Low-code business data non-sensing automatic backflow and knowledge processing method

Abstract

The application discloses a low-code business data non-sensing automatic backflow and knowledge processing method, comprising the following steps: during the operation of a business process of a low-code platform, first business data and first context information generated by the business process are automatically captured based on metadata-driven intelligent point embedding, and the first structured knowledge element is non-invasively transmitted through an asynchronous event bus, intelligently associated to an existing knowledge base through vector similarity matching and knowledge graph query, and the knowledge is supplemented, updated and conflict-resolved based on an incremental learning mechanism to obtain second knowledge after fusion and update; the second knowledge is dynamically fed back to the business system of the low-code platform to form an autonomous closed loop of business operation and knowledge precipitation. The application automatically captures business data and context information based on metadata-driven intelligent point embedding during the operation of the business process of the low-code platform, and finally dynamically feeds back the knowledge after fusion and update to the low-code business system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and enterprise digital transformation, specifically involving a low-code business data seamless automatic backflow and knowledge-based processing method. Background Technology

[0002] In today's context of accelerated digital transformation in enterprises, low-code development platforms have become essential tools for building business applications due to their advantages such as visual modeling, rapid deployment, and lower technical barriers. These platforms enable business users to quickly build applications by dragging and dropping components and configuring rules, greatly improving development efficiency. However, existing low-code platforms generally focus on process automation and have significant shortcomings in data processing depth and intelligent decision support. They struggle to effectively integrate and utilize the massive amounts of heterogeneous data generated during operation and cannot transform them into knowledge assets that drive business optimization.

[0003] With the development of big data and artificial intelligence technologies, knowledge graphs, as structured semantic networks, provide an effective means to extract entities and relationships from multi-source heterogeneous data and construct a global knowledge system. Meanwhile, graph embedding technology can map complex graph structures to low-dimensional vector spaces, preserving their topological structure and semantic relationships, thereby supporting downstream intelligent analysis and decision-making tasks. Nevertheless, these advanced technologies have not yet been deeply integrated with low-code platforms. While current mainstream low-code platforms can handle basic data integration, their capabilities in large-scale data cleaning, complex relationship modeling, dynamic knowledge extraction, and real-time feedback are limited, making it difficult to support data-driven advanced business process automation and optimization.

[0004] Patent CN119338053B proposes a platform management method based on low-code development, aiming to bridge the aforementioned gap. This solution extracts and preprocesses data from multiple heterogeneous data sources to construct a knowledge graph, then uses a graph embedding model to extract global knowledge containing entity relationships and network topology. This global knowledge is then used to evaluate the execution results of business processes and generate guiding decisions. This approach injects knowledge intelligence into low-code platforms, enabling them not only to execute processes but also to autonomously optimize based on a deep understanding of business data. However, this solution focuses on the processing and utilization of static historical data, and its knowledge graph construction relies on prior data extraction and preprocessing. It fails to address the core challenge of how to seamlessly and automatically capture dynamic data during business process execution and instantly transform it into knowledge.

[0005] Therefore, existing technologies, including the solution disclosed in CN119338053B, still face the following key problems: First, there is a lack of a non-intrusive runtime data capture mechanism that seamlessly integrates with low-code platforms, making it difficult to obtain first-hand, high-fidelity business data and its context without interfering with normal business operations; second, there is a lack of intelligent models that can adaptively parse data formats unique to low-code environments (such as forms, rich text, etc.) and accurately extract structured knowledge elements (such as implicit business rules and events); finally, there is a failure to establish a real-time autonomous closed loop from "business operation" to "knowledge accumulation" and then to "dynamic feedback," resulting in lagging knowledge updates and an inability to timely feed back into the immediate optimization of business processes, thus restricting the ability of low-code platforms to evolve towards true intelligence. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a low-code method for seamless automatic data return and knowledge processing of business data, so as to realize automatic data capture, intelligent knowledge extraction and dynamic feedback in business operation.

[0007] The objective of this invention is achieved through the following technical solution: A low-code method for seamless automatic data reflow and knowledge-based processing of business data includes: When the business process runs on the low-code platform, based on metadata-driven intelligent tracking, the first business data and its first context information generated by the business process are automatically captured and transmitted non-intrusively through an asynchronous event bus. The non-intrusive transmission means that the original business code of the low-code platform is not modified, and data packets are captured asynchronously through a listening mechanism. The system receives and parses the first business data, and uses a hybrid intelligent extraction model that integrates semantic and syntactic information to adaptively extract the first structured knowledge element from the first business data. The first structured knowledge element includes entities, relationships, events, and business rules. The first structured knowledge elements are intelligently associated with the existing knowledge base through vector similarity matching and knowledge graph query, and the knowledge is supplemented, updated and conflict resolved based on the incremental learning mechanism to obtain the fused and updated second knowledge. The second knowledge is dynamically fed back to the business system of the low-code platform, forming an autonomous closed loop where business operation is knowledge accumulation.

[0008] As a preferred method, the metadata-driven intelligent data tracking automatically captures the first business data and its first context information, including: When publishing or updating a business process on a low-code platform, the first metadata of the business process is automatically parsed. The first metadata includes node type, form structure, field semantic tags and attachment type. Based on a predefined first knowledge value assessment strategy, high-value knowledge output nodes in the business process are identified and marked. The first knowledge value assessment strategy is used to determine nodes containing preset semantic tags or allowing attachment uploads as high-value knowledge output nodes. The first soft-speech mark is automatically deployed at the high-value knowledge production node. When the execution of the business process triggers the first soft-speech mark, a complete first data packet is captured synchronously. The first data packet contains the first business data, the process execution context in the first context information, and the personnel organization context.

[0009] As a preferred embodiment, the first business data includes structured field values, rich text content, and uploaded files; the process execution context includes a process instance identifier, current node name, handler identifier, timestamp, previous node status, and next node status; the personnel organization context includes the operator's department and operator role.

