A Multi-Agent Collaborative System and Method Based on Knowledge Specification Definition Schema

Abstract

This invention discloses a multi-agent collaboration method and system based on knowledge specification definition, belonging to the field of knowledge-driven multi-agent collaboration technology. This invention defines reusable enterprise standards, business rules, etc., as knowledge specification definition entities carrying complete schemas. The schema defines the knowledge item structure, quality assessment thresholds, and risk triggering conditions. External task execution units complete initialization and execution according to schema constraints, with process data carrying schema version identifiers and flowing back. Platform agents verify and bin the data according to schema constraint rules. The optimal data confirmed by binning is distributed through multiple controlled paths, archived to the knowledge hub, triggering standard upgrade approval, and updating the status of related systems. When a quality risk is identified, a structured risk instruction is broadcast to the agent cluster, and the agents dynamically match and trigger preset risk containment strategies through the capability schema.

Description

Technical Field

[0001] This invention relates to AI-driven enterprise knowledge management and multi-agent collaboration technology, and in particular to a closed-loop system and method that uses a knowledge specification definition schema as the end-to-end driving engine and runs through knowledge encapsulation, consumption, recirculation, cleaning, binning, multi-path distribution and risk broadcasting. It is applicable to enterprise-level application scenarios with high requirements for structured consumption, automated quality assessment, cross-system linkage and continuous self-optimization of knowledge assets. Background Technology

[0002] In traditional enterprise applications, knowledge assets are stored in unstructured or semi-structured forms across various business systems, facing the following problems: First, knowledge consumption lacks structured constraints, making it impossible for external systems or intelligent agents to verify whether the consumption results conform to business specifications when acquiring knowledge; second, process data after task execution lacks structured association with knowledge versions, making accurate quality traceability impossible; third, quality assessment relies on manual sampling, making it impossible to identify execution quality patterns in a full and real-time manner; and fourth, quality assessment results are disconnected from downstream processes such as knowledge updates, inventory management, and parameter optimization, failing to form a closed-loop linkage. Existing standard protocols are mainly used for the interaction between AI models and external tools, but have not yet solved the technical problem of "how to pass the structured constraints of knowledge with the call and run through the entire chain of execution, feedback, evaluation and distribution"; In existing technologies, CN121659985A discloses a multi-agent collaborative decision-making method based on knowledge tokens, which achieves trusted encapsulation of knowledge and multi-agent collaborative decision-making through an original document fidelity layer, an activated knowledge layer, and a unified index layer. However, the knowledge token in this scheme completes its mission after consumption and does not involve post-consumption quality assessment, parameter self-optimization, or cross-system linkage. Its calling contract only constrains the usage conditions of the token and lacks the driving capability to run through the entire chain of consumption, reflow, bucketing, and distribution. Furthermore, the data format definition schema commonly used in the computer field only defines the field names and field types of data, solving the problem of data format consistency, but it does not have the ability to drive quality assessment and cross-system linkage.

