature vector as... ,in The index of the eigenvector of the feedback signal. The dimension is equal to twice the total number of business semantic tags N; due to the feedback signal feature vector Each component of the feature vector corresponds to the adoption frequency and correction frequency of each business semantic tag, while each component of the original feature vector corresponds to the field values ​​of the dataset. The two have different dimensions. To enable the feedback signal to reuse the non-Euclidean potential energy surface constructed in step 2.2, a dimension-aligned projection matrix W is constructed. This matrix has d rows and 2N columns. The element in the u-th row and v-th column of the matrix is ​​the u-th component value of the center position coordinate of the prototype attractor potential well corresponding to the v-th business semantic tag recorded in step 2.3. When v is greater than N, the element in the v-th column corresponds to the u-th component value of the center position coordinate of the potential well of the vN-th business semantic tag, so that the first N columns and the last N columns correspond to the projection basis of the adoption frequency component and the correction frequency component of each business semantic tag, respectively. The feature vector of the feedback signal is then... Perform matrix multiplication with the dimension-aligned projection matrix W = W ,Will Mapping from 2N-dimensional space to d-dimensional space yields the dimension-aligned feature vector of the feedback signal. ; The dimensionally aligned feature vector of the feedback signal Mapping onto a non-Euclidean potential energy surface, we obtain the location of the mapping point on the potential energy surface, denoted as . Obtain the position information of all boundary surface patches on the semantic phase transition hypersurface boundary. For each boundary surface patch, calculate the mapping point. The nearest distance to the boundary surface patch, which is the mapping point. The shortest geodesic path length from the potential energy surface to any point on the boundary surface patch; among all boundary surface patches, select the mapping point. The boundary surface patch with the smallest nearest distance is selected as the target boundary surface patch, and the mapping point is determined on the target boundary surface patch. The point with the smallest nearest distance is denoted as the boundary reference point. .

[0067] Calculate mapping points The local potential gradient direction vector at point is denoted as . First determine the mapping point. The most relevant prototype attractor potential well center Even if the potential energy function value is obtained The center of the potential well that achieves the minimum value among all prototype attractors; then the gradient direction vector. The calculation formula is as follows:

[0068] in Represents the mapping point To the center of the corresponding potential well geodetic distance, The geodesic distance function represents the distance at a point. The gradient direction vector at a point with respect to the surface coordinates. This represents the potential well width parameter corresponding to the prototype attractor. This is the preset potential energy amplitude coefficient; It is an exponential function; for the boundary reference point Using mapping points The same method is used to determine the center of its most relevant prototype attractor potential well. And calculate the boundary reference point. The local potential gradient direction vector at [location] Its calculation expression only requires the above In the expression Replace with That's it; map the point. gradient direction vector at the boundary reference point Subtract the gradient direction vectors at the given location to obtain the potential energy gradient offset. The potential energy gradient offset The direction represents the offset direction of the feedback signal feature vector relative to the boundary of the semantic phase transition hypersurface, and its magnitude is... Indicates the degree of offset, representing the offset of the potential energy gradient. Output as a boundary offset calibration vector.

[0069] Step 5.4: Based on the boundary offset calibration vector, adjust the eigenvectorization operator parameters in the gradient direction to obtain the updated eigenvectorization operator parameters. Specifically, this includes: obtaining the operator configuration parameters recorded in Step 2.3 used in this eigenvectorization process, including the final potential well center position coordinates of each prototype attractor, the potential well width parameter value corresponding to each prototype attractor, and the potential energy amplitude coefficient value; for the potential well center position of each prototype attractor... The surface coordinates of the center position of the potential well are compared with the boundary offset calibration vector. By performing vector addition, the adjusted potential well center position is obtained. ;in Adjust the step size coefficient for the preset parameters. The value ranges from 0 to 1, and it is determined by calculating the magnitude of the boundary offset calibration vector during each iteration update. With respect to the current prototype attractor potential well width parameter The ratio of , if the ratio is greater than 1, then Set to 0.3 to limit the amplitude of a single adjustment to no more than the width of the potential well. This ensures the stability of the iteration process; if the ratio is not greater than 1, then... Set to 0.5.