[0010] As a preferred approach, the method also includes dynamically optimizing the first knowledge value assessment strategy based on the historical knowledge extraction effect; specifically including: For each high-value knowledge-producing node with its first deployed soft-supplement point, calculate the effective knowledge density generated by the corresponding first business data after knowledge processing. The effective knowledge density The calculation formula is: ; in, Effective knowledge density is used to measure the amount of effective knowledge produced by a unit of text information. This represents the total number of first structured knowledge elements that were successfully extracted and validated through fusion from the first business data captured by this node. This indicates the total character length of the first sequence to be processed obtained after multimodal parsing of the first business data captured by the node; This is a smoothing constant, set to 1, with units of characters, used to avoid zero in the denominator; the average effective knowledge density of each high-value knowledge-producing node is calculated at a preset period. If a certain node Continuously below the global average threshold If a node is not marked, it will be removed from the list of high-value knowledge output nodes, and the deployment of new soft-insertions on it will be stopped; conversely, if an unmarked ordinary node is generated in multiple consecutive process instances... Value higher than It will then be automatically upgraded to a high-value knowledge output node and soft-tracked points will be deployed, realizing the adaptive evolution of the tracking strategy.

[0011] As a preferred method, the step of adaptively extracting the first structured knowledge element from the first business data using a hybrid intelligent extraction model that integrates semantic and syntactic information includes: The received first data packet is parsed using multimodal methods. The attachment content in the first data packet is converted into first text data, and the first text data is integrated with the original text data into a unified first sequence to be processed. A joint encoder based on a pre-trained language model and a dependency parsing path is used to encode the first sequence to be processed, so as to deeply integrate semantic features and syntactic structure information to generate the first encoded features; A hybrid expert architecture is adopted in the relation classification layer, in which multiple expert sub-networks learn different types of business relation patterns respectively. Based on the first encoding feature, end-to-end joint extraction of entities and relations is realized to obtain the first structured knowledge elements.

[0012] As a preferred method, it also includes: For scenarios that require summarizing potential business rules from continuous business data streams, a first abductive learning mechanism is introduced; An initial rule pattern hypothesis is set, and the rule pattern hypothesis is iteratively revised and verified by observing continuously flowing business data instances until a first rule set that is logically consistent with all observation results and is concise is derived. The first rule set is then used as the business rules in the first structured knowledge element.

[0013] As a preferred embodiment, the step of intelligently associating the first structured knowledge elements with an existing knowledge base and completing the supplementation, updating, and conflict resolution of knowledge based on an incremental learning mechanism to obtain the fused and updated second knowledge includes: For the first knowledge entity in the first structured knowledge element, a first similarity search is performed in the vectorized semantic index of the existing knowledge base, and the existing relationship network of the potential associated nodes of the first knowledge entity is queried in the knowledge graph of the existing knowledge base. Accurate entity linking is achieved through dual verification of semantic similarity and graph structure, and the first linking result is obtained. If the first link result indicates that the first relation in the first knowledge entity or the first structured knowledge element is new content, then a corresponding first node or first edge is created in the knowledge graph. If the first link result indicates that the first knowledge entity is an existing entity and the first structured knowledge element contains new attribute information of the existing entity, then the historical version of the existing entity is retained to achieve traceability, and the new attribute information is marked as the current valid value to form the first update record; When a first business rule in the first structured knowledge element conflicts with a second business rule in the existing knowledge base, a multi-dimensional automatic adjudication is performed based on the authority score of the rule source, the time freshness value, and the amount of supporting data to obtain a first adjudication result, or conflicts that cannot be automatically adjudicated are marked as first abnormal records pending manual confirmation.

[0014] As a preferred method, the vectorized semantic index of the existing knowledge base is generated by encoding existing knowledge entities in the existing knowledge base through a pre-trained language model. The first similarity search uses cosine similarity calculation, and the formula for calculating cosine similarity is: ; in, The encoding vector representing the first knowledge entity; This represents the encoding vector of any existing knowledge entity in the existing knowledge base; Represents the dot product of two vectors; Representing vectors The Euclidean norm (modulus). Representing vectors The Euclidean norm (modulus) of the cosine similarity, the range of values ​​for which the cosine similarity is given is... The closer the value is The more similar the meanings, the better.

[0015] As a preferred approach, retaining the historical versions of the existing entity specifically includes: maintaining a version list for the existing entity, wherein each node in the version list records a version number, an update timestamp, a source process instance identifier, and a change type, wherein the change type includes addition, modification, or deletion.

[0016] As a preferred approach, after performing the cosine similarity search, graph structure similarity is further introduced for double verification, resulting in a final entity link score. It is obtained by weighted fusion of semantic similarity and structural similarity, and its calculation formula is as follows: ; in, This represents the final entity link score, with a value range of [0, 1]. Cosine similarity; Jaccard The Jaccard similarity coefficient, representing the set of neighbor nodes between the first knowledge entity and existing knowledge entities in the knowledge graph, is defined as: ; in For vectors The set of direct neighbor nodes of the corresponding entity in the knowledge graph. For vectors The set of direct neighbor nodes of the corresponding entity; For semantic weight coefficients, Only when the final entity link score is... Greater than the preset threshold Only when the first knowledge entity and the existing knowledge entity are confirmed to be the same entity is the first knowledge entity confirmed to be the same entity, thus achieving precise linking.

[0017] The present invention has at least the following beneficial effects: The present invention provides a method for seamless automatic feedback and knowledge processing of low-code business data. By using metadata-driven intelligent tracking during the business process of the low-code platform, business data and its contextual information are automatically captured and transmitted in a non-intrusive manner via an asynchronous event bus. Subsequently, a hybrid intelligent extraction model that integrates semantic and syntactic information is used to adaptively extract structured knowledge elements containing entities, relationships, events, and business rules. Then, through vector similarity matching and knowledge graph query, these knowledge elements are intelligently associated with the existing knowledge base, and knowledge is supplemented, updated, and conflict resolved based on an incremental learning mechanism. Finally, the integrated and updated knowledge is dynamically fed back to the low-code business system, forming an autonomous closed loop of "business operation is knowledge accumulation". Attached Figure Description

[0018] To reveal the technical details of the embodiments of the present invention, the accompanying drawings involved in the embodiments will be briefly described below. It should be emphasized that these drawings only present several embodiments of the present invention and should not be considered as defining the scope of the invention. For those skilled in the art, other related drawings can still be derived based on these drawings without inventive effort.