[0003] The technical problem to be solved by this invention is to provide a closed-loop system and method that uses the knowledge specification definition body Schema as the end-to-end driving engine and runs through knowledge encapsulation, consumption, reflow, cleaning, bucketing, multi-path distribution and risk broadcasting. It should be noted that the knowledge specification definition schema described in this invention differs from the data format definition schema commonly used in the computer field. The general schema only defines the field names and field types of data, solving the problem of data format consistency; the schema of this invention not only defines the structure of knowledge items, but also defines bucketing thresholds, risk levels, and approval triggering conditions, and possesses self-description and lifecycle management capabilities, serving as a full-link driving engine throughout knowledge consumption, quality assessment, risk broadcasting, and parameter updates. This difference upgrades the schema of this invention from passive data format constraints to proactive quality decision-driven processes, which is the key feature distinguishing this invention from existing technologies. It should be noted that the knowledge specification definition body described in this invention is a structured carrier of all reusable standard document specifications, business rules, and technical data standards for an enterprise. Unlike document templates or data templates commonly used in the computer field, the core feature of the knowledge specification definition body is that its complete schema not only defines the name and type of knowledge items, but also defines constraint rules, bucketing thresholds, and risk levels, thus upgrading the knowledge specification definition body from a passive format definition to an active, end-to-end driving engine. Unlike the knowledge token mechanism in CN121659985A, the Schema of the knowledge specification definition body in this invention is not only a definition format for knowledge, but also a driving engine that runs through the entire chain. The bucketing thresholds and risk levels defined in the Schema directly drive quality assessment, the Schema version identifier associates the version ownership of the returned data, the Schema constraint rules drive data cleaning and verification, and the bucketing results trigger cross-system linkage for PLM upgrade approval and ERP inventory updates. This end-to-end Schema-driven mechanism is the core feature that distinguishes this invention from existing technologies. To achieve the above objectives, this invention proposes the following technical solution: All reusable knowledge assets of an enterprise are defined as knowledge specification definitions carrying a complete schema, and encapsulated as knowledge consumption services through a standard protocol; external task execution units complete knowledge consumption and task execution according to schema constraints; multi-dimensional process data after execution is fed back with a schema version identifier; platform agents perform legality verification and bucketing analysis on the data according to schema constraint rules; the optimal data confirmed by bucketing is distributed through multiple controlled channels and archived to the knowledge hub for subsequent consumption, comparison with PLM atomic standards to trigger upgrade approval, and updating ERP inventory status; when serious quality risks or fatal errors are identified, a risk notification carrying structured instruction parameters is directly broadcast to the agent cluster, and the receiving agent can directly parse and trigger a preset risk containment strategy. It should be noted that in the architecture of this invention, communication between intelligent agents and collaboration between intelligent agents and users are achieved through different channels. Intelligent agents directly transmit structured parameters via standard protocols (such as MCP, WebSocket, HTTP, etc.), including verified parameter ranges, unverified parameter ranges, risk level parameters, and instruction parameters. The receiving intelligent agent can directly parse the structured parameters and trigger a preset automatic response strategy. Collaboration between intelligent agents and users is achieved through an approval engine middleware, including triggering user confirmation processes, change approval processes, and feedback collection processes. These two types of communication channels operate independently and each performs its own function. It should be noted that the multi-path controlled distribution described in this invention is not limited to product lifecycle management platforms and enterprise resource planning systems. In specific implementations, the optimal data confirmed by bucketing can be distributed to any enterprise-level application platform that needs the data, including but not limited to manufacturing execution systems, quality management systems, and warehouse management systems. The distribution path is determined by the linkage rules defined in the knowledge specification definition schema. New application platforms only need to access the knowledge hub through a standard protocol to join the linkage system. It should be noted that the dynamic technical knowledge nodes of the product lifecycle management platform in collaboration with the knowledge hub in this invention are special structured storage carriers under the architecture of this invention. Unlike ordinary database storage, the technical standard data stored in the dynamic technical knowledge nodes all come from approved and archived tasks. Each data record is associated with the source task identifier, approval number, archive time, and version number, ensuring full-link traceability and possessing strong process control and version traceability capabilities not available in ordinary databases. This dynamic technical knowledge node and the structured knowledge base belong to the same structured