[0070] For the potential well width parameter of each prototype attractor Calculate the boundary offset calibration vector Length of the module Geodesic distance from the center of the prototype attractor potential well to the target boundary surface patch The ratio between them yields the width adjustment coefficient. The potential well width parameter is adjusted according to the width adjustment coefficient to obtain the adjusted potential well width parameter. in For the first Width adjustment coefficients corresponding to each prototype attractor; potential energy amplitude coefficients in the eigenvectorization operator parameters. The average value of the updated potential well width parameter is synchronously adjusted to match the updated potential well width. The adjustment formula is as follows: ,in To adjust the mean value of the foreground well width parameter, The mean value of the potential well width parameter after adjustment; the convergence threshold and the actual number of iterations are not parameters that need to be adjusted in this iteration update and remain unchanged in this update; the center position of the potential well after adjusting all prototype attractors. Adjusted potential well width parameters and the potential energy amplitude coefficient after synchronization adjustment The parameters are summarized and combined with the iterative convergence threshold and the actual number of iterations to obtain the updated feature vectorization operator parameters.

[0071] Step 5.5: Based on the boundary offset calibration vector, adaptively correct the preset semantic routing rule matching threshold along the information entropy gradient descent direction to obtain the updated semantic routing rule matching threshold. Specifically, the preset semantic routing rule matching threshold is generated using the default configuration value when the agent low-code orchestration system is first deployed, or dynamically calibrated based on the initial data distribution in the first iteration, with subsequent iterations using the output value of the previous iteration as the update benchmark; obtain the information entropy gradient distribution calculated in step 2.4, and read the boundary offset calibration vector from the information entropy gradient distribution. The entropy gradient value corresponding to the direction closest to the entropy gradient direction is denoted as . Get the current preset semantic routing rule matching threshold, denoted as . ; calibrate vector based on boundary offset Modulus and information entropy gradient value The product determines the threshold correction amount. ;in This is the preset threshold correction coefficient. A value greater than 0 is determined by obtaining the standard deviation of all information entropy gradient values ​​in the information entropy gradient distribution during the system initialization phase, denoted as . To obtain the median of the information entropy gradient distribution, denoted as... ,Will The initial value is set to This allows the threshold correction amount to be adjusted proportionally according to the dispersion of the information entropy gradient.

[0072] If the boundary offset calibration vector If the direction points outward from the boundary of the semantic phase transition hypersurface, then the threshold correction amount will be... Compared with the current threshold Add them together to get the semantic routing rule matching threshold updated in the outward direction. If the boundary offset calibration vector If the direction points inside the semantic phase transition hypersurface boundary, then the current threshold will be... Subtract threshold correction amount The semantic routing rule matching threshold updated in the inward direction is obtained. Replace the current threshold with the updated semantic routing rule matching threshold. .

[0073] Step 5.6: Based on the updated feature vectorization operator parameters and the updated semantic routing rule matching threshold, replace the corresponding configuration items in the executable agent orchestration instance to obtain the optimized agent low-code orchestration model. Specifically, this includes: replacing the corresponding configuration items in the executable agent orchestration instance based on the updated feature vectorization operator parameters and the updated semantic routing rule matching threshold; writing the updated feature vectorization operator parameters back to the operator parameter storage area of ​​the feature semantic mapping module in the executable agent orchestration instance, overwriting the original operator configuration parameters; writing the updated semantic routing rule matching threshold back to the threshold configuration area of ​​the intent routing execution module in the executable agent orchestration instance, overwriting the original matching threshold parameters; after completing the above two configuration replacements, reload the configuration description file of the executable agent orchestration instance so that the updated operator parameters and matching thresholds take effect immediately in subsequent orchestration instance runs; store the executable agent orchestration instance after configuration replacement as the optimized agent low-code orchestration model and assign it a new version identifier to distinguish the orchestration model versions before and after optimization.

[0074] like Figure 2 As shown, embodiments of the present invention also provide an intelligent agent low-code orchestration system for multi-source data, comprising: The data pipeline construction module is used to construct a data processing pipeline by dragging and dropping operators in a preset low-code orchestration canvas based on a pre-acquired standardized initial dataset; and to perform data cleaning on the standardized initial dataset according to the data processing pipeline to obtain cleaned data. The feature semantic mapping module is used to perform feature vectorization processing on the cleaned data in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; and to obtain the corresponding semantic routing rules based on the business semantic labels of the intermediate analysis set. The decision graph orchestration and verification module is used to configure intent recognition nodes and tool invocation nodes in a preset low-code orchestration canvas according to the semantic routing rules to obtain an agent decision graph; and to perform logical verification on the agent decision graph to obtain an executable agent orchestration instance. The intent routing execution module is used to perform intent parsing on the received natural language query based on the executable intelligent agent orchestration instance, obtain the intent parsing result, activate the corresponding semantic routing rule based on the intent parsing result, so as to trigger the association operator in the data processing pipeline to perform targeted query and model inference on the intermediate analysis set, and obtain the structured analysis result set. The feedback iterative optimization module is used to obtain the target display view based on the structured analysis result set; collect user interaction feedback data on the target display view; and iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold based on the user interaction feedback data to obtain the optimized agent low-code orchestration model.