[0019] Figure 1 Flowchart of a low-code, seamless, automatic data reflow and knowledge-based processing method for business data. Figure 2 Flowchart of a hybrid intelligent extraction model that integrates semantic and syntactic information. Detailed Implementation

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0021] It should be noted that while the embodiments described below contain many specific details, intended to facilitate a comprehensive understanding of the exemplary implementations of the present invention, those skilled in the art should understand that the embodiments can still be implemented even without these specific details. For example, the system may be illustrated in the form of a block diagram to avoid affecting the clear expression of the overall solution due to too many implementation details; in other cases, to highlight the core concept, certain well-known common-sense process, structural, or technical details may also be omitted.

[0022] like Figure 1 As shown, a low-code method for seamless automatic data reflow and knowledge-based processing of business data includes: When the business process runs on the low-code platform, based on metadata-driven intelligent tracking, the first business data and its first context information generated by the business process are automatically captured and transmitted non-intrusively through an asynchronous event bus. The non-intrusive transmission means that the original business code of the low-code platform is not modified, and data packets are captured asynchronously through a listening mechanism. The system receives and parses the first business data, and uses a hybrid intelligent extraction model that integrates semantic and syntactic information to adaptively extract the first structured knowledge element from the first business data. The first structured knowledge element includes entities, relationships, events, and business rules. The first structured knowledge elements are intelligently associated with the existing knowledge base through vector similarity matching and knowledge graph query, and the knowledge is supplemented, updated and conflict resolved based on the incremental learning mechanism to obtain the fused and updated second knowledge. The second knowledge is dynamically fed back to the business system of the low-code platform, forming an autonomous closed loop where business operation is knowledge accumulation.

[0023] This solution utilizes metadata-driven intelligent tracking to seamlessly capture and asynchronously transmit data during low-code business process execution. It adaptively extracts structured knowledge such as entities, relationships, events, and rules using a hybrid model that integrates semantics and syntax. Then, it achieves intelligent association with the existing knowledge base through vector similarity and knowledge graph queries, and completes knowledge supplementation, updates, and conflict resolution based on incremental learning. Finally, the updated knowledge is dynamically fed back to the business system, forming an autonomous closed loop where "business operation is knowledge accumulation, and knowledge updates are business optimization." This method requires no modification to the original code, achieving high-precision, low-overhead, and self-evolving automatic knowledge production and application, significantly improving the intelligence level of the low-code platform.

[0024] In a preferred embodiment, if there is a delay in knowledge fusion, newly initiated process instances may still use old rules, leading to loop failure. To address this issue, this invention introduces a process initiation delay control method based on knowledge update timeliness at the task scheduling layer of a low-code platform, implementing a "wait until knowledge is ready before execution" strategy for highly sensitive processes.

[0025] The specific implementation is as follows: Let a certain type of process This falls under the category of highly sensitive information (such as risk control approvals and compliance verifications). The system records the time of the most recent effective knowledge fusion completion. (Unit: seconds). When there is a new When the process request is started, the current time is (Unit: seconds), and check within the time interval Does it exist within the process? When relevant business rules are added or modified, it is recorded as "update with rules" or "update without rules".

[0026] The system presets two parameters: This serves as a knowledge stabilization window (unit: seconds, typical value 5~10 seconds) to ensure that newly integrated knowledge fully takes effect. Sets the maximum waiting time (in seconds, typically 60 seconds) to prevent indefinite blocking.

[0027] The scheduling decision rules are as follows: If in If no rules are updated within the interval, the process will start immediately; otherwise... If a rule is updated within the interval, the process is placed in the waiting queue, and a timer (upper limit) is started. Once the system completes a new round of knowledge fusion (the completion time is recorded as...), ), and at the current moment Then, release all waiting queues in batches. Class process instance; if in If the release conditions are not met within the time limit, the process will be forcibly started, and a "knowledge delayed execution" alarm log will be recorded.

[0028] This example uses "whether there are pending rule updates" as the primary scheduling criterion, rather than solely relying on the time elapsed since the last fusion, thus ensuring that all highly sensitive processes are only executed after the latest knowledge has been actually fused and stabilized. This method uses... Using the completion time of the latest knowledge fusion as the waiting benchmark, this method ensures that highly sensitive processes are only executed after the latest knowledge has been actually fused and passed through a stable window. By embedding knowledge timeliness judgment logic in the task scheduling stage, this method effectively bridges the time gap between knowledge accumulation and business execution, truly achieving an intelligent closed loop of "optimization on run," a dynamic collaborative capability that traditional low-code platforms do not possess.

[0029] In a preferred embodiment, the metadata-driven intelligent tracking automatically captures the first business data and its first context information, including: When publishing or updating a business process on a low-code platform, the first metadata of the business process is automatically parsed. The first metadata includes node type, form structure, field semantic tags and attachment type. Based on a predefined first knowledge value assessment strategy, high-value knowledge output nodes in the business process are identified and marked. The first knowledge value assessment strategy is used to determine nodes containing preset semantic tags or allowing attachment uploads as high-value knowledge output nodes. The first soft-speech mark is automatically deployed at the high-value knowledge production node. When the execution of the business process triggers the first soft-speech mark, a complete first data packet is captured synchronously. The first data packet contains the first business data, the process execution context in the first context information, and the personnel organization context.

[0030] In a preferred embodiment, the presence or absence of semantic tags alone cannot distinguish the type of knowledge a node might generate. For example, a form containing the tags "contract amount," "contracting party," and "effective date" is more likely to generate a "contract signing event," while a node with only the tags "applicant" and "department" is more likely to generate a "personnel entity." Therefore, this invention introduces a knowledge element type prediction mechanism based on field tag combination patterns during the node deployment phase. This mechanism pre-determines the most likely main knowledge category generated by the node and configures the subsequent extraction model's parsing path accordingly.