storage system under the architecture of this invention, sharing the same schema definition specifications and version management mechanism. It should be noted that in the architecture of this invention, the platform intelligent agent plays three roles: firstly, a knowledge consumption scheduler—retrieves the knowledge specification definition from the knowledge hub through standard protocols and aggregates multi-source heterogeneous knowledge to complete task initialization; secondly, a quality assessment executor—performs bucket analysis on the returned data according to the bucketing thresholds and risk levels defined in the Schema, and identifies execution quality patterns; and thirdly, a risk monitoring manager—monitors risk indicators during the execution process, and when a risk indicator is detected to trigger a preset threshold, automatically generates an early warning notification with structured instruction parameters, broadcasts it to relevant terminals, and triggers the approval process. It should be noted that in the risk handling architecture of this invention, there are two independent risk types and processing channels: the first is operational risk—triggered by operational indicators such as changes in inventory status, where the platform agent generates an early warning notification and broadcasts it to relevant terminals; the second is quality risk—triggered by severe quality patterns identified by bucket analysis, where the platform agent directly broadcasts a risk notification carrying structured instruction parameters to the agent cluster and triggers the approval engine middleware to create an approval process. The processing channels, receiving objects, and response mechanisms for the two types of risks are independent of each other and each performs its own function. It should be noted that in the bucket architecture of this invention, there are two independently operating layers: the quality bucket boundaries of the quality assessment layer and the operation window boundaries of the operation parameter layer. The quality bucket boundaries are based on objective quality result indicators, remain fixed in the manufacturing scenario, and continuously narrow as data accumulates, with statistical confidence continuously improving; the operation window boundaries are the distribution range of operation parameters within each quality bucket, automatically shifting with changes in environmental conditions. The two layers operate independently and do not interfere with each other—the quality assessment standards remain unchanged, while the optimal operation range automatically adapts to environmental changes; It should be noted that the bucketing threshold can be defined directly by the schema, or it can be referenced to an external configuration source through threshold referencing rules defined in the schema. Regardless of the method used, the bucketing threshold is associated with the schema version identifier. When the schema version changes, the associated bucketing threshold is updated synchronously to ensure version consistency and traceability of bucketing analysis. It should be noted that the bucketing analysis described in this invention refers to discretizing and grouping execution records according to result quality index values ​​to form statistically significant discrete quality levels. The discretization essence of bucketing lies in mapping continuous quality index values ​​to a finite number of quality levels, with each quality level corresponding to a range of operational parameter distributions. Even if execution records are scored using a continuous scoring function and then grouped according to a scoring threshold, as long as discrete quality levels are formed after grouping and the operational parameter distributions of each level are extracted retrospectively, it still falls within the scope of bucketing analysis. The core technical characteristics of bucketing analysis are that it uses result quality indices as the grouping basis, discretization as the analysis method, and retrospective extraction of operational parameter distributions as the optimized output—rather than the specific implementation form of bucketing. It should be noted that the platform's intelligent agent automatically triggers the parameter upgrade approval process and automatically updates the inventory status of the enterprise resource planning system. The technical trigger condition is that the difference between the optimal parameters confirmed by the bucketing and the atomic technical standard parameters exceeds a preset threshold—this is automated cross-system linkage driven by technical parameters, distinct from manual operation based on human judgment. The core technical effect of automated linkage lies in eliminating the response delay and subjective bias of human decision-making, ensuring the timeliness and consistency of parameter optimization. When the difference exceeds the threshold, the platform's intelligent agent automatically creates a structured upgrade change request carrying a bucketing analysis report and difference comparison data. After approval, it automatically triggers parameter updates—the entire process is driven by technical events, requiring no manual judgment on whether a change request needs to be created. It should be noted that the schema also includes its own meta-description information, including schema version identifier, effective date, expiration conditions, and applicable boundaries. When a schema reaches its expiration conditions, the system automatically triggers a schema migration or deactivation process. It should be noted that the "layered fixed architecture" described in this invention refers to a top-down hierarchical constraint system: the knowledge item structure of the upper-level classification hierarchy is preset by the system and cannot be changed, while the knowledge item names and types of the lower-level classification hierarchy are filled in by the system under the constraints of the upper-level classification