[0075] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0076] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A low-code orchestration method for intelligent agents oriented towards multi-source data, characterized in that, The method includes: Step 1: Based on the pre-acquired standardized initial dataset, construct a data processing pipeline in a pre-set low-code orchestration canvas by dragging and dropping operators; perform data cleaning on the standardized initial dataset according to the data processing pipeline to obtain cleaned data; Step 2: Based on the cleaned data, feature vectorization processing is performed in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; based on the business semantic labels of the intermediate analysis set, the corresponding semantic routing rules are obtained. Step 3: Based on the semantic routing rules, configure intent recognition nodes and tool invocation nodes in the preset low-code orchestration canvas to obtain the agent decision graph; perform logical verification on the agent decision graph to obtain an executable agent orchestration instance; Step 4: Based on the executable intelligent agent orchestration instance, perform intent parsing on the received natural language query to obtain the intent parsing result. Activate the corresponding semantic routing rule based on the intent parsing result to trigger the association operator in the data processing pipeline to perform targeted querying and model reasoning on the intermediate analysis set to obtain the structured analysis result set. Step 5: Based on the structured analysis result set, obtain the target display view; collect user interaction feedback data for the target display view; based on the user interaction feedback data, iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold to obtain the optimized agent low-code orchestration model.

2. The low-code orchestration method for intelligent agents oriented towards multi-source data according to claim 1, characterized in that, Based on a pre-acquired standardized initial dataset, a data processing pipeline is built in a pre-built low-code orchestration canvas by dragging and dropping operators; The standardized initial dataset is cleaned according to the data processing pipeline to obtain cleaned data, including: The standardized initial dataset is parsed to obtain its field structure and data type; based on the field structure and data type, the corresponding target operator is matched from the preset low-code orchestration canvas operator library to obtain the operator matching result; Based on the operator matching results, the target operator is added to the canvas editing area by dragging and dropping, and the connection relationship between the operators is established to obtain the initial data processing pipeline; Based on the initial data processing pipeline, obtain the parameter configuration items of each operator in the initial data processing pipeline, set the data cleaning rules according to the parameter configuration items, and obtain the configured data processing pipeline. According to the configured data processing pipeline, a data cleaning operation is performed on the standardized initial dataset to obtain cleaned data.

3. The low-code orchestration method for intelligent agents oriented towards multi-source data according to claim 2, characterized in that, Based on the cleaned data, feature vectorization processing is performed in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; based on the business semantic labels of the intermediate analysis set, corresponding semantic routing rules are obtained, including: Based on the cleaned data, the feature vectorization operator in the data processing pipeline is invoked to perform vector mapping on the cleaned data to obtain an initial feature vector set; Based on the initial feature vector set, the pre-constructed implicit topological metric space is invoked to map the initial feature vector set to the non-Euclidean state potential energy surface within the implicit topological metric space, thereby obtaining the state potential energy surface mapping result. Based on the state potential energy surface mapping results, using business semantic labels as prototype attractors, adaptive semantic field discretization processing is performed to construct semantic phase transition hypersurface boundaries, and intermediate analysis sets and feature vectorization operator parameters with business semantic labels are obtained. Based on the intermediate analysis set, the state transition topological correlation and information entropy gradient distribution of the intermediate analysis set are extracted to obtain a structured feature description of the intermediate analysis set. Based on the structured feature description of the intermediate analysis set, the corresponding semantic routing rules are obtained.

4. The low-code orchestration method for intelligent agents oriented towards multi-source data according to claim 3, characterized in that, Based on the semantic routing rules, intent recognition nodes and tool invocation nodes are configured in a preset low-code orchestration canvas to obtain an agent decision graph; Logical verification of the agent decision graph yields executable agent orchestration instances, including: Based on the semantic routing rules, the routing conditions and target actions in the semantic routing rules are parsed to obtain the routing rule parsing results; Based on the routing rule parsing results, intent recognition nodes are configured in the preset low-code orchestration canvas, and the intent category and triggering conditions corresponding to the intent recognition nodes are set to obtain the intent recognition node configuration. Based on the routing rule parsing results, tool call nodes are configured in the preset low-code orchestration canvas, and the association between the tool call nodes and the target actions in the semantic routing rules is established to obtain the tool call node configuration; Based on the intent recognition node configuration and the tool call node configuration, the execution dependency path between nodes is constructed to obtain the initial agent decision graph; Based on the initial agent decision graph, the execution dependency path is logically verified, and an executable agent orchestration instance is obtained after the verification is passed.