[0031] The specific implementation is as follows: Suppose a high-value node contains... A field with semantic tags, whose tag set is denoted as . The system pre-configures several knowledge templates (such as "Procurement Events", "Customer Entities", and "Approval Rules"). Each template... Define two parameters: The minimum number of tags required for this template (in units of tags); This represents the ideal total number of tags for this template (unit: tags).

[0032] For each template Calculate its matching score : ; in, For a set of node labels In and template The required number of overlapping labels (unit: labels); Indicates the required label coverage (value range [0, 1]); This indicates the suitability of the number of tags (to prevent too few or too many fields from interfering with the judgment).

[0033] Finally, the template with the highest score is selected as the knowledge type for that node. : ; in, Indicates that The template index that retrieves the maximum value This result will be appended as metadata to the event tracking data package to guide the backend hybrid intelligent model in prioritizing the use of [specific technologies / methods]. Corresponding extraction sub-modules (such as event extractors or entity recognizers) avoid parallel operation of the entire model. This mechanism analyzes the combination features of semantic tags in low-code forms, accurately predicts the potential knowledge output types of nodes without increasing the user's operational burden, significantly improving the targeting and efficiency of knowledge extraction, and is especially suitable for typical scenarios in enterprise-level business processes where field semantics are clear and the structure is standardized.

[0034] In a preferred embodiment, the first business data includes structured field values, rich text content, and uploaded files; the process execution context includes a process instance identifier, current node name, handler identifier, timestamp, previous node status, and next node status; the personnel organization context includes the operator's department and operator role.

[0035] In a preferred embodiment, the method further includes dynamically optimizing the first knowledge value assessment strategy based on the historical knowledge extraction effect; specifically including: For each high-value knowledge-producing node with its first deployed soft-supplement point, calculate the effective knowledge density generated by the corresponding first business data after knowledge processing. The effective knowledge density The calculation formula is: ; in, Effective knowledge density is used to measure the amount of effective knowledge produced by a unit of text information. This represents the total number (in units: elements) of the first structured knowledge elements successfully extracted and validated from the first business data captured by this node. This indicates the total character length (in characters) of the first sequence to be processed obtained after multimodal parsing of the first business data captured by the node. This is a smoothing constant, set to 1, with units of characters, used to avoid zero in the denominator; the average effective knowledge density of each high-value knowledge-producing node is calculated at a preset period. If a certain node Continuously below the global average threshold If a node is not marked, it will be removed from the list of high-value knowledge output nodes, and the deployment of new soft-insertions on it will be stopped; conversely, if an unmarked ordinary node is generated in multiple consecutive process instances... Value higher than It will then be automatically upgraded to a high-value knowledge output node and soft-tracked points will be deployed, realizing the adaptive evolution of the tracking strategy.

[0036] In another preferred embodiment, effective knowledge density relies solely on ex post evaluation. This could still lead to a large number of low-quality attachments being sent to the parsing pipeline, resulting in wasted resources. To address this, the present invention further introduces a point-triggered suppression strategy based on the semantic density of multimodal attachments at the soft-point triggering time. This strategy intercepts attachments that clearly lack information in advance without blocking the main business process, thus avoiding the generation and transmission of invalid data packets.

[0037] Specifically, when the soft-embedded point of a high-value knowledge production node is triggered, the system first performs a lightweight semantic density pre-check on the uploaded attachment. Let the semantic density of the attachment be... The calculation formula is as follows: ; in, This indicates the semantic density of the attachment, expressed in characters per KB. This is the number of text characters that can be quickly extracted directly from the file metadata, in characters (e.g., PDFs are estimated using the ToUnicode table, and DOCs are estimated using the number of paragraphs and the average number of characters, without the need for full-text parsing). λ represents the number of potential text regions detected in the images in the attachment, expressed in units of text, estimated quickly by running a lightweight text detection model (such as EAST) on the thumbnails; λ is the equivalent character coefficient of the image text, expressed in units of characters per character, taken as an empirical value. This indicates that each detected text region is equivalent to approximately 50 recognizable characters; This refers to the size of the attached file, in KB.

[0038] If the calculation yields (e.g., setting a threshold) If the number of characters (in KB) is too low, the attachment is deemed to have too little information (e.g., a blank template, a purely decorative image, or a low-quality scan). In this process, the first data packet will not be generated, even if the node has been marked as a high-value knowledge output node.

[0039] This strategy embeds lightweight attachment content pre-inspection logic into the event listener. By quickly extracting the amount of file metadata text and the number of image text regions, it calculates the semantic density index and directly suppresses the generation of data packets for attachments that obviously lack information (such as blank templates and pure image scans). This reduces invalid transmission and parsing overhead at the source and forms a two-level efficiency governance system that works in synergy with the post-evaluation mechanism based on the D value, which greatly improves the overall throughput efficiency of the system.

[0040] In a preferred embodiment, the asynchronous event bus is implemented using a high-throughput message queue, and the non-intrusive transmission specifically means that when the first soft-slot is triggered, the first data packet is asynchronously sent to the high-throughput message queue without blocking the execution of the main business process of the low-code platform.

[0041] The asynchronous event bus achieves non-intrusive data transmission through a high-throughput message queue, completely decoupling data acquisition from business execution. When the first soft-access event in the low-code platform is triggered, the system immediately generates a first data packet containing key business data and context information. However, this packet is not directly processed or sent in the main business process; instead, it is quickly delivered to the high-throughput message queue by an asynchronous task in the background. This process consumes almost no computing resources of the main process and does not wait for network responses or confirmations, thus having no impact on the execution speed and stability of the original business logic. The message queue, acting as an intermediate buffer layer, can efficiently handle large amounts of concurrent data and ensure its orderly and reliable delivery to subsequent knowledge processing modules. This design not only guarantees the smooth operation of the business system but also achieves complete and timely data feedback, truly achieving seamless acquisition and reliable transmission.