hierarchy. The core design concept of this architecture is "the upper layer ensures structural consistency and a safety baseline, while the lower layer flexibly adapts within the constraints." The hierarchical inheritance mechanism of the schema and the automatic generation mechanism of the standard template self-learning system in this invention both follow this design concept. It should be noted that the standard template self-learning system referenced in this invention refers to a system that continuously expands the knowledge extraction template library through an automatic generation and registration mechanism under a hierarchical fixed architecture. The core mechanism of this system includes: when no extraction template corresponding to the current category identifier exists in the mapping table, a new extraction template is automatically generated according to the preset hierarchical fixed template architecture; the newly generated template is added to the template library after registration and review for use by subsequent similar documents. The schema lifecycle management in this invention adopts the automatic registration, version tracking, and failure detection mechanisms of this system, the difference being that the managed object is expanded from knowledge extraction templates to knowledge specification definition bodies (Schemas). It should be noted that the end-to-end traceability is recorded in the form of a structured traceability tree. Each node in the traceability tree includes an operation step identifier, an executing entity identifier, an application schema version identifier, an operation timestamp, and a hash signature of the operation result. The nodes are linked through a hash chain to form an immutable audit chain. The hash signature and the vector storage encryption scheme share the same key management system—this same key system protects both the storage security of the vector data and the tamper-proof traceability chain, thus creating a complete security and compliance closed loop between secure knowledge storage and reliable knowledge traceability. It should be noted that each agent in the agent cluster carries an agent capability schema, defining the agent's capability type, permission level, and responsive instruction schema. When a platform agent broadcasts a risk notification, it includes the target capability requirement in the notification. The receiving agent matches the target capability requirement with its own capability schema; only agents with matching capabilities parse and execute the corresponding risk mitigation strategy. This mechanism ensures that the broadcast recipients are dynamically determined—through structured matching rules of capabilities and permissions, rather than a hard-coded list of recipients. It should be noted that the schema adopts a hierarchical inheritance structure, including at least three levels: global schema, domain schema, and project schema. Lower-level schemas automatically inherit all constraint rules, bucketing thresholds, and risk levels from upper-level schemas; lower-level schemas can tighten constraints within the limits allowed by the upper-level schema, but cannot loosen the constraints defined by the upper-level schema. This mechanism is similar to a layered fixed architecture—the upper layer ensures structural consistency and a safety baseline, while the lower layer flexibly adapts within constraints. It's important to note that the Agent performs self-reasoning and summarizing on user feedback—regardless of whether the feedback content is contradictory, the Agent analyzes the feedback perspectives, supporting evidence, and conclusions from each user to identify consensus and disagreement, and extract factual conclusions. When feedback content is contradictory, the Agent performs a weighted analysis of the points of contention based on the feedback evidence—determining the source of the contradiction and the truth based on the historical reliability of each user's feedback, the objectivity of the feedback evidence, and the consistency of the feedback conclusions. This continuous self-reasoning ability is the core characteristic that distinguishes the Agent from traditional systems that passively receive feedback signals—the Agent does not wait for contradictions to appear before initiating reasoning, but is always in an active and continuous process of state assessment and knowledge extraction. It should be noted that in the ecological architecture of this invention, the self-learning capability is reflected in three levels: firstly, the self-learning of standard templates—through automatic generation and registration mechanisms under a layered fixed architecture, the knowledge extraction template library is continuously expanded; secondly, the self-learning of agent execution quality—through broadcast feedback, continuous reasoning, and bucketing / blacklisting mechanisms, the execution strategy is continuously evolved; and thirdly, the self-learning of the schema—through self-description, self-verification, and layered inheritance mechanisms, the knowledge specification definition body is self-evolved. These three levels of self-learning loops are unified under the design concept of the layered fixed architecture, forming a complete self-learning ecosystem from knowledge definition to knowledge consumption to quality assessment. It should be noted that the "negative entropy increase in data quality" mentioned in this invention refers to the fact that, under the natural trend of data scale increasing with usage time, general quality assessment methods usually face the problem of assessment accuracy decay. This invention, through schema version grouping, time decay weight calculation, and version change recalibration mechanism, enables the bucket analysis results to continuously converge rather than diverge with the accumulation of data volume, and the assessment accuracy to continuously improve rather than degrade, thus achieving negative entropy increase in data quality management.