5. The low-code orchestration method for intelligent agents oriented towards multi-source data according to claim 4, characterized in that, Based on the executable agent orchestration instance, the received natural language query is parsed to obtain the intent parsing result. The corresponding semantic routing rule is activated based on the intent parsing result to trigger the association operator in the data processing pipeline to perform targeted querying and model inference on the intermediate analysis set, resulting in a structured analysis result set, including: Based on the executable intelligent agent orchestration instance, a natural language query sent by an external terminal is received to obtain a query statement to be processed; Based on the query statement to be processed, the intent recognition node in the executable intelligent agent orchestration instance is invoked to perform semantic parsing and obtain the intent parsing result; Based on the intent parsing result, the semantic routing rules in the executable agent orchestration instance are traversed, and the target semantic routing rule corresponding to the intent parsing result is matched to obtain the routing matching result; Based on the routing matching results, the target semantic routing rule is activated, and the associated operators to be triggered in the data processing pipeline are determined based on the target semantic routing rule to obtain the associated operator trigger list; Based on the association operator trigger list, the corresponding association operators are triggered sequentially to perform targeted queries and model inference on the intermediate analysis set, thereby obtaining a structured analysis result set.

6. The low-code orchestration method for intelligent agents oriented towards multi-source data according to claim 5, characterized in that, Based on the structured analysis result set, a target display view is obtained; user interaction feedback data on the target display view is collected; based on the user interaction feedback data, the feature vectorization operator parameters and the preset semantic routing rule matching threshold are iteratively updated to obtain an optimized agent low-code orchestration model, including: Based on the structured analysis results set, the visualization rendering engine is invoked to generate the target display view, and user interaction feedback data for the target display view is collected to obtain the user interaction feedback dataset; Based on the user interaction feedback dataset, extract the user's adoption behavior and correction operation records for the target display view to obtain the feedback signal feature vector; Based on the feedback signal feature vector, the potential energy gradient offset of the feedback signal feature vector relative to the semantic phase transition hypersurface boundary is calculated to obtain the boundary offset calibration vector. Based on the boundary offset calibration vector, the parameters of the eigenvectorization operator are adjusted in the gradient direction to obtain the updated eigenvectorization operator parameters; Based on the boundary offset calibration vector, the preset semantic routing rule matching threshold is adaptively corrected along the direction of information entropy gradient descent to obtain the updated semantic routing rule matching threshold. Based on the updated feature vectorization operator parameters and the updated semantic routing rule matching threshold, the corresponding configuration items in the executable agent orchestration instance are replaced to obtain the optimized agent low-code orchestration model.

7. A low-code orchestration system for intelligent agents oriented to multi-source data, the system implementing the method as described in any one of claims 1 to 6, characterized in that, include: The data pipeline building module is used to build a data processing pipeline by dragging and dropping operators in a pre-built low-code orchestration canvas based on a pre-acquired standardized initial dataset. The standardized initial dataset is cleaned according to the data processing pipeline to obtain cleaned data. The feature semantic mapping module is used to perform feature vectorization processing on the cleaned data in the data processing pipeline to obtain an intermediate analysis set with business semantic labels and feature vectorization operator parameters; and to obtain the corresponding semantic routing rules based on the business semantic labels of the intermediate analysis set. The decision graph orchestration and verification module is used to configure intent recognition nodes and tool invocation nodes in a preset low-code orchestration canvas according to the semantic routing rules to obtain an agent decision graph; and to perform logical verification on the agent decision graph to obtain an executable agent orchestration instance. The intent routing execution module is used to perform intent parsing on the received natural language query based on the executable intelligent agent orchestration instance, obtain the intent parsing result, activate the corresponding semantic routing rule based on the intent parsing result, so as to trigger the association operator in the data processing pipeline to perform targeted query and model inference on the intermediate analysis set, and obtain the structured analysis result set. The feedback iterative optimization module is used to obtain the target display view based on the structured analysis result set; Collect user interaction feedback data for the target display view, and iteratively update the feature vectorization operator parameters and the preset semantic routing rule matching threshold based on the user interaction feedback data to obtain the optimized agent low-code orchestration model.

Images