[0042] In a preferred embodiment, such as Figure 2 As shown, the method of using a hybrid intelligent extraction model that integrates semantic and syntactic information to adaptively extract the first structured knowledge element from the first business data includes: The received first data packet is parsed using multimodal methods. The attachment content in the first data packet is converted into first text data, and the first text data is integrated with the original text data into a unified first sequence to be processed. A joint encoder based on a pre-trained language model and a dependency parsing path is used to encode the first sequence to be processed, so as to deeply integrate semantic features and syntactic structure information to generate the first encoded features; A hybrid expert architecture is adopted in the relation classification layer, in which multiple expert sub-networks learn different types of business relation patterns respectively. Based on the first encoding feature, end-to-end joint extraction of entities and relations is realized to obtain the first structured knowledge elements.

[0043] In a preferred embodiment, the method further includes: introducing a first abductive learning mechanism for scenarios that require summarizing potential business rules from continuous business data streams; An initial rule pattern hypothesis is set, and the rule pattern hypothesis is iteratively revised and verified by observing continuously flowing business data instances until a first rule set that is logically consistent with all observation results and is concise is derived. The first rule set is then used as the business rules in the first structured knowledge element.

[0044] In scenarios requiring the automatic discovery of potential business rules from continuous business data streams, the system introduces a first abductive learning mechanism. Starting from observed specific business behaviors, it reverse-engineers the most likely general rules supporting these behaviors. This mechanism first establishes an initial set of rule pattern assumptions based on domain common sense or historical experience, such as "if a customer's level is VIP and the order amount exceeds a threshold, then automatic expedited processing is implemented." Subsequently, as business data instances continuously flow in, the system continuously compares the actual operations with the results predicted by the current rules. If discrepancies arise, the mechanism automatically adjusts the conditions, actions, or logical structure of the rules, continuously revising and verifying these assumptions through multiple iterations, eliminating redundant or contradictory rules, and retaining rules that can explain all observed cases and are concise in form. The final converged first set of rules not only maintains logical consistency with all observed data but also possesses good generalization ability. It can be directly used as business rules in structured knowledge elements for subsequent knowledge fusion and intelligent decision-making, thereby achieving automated extraction from data behavior to interpretable rules.

[0045] In a preferred embodiment, the step of intelligently associating the first structured knowledge elements with an existing knowledge base and supplementing, updating, and resolving conflicts of knowledge based on an incremental learning mechanism to obtain the fused and updated second knowledge includes: For the first knowledge entity in the first structured knowledge element, a first similarity search is performed in the vectorized semantic index of the existing knowledge base, and the existing relationship network of the potential associated nodes of the first knowledge entity is queried in the knowledge graph of the existing knowledge base. Accurate entity linking is achieved through dual verification of semantic similarity and graph structure, and the first linking result is obtained. If the first link result indicates that the first relation in the first knowledge entity or the first structured knowledge element is new content, then a corresponding first node or first edge is created in the knowledge graph. If the first link result indicates that the first knowledge entity is an existing entity and the first structured knowledge element contains new attribute information of the existing entity, then the historical version of the existing entity is retained to achieve traceability, and the new attribute information is marked as the current valid value to form the first update record; When a first business rule in the first structured knowledge element conflicts with a second business rule in the existing knowledge base, a multi-dimensional automatic adjudication is performed based on the authority score of the rule source, the time freshness value, and the amount of supporting data to obtain a first adjudication result, or conflicts that cannot be automatically adjudicated are marked as first abnormal records pending manual confirmation.

[0046] In a preferred embodiment, when performing the above-mentioned multi-dimensional automatic adjudication, if the conflict rules originate from different process nodes, the semantic weight differences in their respective business process topology must also be considered. To this end, this invention introduces a knowledge element time-lapse compensation method based on process topology awareness. This method dynamically adjusts the effective lifespan of the knowledge element according to the topological depth and path criticality of the knowledge-producing node in the directed acyclic graph (DAG) of the process, thereby assigning higher priority to knowledge that is more significant for business decisions during the knowledge fusion and conflict resolution stages.

[0047] Specifically, suppose a certain structured knowledge element consists of process nodes. Output, then its effective life cycle Calculate using the following formula: ; in, This indicates the effective lifespan of the knowledge element, in seconds. This is the basic lifecycle constant, measured in seconds, and is preset by the knowledge type (e.g., business rules). Seconds, event Second); This is the depth enhancement coefficient, with a value range of [0, 1], used to adjust the strength of the positive impact of process depth on knowledge value; For nodes The longest path depth in a process DAG, measured in jumps (i.e., the longest edge from the start of the process to that node); The total longest path length defined for the current process (static attribute), in units of jumps, used to normalize the depth; This is a path criticality factor, and its value is determined based on the path type where the node is located: if the node is located in the "default flow path" (main path) marked in the process metadata, then... If it is located on a normal branch path, then If it is located in an abnormal rollback or rejection path, then .

[0048] This example analyzes the node connection relationships and default path markers in the low-code process definition, calculates the normalized topology depth and path criticality factor of knowledge-producing nodes, dynamically extends the effective lifespan of knowledge produced by the terminal nodes of the main path, and prioritizes the adoption of knowledge with high timeliness weight during conflict resolution. This ensures that the content accumulated in the knowledge base is closer to real business decision-making scenarios, and significantly improves the accuracy and practicality of knowledge feedback.

[0049] In a preferred embodiment, the vectorized semantic index of the existing knowledge base is generated by encoding existing knowledge entities in the existing knowledge base using a pre-trained language model. The first similarity search uses cosine similarity calculation, and the formula for calculating cosine similarity is: ; in, The encoding vector representing the first knowledge entity; This represents the encoding vector of any existing knowledge entity in the existing knowledge base; Represents the dot product of two vectors; Representing vectors The Euclidean norm (modulus). Representing vectors The Euclidean norm (modulus) of the cosine similarity, the range of values ​​for which the cosine similarity is given is... The closer the value is The more similar the meanings, the better.