[0004] Compared with existing solutions, the present invention has the following advantages: 1. Structured constraints on knowledge consumption: External task execution units must populate knowledge items according to the schema definition, and the consumption results can be automatically verified; 2.* Fully Automated Quality Assessment: The bucketing engine automatically assesses the quality of each task execution based on the bucketing thresholds and risk levels defined in the schema, eliminating the need for manual sampling. The core technical feature of this bucketing analysis lies in using result quality indicators as the grouping basis, discretization as the analysis method, and backtracking to extract the distribution of operation parameters as the optimized output. The discretization nature of bucketing ensures that each quality level has statistical significance, avoiding quality assessment distortion caused by design biases in the scoring function of continuous scoring models. 3. Adaptive Drift of Operation Window and Continuous Convergence of Quality Buckets: The operation window boundary automatically follows the drift—the distribution of operation parameters within the optimal quality bucket naturally reflects the best operation range under the current environment; the quality bucket boundary is based on objective quality result indicators, continuously narrowing with data accumulation, and the statistical confidence continuously improves without degrading over time. When the objective quality result indicators correspond to a fixed physical standard, the quality bucket boundary continuously narrows with data accumulation, the statistical confidence continuously improves, and the operation window continuously converges around the fixed quality standard. Bucket analysis is grouped by schema version, and data within the same version is statistically analyzed with time decay weights; when the schema version changes, the data of the old and new versions are recalibrated based on the change records. This mechanism ensures that the bucket analysis results continuously converge with the increase in usage time and data volume, achieving negative entropy increase in data quality—the more data, the more accurate the evaluation, and there is no decay; 4. Self-optimization closed loop for technical parameters: The optimal parameters confirmed by each bucket automatically trigger parameter upgrade approval. After approval and archiving, the optimized version is automatically retrieved for the next consumption. The automated triggering of the parameter upgrade approval process is driven by technical parameters—when the difference between the optimal parameters confirmed by each bucket and the atomic standard exceeds a preset threshold, the platform's intelligent agent automatically creates a structured change request without requiring manual judgment on whether creation is necessary. This technical parameter-driven automation mechanism eliminates the response delay and subjective bias of manual decision-making, ensuring the timeliness and consistency of parameter optimization; 5. Risk Identification and Real-Time Containment: When a bucket identifies a serious quality risk or fatal error, the platform agent directly broadcasts a risk notification carrying structured instruction parameters to the agent cluster. The receiving agent can directly parse the structured parameters and trigger the preset risk containment strategy. At the same time, it triggers the approval engine middleware to create a risk approval process, realizing real-time risk perception, automatic response, and process traceability. 6. Compound accumulation of knowledge assets: Each task execution contributes structured experience data that has been validated for quality to the knowledge center, and the quality of knowledge continues to improve with the number of executions; 7. Data cleaning and binning analysis together constitute a dual quality assurance system: Data cleaning, as a preprocessing step before binning analysis, filters out obviously abnormal data through schema constraint rules, providing clean data input for binning analysis; binning analysis, as the core step of quality assessment, identifies execution quality patterns based on result quality indicators. The two work in sequence and collaboratively to form a dual assurance system from data validity to execution quality. 8. Schema Self-Description and Lifecycle Management: The schema possesses self-description capabilities, carrying its own version identifier, effective date, expiration conditions, and applicable boundaries. When a schema reaches its expiration conditions, the system automatically triggers a migration or deactivation process, eliminating the need for manual management of schema version evolution. This capability makes the schema a "living document" with self-lifecycle management capabilities. 9. Structured and in-depth end-to-end traceability: Traceability information is recorded in the form of a structured traceability tree. The operator, rules, inputs, and outputs of each node are protected by hash signatures. Each node is linked through a hash chain to form an immutable audit chain. The hash signature and vector storage encryption scheme share a key management system, forming a complete closed loop from secure knowledge storage to trusted knowledge traceability. 10. Schema-based matching of agent capabilities: Each agent in the agent cluster carries a capability schema. The receivers of broadcasts are dynamically determined through structured matching rules between capabilities and permissions, rather than a hard-coded list of recipients. This mechanism allows the system to add or remove agents without modifying the broadcast configuration, improving the system's robustness and scalability. 