[0050] In a preferred embodiment, retaining the historical versions of the existing entity specifically includes: maintaining a version list for the existing entity, wherein each node in the version list records a version number, an update timestamp, a source process instance identifier, and a change type, wherein the change type includes addition, modification, or deletion.

[0051] In a preferred embodiment, after performing the cosine similarity search, graph structure similarity is further introduced for double verification, resulting in a final entity link score. It is obtained by weighted fusion of semantic similarity and structural similarity, and its calculation formula is as follows: ; in, This represents the final entity link score, with a value range of [0, 1]. Cosine similarity; Jaccard The Jaccard similarity coefficient, representing the set of neighbor nodes between the first knowledge entity and existing knowledge entities in the knowledge graph, is defined as: ; in For vectors The set of direct neighbor nodes of the corresponding entity in the knowledge graph. For vectors The set of direct neighbor nodes of the corresponding entity; For semantic weight coefficients, Only when the final entity link score is... Greater than the preset threshold Only when the first knowledge entity and the existing knowledge entity are confirmed to be the same entity is the first knowledge entity confirmed to be the same entity, thus achieving precise linking.

[0052] In a preferred embodiment, the step of dynamically feeding back the second knowledge to the business system of the low-code platform includes: The knowledge base content in the second knowledge is pushed to the operation interface associated with the business process in real time, providing business personnel with context-aware intelligent assistance prompts and decision suggestions; The evolved knowledge tags and business rules from the second knowledge are applied in real time to the strategy engine of the downstream intelligent agent, driving the downstream intelligent agent to perform automated decision-making for knowledge optimization and realize business closed-loop optimization.

[0053] The dynamic feedback mechanism of the second knowledge layer aims to reinject the integrated and evolved knowledge into the business system, enabling the immediate transformation of knowledge value. Specifically, the system pushes updated knowledge base content to the operation interface associated with the current business process in real time. This allows frontline business personnel to receive intelligent assistance prompts and decision-making suggestions highly matched to their specific scenarios, such as automatically recommending similar historical cases, warning of potential risks, or prompting compliance requirements, thereby improving decision-making efficiency and accuracy. Simultaneously, the system also injects newly evolved knowledge tags and business rules into the strategy engine of downstream intelligent agents. These agents can be automated approval robots, intelligent scheduling modules, or customer service assistants, etc. They dynamically adjust their behavioral logic based on the latest knowledge rules, executing more accurate and intelligent automated decisions without human intervention. Through this dual-channel feedback approach, the system enhances the contextual awareness of human operations on the one hand, and drives the continuous evolution of machine intelligence on the other, ultimately achieving closed-loop optimization and self-improvement of business processes under human-machine collaboration.

[0054] A low-code, seamless, automatic data feedback and knowledge processing system for business data includes: The first data capture module is deeply integrated with the low-code platform. It is used to automatically capture key business data and complete context information in the business process based on metadata-driven intelligent tracking, and transmit them non-intrusively through an asynchronous event bus. The first knowledge processing module is communicatively connected to the first data capture module and is used to receive and parse the business data, and adaptively extract structured knowledge elements using a hybrid intelligent extraction model that integrates semantic and syntactic information. The first knowledge fusion and feedback module is communicatively connected to the first knowledge processing module. It is used to intelligently associate and fuse the structured knowledge elements into the existing knowledge base, and dynamically feed the updated knowledge back to the business system.

[0055] The low-code business data seamless automatic feedback and knowledge processing system achieves a closed loop from business operations to knowledge accumulation and business empowerment through three closely cooperating modules. First, the system's first data capture module is deeply integrated within the low-code platform. When a business process is released or changed, it automatically parses its metadata, intelligently identifies key nodes with knowledge value, and automatically deploys soft-skill points without modifying the original business logic. This allows for the seamless capture of complete business data and its contextual information during process execution. This data is then transmitted efficiently and reliably in a non-intrusive manner via an asynchronous event bus, avoiding any impact on business system performance. Next, the first knowledge processing module receives this raw data, performs unified parsing, and then uses a hybrid intelligent extraction model that integrates semantic understanding and syntactic structure information to adaptively identify and extract structured knowledge elements such as entities, relationships, and events. The model can dynamically adjust the extraction strategy according to the business scenario, ensuring the accuracy and generalization ability of knowledge extraction. Finally, the first knowledge fusion and feedback module intelligently associates these newly extracted knowledge elements with the existing knowledge base, completes entity alignment through vector similarity matching and graph structure verification, and handles version updates and potential conflicts during incremental fusion to ensure the consistency and timeliness of the knowledge base. After fusion, the updated knowledge will be fed back to the business system in real time, such as being pushed to the operation interface to assist decision-making or injected into the strategy engine of downstream intelligent agents, so that the business system becomes smarter with use, forming a virtuous cycle of data-driven knowledge and knowledge feeding back to business.

[0056] In a preferred embodiment, the first data capture module includes: The metadata analysis unit is used to parse the metadata of the business process when it is published or updated on the low-code platform, and to identify high-value knowledge output nodes based on a predefined knowledge value assessment strategy. The event tracking and monitoring unit is used to automatically deploy soft event tracking points at the high-value knowledge output nodes and to monitor and capture complete data packets generated during runtime. An asynchronous transmission unit is used to transmit the data packets asynchronously and reliably through a high-throughput message queue.

[0057] In a preferred embodiment, the first knowledge processing module includes: The multimodal parsing unit is used to uniformly parse text content and various attachments, converting attachments into processable text data; The hybrid intelligent extraction unit includes a joint encoder that integrates semantic and syntactic information and a relation classifier that adopts a hybrid expert architecture, used to extract entities, relations and events from the business data; The rule discovery unit summarizes and verifies business rules from continuous business data streams based on abductive learning mechanisms.