11. Hierarchical Inheritance and Controlled Coverage of Schema: The schema adopts a three-tier inheritance structure: global, domain, and project. Lower layers automatically inherit constraints from upper layers, which can only be tightened, not loosened. This mechanism enables enterprises to establish a knowledge governance system with "one global standard, multiple domain adaptations, and countless project instances"—standards are uniformly set, adaptations are flexibly implemented, and security is never compromised. 12. Agent Continuous Autonomous Reasoning: The agent's self-reasoning is a continuous and proactive state assessment process—regardless of whether user feedback is contradictory, the agent performs a weighted analysis based on the feedback's supporting evidence, comprehensively analyzing and extracting facts based on the historical reliability of each user's feedback, the objectivity of the feedback's supporting evidence, and the consistency of the feedback conclusions. This continuous reasoning mechanism ensures the agent's performance quality is continuously improved, rather than only triggering optimization when contradictions occur. Attached Figure Description

[0005] 【 Figure 1 [This is a diagram showing the overall architecture of the system of the present invention;] 【 Figure 2 This is a flowchart illustrating the schema-driven process of knowledge encapsulation, consumption, and reflow. 【 Figure 3 This is a schematic diagram illustrating the multi-path controlled distribution and risk broadcasting of a schema-driven bucketing engine. Detailed Implementation

[0006] To make the objectives, technical solutions, and advantages of the present invention clearer, further detailed descriptions are provided below in conjunction with the accompanying drawings and embodiments; Example 1: Schema-Driven Knowledge Encapsulation and Consumption A manufacturing company defines project review reports as knowledge specification definitions, whose schema includes project name (string, required), project period (string, required), target achievement rate (numerical value, range 0-100), key issues (list), lessons learned (text), and improvement measures (text). This knowledge specification definition is encapsulated as a knowledge consumption service using the MCP protocol. After receiving the task, the platform agent retrieves the knowledge specification definition body through the MCP protocol, populates the knowledge items according to the schema definition, and completes the structured initialization of the report. After the populated result passes the schema validation, it is submitted to the approval engine middleware for approval. Example 2: Multidimensional process data backflow and schema-driven bucketing The embodied intelligent robot obtains assembly torque parameters from dynamic technical knowledge nodes (such as the technical database in a PLM platform) via the MCP protocol. Its schema definition includes: parameter name (assembly torque), data type (float), value range (0-20 N·m), bucket threshold (each 0.5 N·m represents a bucket interval), and risk level (the acceptable range of 5.5-6.5 N·m is considered a high-quality bucket, and 3-4 N·m is considered a risky bucket). The robot executes the assembly task according to the schema constraints. After assembly, multi-dimensional process data—process data (actual torque value), inspection data (torque test results), and material consumption data (screw usage)—are sent back to the platform Agent along with the Schema version identifier v2.1. The platform agent first verifies the validity of the returned data based on the value range constraints and data type constraints defined in the schema, removing abnormal data that exceeds the value range or has a mismatched data type. The removed abnormal data is stored in association with the cleaned data and carries an identifier indicating the reason for removal. The cleaned data then enters the bucketed analysis. The binning engine uses the final inspection pass rate as the result quality indicator to categorize bins: a pass rate >98% is classified as a high-quality bin, 90-98% as an intermediate bin, and <90% as a risk bin. It should be noted that the >98% pass rate boundary is set based on an objective quality result indicator and remains constant in the manufacturing scenario. Then, by backtracking and extracting the torque value distribution range corresponding to each bin, it was found that the torque values ​​in the high-quality bin are concentrated in the 5.5-6.5 N·m range, which represents the optimal operating window under the current equipment and environmental conditions. When equipment aging causes the optimal torque window to shift, the torque distribution (operating window boundary) within the high-quality bucket automatically follows suit. This is because bucket division is based on the objective quality indicator of final inspection pass rate; torque values ​​that generate a high pass rate are naturally assigned to the high-quality bucket, and its distribution range reflects the latest equipment condition. Meanwhile, the boundary of the quality bucket (pass rate > 98%) remains fixed, continuously narrowing with data accumulation, and the statistical confidence level continuously improves. These two levels operate independently and without interference. Example 3: Controlled Multi-channel Distribution and Risk Broadcasting The optimal range of 5.5-6.5 N·m, confirmed by bucketing, triggers a three-way distribution: The first path archives the data to the historical experience database of the knowledge center, serving as highly reliable structured data with schema version identifiers and bucketing verification, for direct consumption and reuse by subsequent task execution units; the second path compares the data with atomic technical standard parameters in dynamic technical knowledge nodes, which are dedicated storage carriers