[0058] When the first data capture module publishes or updates business processes on the low-code platform, the metadata analysis unit first parses the metadata of the process and automatically identifies key nodes with high knowledge output potential based on a predefined knowledge value assessment strategy. Subsequently, the event tracking and monitoring unit automatically deploys soft event tracking on these nodes, and monitors and fully captures the generated data packets in real time during business operation. To ensure efficient and reliable data flow, the asynchronous transmission unit transmits these data packets asynchronously through a high-throughput message queue, using message confirmation and retry mechanisms to prevent data loss. Building upon this foundation, the first knowledge processing module performs structured transformation on the raw data: the multimodal parsing unit uniformly processes text and various attachments, converting non-text files such as PDFs and images into analyzable text formats; the hybrid intelligent extraction unit combines semantic understanding and syntactic structure information, accurately identifies entities through a joint encoder, and models different business relationship patterns separately using a hybrid expert architecture relationship classifier, thereby efficiently extracting entities, relationships, and events; finally, the rule discovery unit, based on abductive learning mechanisms, automatically summarizes, verifies, and optimizes potential business rules from the continuously flowing business data, enabling the system to continuously extract reusable logical knowledge from actual operations, forming a complete closed loop from data capture to knowledge generation.

[0059] In a preferred embodiment, the first knowledge fusion and feedback module includes: The intelligent association unit is used to link and align entities of new knowledge through vector similarity calculation and knowledge graph structure query; The incremental fusion and conflict resolution unit uses an incremental learning mechanism to handle knowledge supplementation, version updates, and conflict coordination and resolution. The knowledge feedback unit is used to push updated knowledge to the business operation interface or the strategy engine of downstream intelligent agents in real time.

[0060] The first knowledge fusion and feedback module effectively integrates newly extracted knowledge into the existing knowledge system and promptly feeds its value back to the business system. This module first connects new knowledge with the existing knowledge base through an intelligent association unit. Specifically, it combines vector similarity calculation and structured querying of the knowledge graph: on the one hand, semantic vectors are used to measure the semantic similarity between new entities and existing entities; on the other hand, topological relationships in the knowledge graph (such as parent-child, membership, and dependency) are used to verify logical consistency, thus achieving accurate entity linking and alignment. Building on this, the incremental fusion and conflict resolution unit handles the dynamic fusion of new and old knowledge. It employs an incremental learning mechanism to efficiently absorb new facts and update outdated information without retraining the entire system. It also automatically identifies and coordinates contradictions (such as two conflicting attribute values ​​for the same entity), ensuring the knowledge base remains consistent and timely. Finally, the knowledge feedback unit pushes the latest fused and verified knowledge to the business operation interface or the strategy engine of downstream intelligent agents in real time, enabling frontline personnel or automated systems to immediately utilize the updated knowledge to optimize decisions, adjust processes, or improve the user experience, forming a positive cycle from knowledge generation to application.

[0061] In a preferred embodiment, the asynchronous transmission unit uses a high-throughput message queue that supports at least one message delivery semantics and is configured with a message acknowledgment mechanism to ensure that data is not lost; the joint encoder in the hybrid intelligent extraction unit is built based on a pre-trained language model and a dependency parsing path, and the hybrid expert architecture in the relation classifier contains multiple expert sub-networks, each of which learns a preset business relationship pattern.

[0062] In this embodiment, the asynchronous transmission unit employs a high-throughput message queue to carry the data stream captured from the business process. This message queue supports at least-once message delivery semantics and is coupled with a message acknowledgment mechanism. That is, a message is only marked as complete and removed from the queue after the receiving end successfully processes it and returns an acknowledgment signal. If processing fails or no acknowledgment is received, the system automatically retryes delivery, effectively preventing data loss during transmission and ensuring that all business events are fully recorded and subsequently processed. Based on this, the hybrid intelligent extraction unit is responsible for transforming the raw business data into structured knowledge. Its core is a joint encoder, which integrates the deep semantic understanding capabilities of a pre-trained language model with the internal structural relationships revealed by dependency parsing, enabling more accurate identification of key components and their interactions in the text. Furthermore, the relation classifier adopts a hybrid expert architecture, which includes multiple specialized sub-networks. Each expert performs optimization learning for a specific type of business relationship pattern (such as approval, applicant, contract, signatory, etc.). When extracted information is input, the system dynamically activates the most relevant expert combination based on the context, thereby improving the accuracy and generalization ability for identifying complex and variable business relationships. The entire process, from reliable transmission to accurate parsing, progresses step by step, laying a high-quality data foundation for subsequent knowledge integration and closed-loop feedback.

[0063] Although preferred embodiments of the present invention have been described above, those skilled in the art, based on their understanding of the core concept of the invention, can make various changes, substitutions, or improvements. Therefore, the appended claims should be interpreted as covering the preferred embodiments, as well as all equivalent substitutions, modifications, and improvements falling within the spirit and scope of the invention. It should be emphasized that the above description is merely a preferred embodiment of the invention and is not a limitation thereof; any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be considered as included within the scope of protection of the invention.

Claims

1. A low-code, seamless automatic data reflow and knowledge-based processing method for business data, characterized in that, include: When the business process runs on the low-code platform, based on metadata-driven intelligent tracking, the first business data and its first context information generated by the business process are automatically captured and transmitted non-intrusively through an asynchronous event bus. The non-intrusive transmission means that the original business code of the low-code platform is not modified, and data packets are captured asynchronously through a listening mechanism. The system receives and parses the first business data, and uses a hybrid intelligent extraction model that integrates semantic and syntactic information to adaptively extract the first structured knowledge element from the first business data. The first structured knowledge element includes entities, relationships, events, and business rules. The first structured knowledge elements are intelligently associated with the existing knowledge base through vector similarity matching and knowledge graph query, and the knowledge is supplemented, updated and conflict resolved based on the incremental learning mechanism to obtain the fused and updated second knowledge. The second knowledge is dynamically fed back to the business system of the low-code platform, forming an autonomous closed loop where business operation is knowledge accumulation.

2. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 1, characterized in that, Based on metadata-driven intelligent event tracking, the system automatically captures the first business data and its first context information, including: When publishing or updating a business process on a low-code platform, the first metadata of the business process is automatically parsed. The first metadata includes node type, form structure, field semantic tags and attachment type. Based on a predefined first knowledge value assessment strategy, high-value knowledge output nodes in the business process are identified and marked. The first knowledge value assessment strategy is used to determine nodes containing preset semantic tags or allowing attachment uploads as high-value knowledge output nodes. The first soft-speech mark is automatically deployed at the high-value knowledge production node. When the execution of the business process triggers the first soft-speech mark, a complete first data packet is captured synchronously. The first data packet contains the first business data, the process execution context in the first context information, and the personnel organization context.

3. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 2, characterized in that, The first business data includes structured field values, rich text content, and uploaded files; the process execution context includes process instance identifier, current node name, handler identifier, timestamp, previous node status, and next node status; the personnel organization context includes the operator's department and operator role.

4. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 2, characterized in that, It also includes dynamically optimizing the first knowledge value assessment strategy based on the effectiveness of historical knowledge extraction; specifically including: For each high-value knowledge-producing node with its first deployed soft-supplement point, calculate the effective knowledge density generated by the corresponding first business data after knowledge processing. The effective knowledge density The calculation formula is: ； in, Effective knowledge density is used to measure the amount of effective knowledge produced by a unit of text information. This represents the total number of first structured knowledge elements that were successfully extracted and validated through fusion from the first business data captured by this node. This indicates the total character length of the first sequence to be processed obtained after multimodal parsing of the first business data captured by the node; This is a smoothing constant, set to 1, with units of characters, used to avoid zero in the denominator; the average effective knowledge density of each high-value knowledge-producing node is calculated at a preset period. If a certain node Continuously below the global average threshold If a node is not marked, it will be removed from the list of high-value knowledge output nodes, and the deployment of new soft-insertions on it will be stopped; conversely, if an unmarked ordinary node is generated in multiple consecutive process instances... Value higher than It will then be automatically upgraded to a high-value knowledge output node and soft-tracked points will be deployed, realizing the adaptive evolution of the tracking strategy.

5. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 1, characterized in that, Using a hybrid intelligent extraction model that integrates semantic and syntactic information, the first structured knowledge element is adaptively extracted from the first business data, including: The received first data packet is parsed using multimodal methods. The attachment content in the first data packet is converted into first text data, and the first text data is integrated with the original text data into a unified first sequence to be processed. A joint encoder based on a pre-trained language model and a dependency parsing path is used to encode the first sequence to be processed, so as to deeply integrate semantic features and syntactic structure information to generate the first encoded features; A hybrid expert architecture is adopted in the relation classification layer, in which multiple expert sub-networks learn different types of business relation patterns respectively. Based on the first encoding feature, end-to-end joint extraction of entities and relations is realized to obtain the first structured knowledge elements.

6. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 5, characterized in that, Also includes: For scenarios that require summarizing potential business rules from continuous business data streams, a first abductive learning mechanism is introduced; An initial rule pattern hypothesis is set, and the rule pattern hypothesis is iteratively revised and verified by observing continuously flowing business data instances until a first rule set that is logically consistent with all observation results and is concise is derived. The first rule set is then used as the business rules in the first structured knowledge element.

7. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 1, characterized in that, The first structured knowledge elements are intelligently linked to the existing knowledge base, and knowledge is supplemented, updated, and conflict resolved based on an incremental learning mechanism to obtain the fused and updated second knowledge, including: For the first knowledge entity in the first structured knowledge element, a first similarity search is performed in the vectorized semantic index of the existing knowledge base, and the existing relationship network of the potential associated nodes of the first knowledge entity is queried in the knowledge graph of the existing knowledge base. Accurate entity linking is achieved through dual verification of semantic similarity and graph structure, and the first linking result is obtained. If the first link result indicates that the first relation in the first knowledge entity or the first structured knowledge element is new content, then a corresponding first node or first edge is created in the knowledge graph. If the first link result indicates that the first knowledge entity is an existing entity and the first structured knowledge element contains new attribute information of the existing entity, then the historical version of the existing entity is retained to achieve traceability, and the new attribute information is marked as the current valid value to form the first update record; When a first business rule in the first structured knowledge element conflicts with a second business rule in the existing knowledge base, a multi-dimensional automatic adjudication is performed based on the authority score of the rule source, the time freshness value, and the amount of supporting data to obtain a first adjudication result, or conflicts that cannot be automatically adjudicated are marked as first abnormal records pending manual confirmation.

8. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 7, characterized in that, The vectorized semantic index of the existing knowledge base is generated by encoding existing knowledge entities in the existing knowledge base using a pre-trained language model. The first similarity search uses cosine similarity calculation, and the formula for calculating cosine similarity is: ； in, The encoding vector representing the first knowledge entity; This represents the encoding vector of any existing knowledge entity in the existing knowledge base; Represents the dot product of two vectors; Representing vectors The Euclidean norm, Representing vectors The Euclidean norm, the range of values ​​for the cosine similarity is... The closer the value is The more similar the meanings, the better.

9. The method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 7, characterized in that, Retaining the historical versions of the existing entity specifically includes: maintaining a version list for the existing entity, wherein each node in the version list records a version number, an update timestamp, a source process instance identifier, and a change type, wherein the change type includes addition, modification, or deletion.

10. A method for seamless automatic backflow and knowledge-based processing of low-code business data according to claim 8, characterized in that, After performing the cosine similarity search, graph structure similarity is further introduced for double verification, resulting in a final entity link score. It is obtained by weighted fusion of semantic similarity and structural similarity, and its calculation formula is as follows: ； in, This represents the final entity link score, with a value range of [0, 1]. The cosine similarity as defined in claim 8; Jaccard The Jaccard similarity coefficient, representing the set of neighbor nodes between the first knowledge entity and existing knowledge entities in the knowledge graph, is defined as: ； in For vectors The set of direct neighbor nodes of the corresponding entity in the knowledge graph. For vectors The set of direct neighbor nodes of the corresponding entity; For semantic weight coefficients, Only when the final entity link score is... Greater than the preset threshold Only then is it confirmed that the first knowledge entity and the existing knowledge entity are the same entity, thus achieving precise linking.

Images