for structured standard data. The technical standard data stored in these nodes all carry version identifiers and approval traceability chains. If the difference exceeds a 20% threshold, an upgrade change application is automatically created and submitted for approval; the third path deducts the number of screws in the ERP inventory. The platform's intelligent agent, acting as a monitoring and early warning manager for operational risks, automatically generates a procurement warning with structured instruction parameters and broadcasts it to relevant terminals when it detects a change in inventory status that triggers a preset threshold. Meanwhile, the bucketing analysis identified the 3-4 N·m range as a risk bucket, where the pass rate (87.5%) was lower than the risk level threshold (90%) defined by the schema. This was a serious quality risk identified by the bucketing analysis. The platform agent directly broadcast a risk notification carrying structured risk parameters to the agent cluster. The content included the schema identifier that triggered the risk (assembly torque parameter v2.1), the risk level (high risk), the risk pattern description (abnormal pass rate in the low torque range), the scope of affected tasks (all assembly tasks in the current batch), and structured instruction parameters that could directly trigger the receiving agent to execute a preset containment strategy (such as pausing low torque range operations or switching to high-quality range parameters). At the same time, the approval engine middleware was triggered to create a risk approval process. After parsing the structured instruction parameters, the receiving agent automatically executes the preset risk containment strategy—pausing the assembly operation in the low torque range and adjusting the torque parameters to the optimal range. After the upgrade change application is approved by the relevant approver and archived, the Agent is triggered to pull the updated parameters (5.5-6.5 N·m), write them into the dynamic technical knowledge node, and generate version 2.2. Subsequent assembly tasks automatically consume version 2.2 to achieve parameter self-optimization; Example 4: ERP Consumption and Supplier Behavior Quality Segmentation The platform agent periodically retrieves historical procurement data from the ERP system and extracts supplier behavior quality indicators (such as on-time delivery rate and quality pass rate). The data is then divided into buckets based on preset schema intervals. When a supplier's behavior quality indicator consistently falls into a risk bucket, a supplier behavior quality assessment report is automatically generated and broadcast to relevant users. Simultaneously, the approval engine middleware is triggered to create a supplier status adjustment approval process. The above disclosure is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A Schema-driven multi-agent collaboration method based on knowledge specification definition, characterized in that, include: Establish a structured knowledge base to store reusable standard document specifications, business rules and technical data standards for the enterprise. Each knowledge entry is defined as a knowledge specification definition body carrying a complete schema. The schema defines the knowledge item name, knowledge item type, the relationship between knowledge items, constraint rules, bucketing threshold and risk level. The constraint rules include data type constraints, value range constraints, and mandatory constraints. The knowledge specification definition body is encapsulated into a knowledge consumption service that can be invoked by external task execution units through a standard protocol; The external task execution unit retrieves the corresponding knowledge specification definition body through the standard protocol, and completes the knowledge item filling and structured initialization of the task instance according to the Schema definition; After the external task execution unit executes the task, it will send the multidimensional process data carrying the Schema version identifier back to the platform agent through a standard protocol. After receiving the multi-dimensional process data, the platform agent first performs legality verification and integrity checks on the data according to the constraint rules defined in the Schema, removes abnormal data, and stores the cleaned data in association with the original data to form a traceable data cleaning closed loop; the cleaned data then enters the bucket analysis. The platform agent performs bucket analysis on the cleaned data according to the bucketing threshold and risk level defined in the Schema, and identifies high-quality execution mode and low-quality execution mode. The optimal data confirmed by bucketing is distributed in a controlled multi-path manner: one copy is sent back to the historical experience database of the structured knowledge base as a high-confidence structured data archive carrying schema version identifiers and verified by bucketing, for direct consumption and reuse by subsequent task execution units; another copy is compared with the atomic technical standard parameters in the dynamic technical knowledge nodes of the product lifecycle management platform. The dynamic technical knowledge nodes are dedicated storage carriers for structured standard data. The technical standard data stored in them all carry version identifiers, approval traceability chains, and change history, which are different from ordinary database storage. When the difference exceeds a preset threshold, the parameter upgrade approval process is triggered. An updated inventory status of the enterprise resource planning system; the platform's intelligent agent, acting as a monitor and early warning manager for operational risks, automatically generates an early warning notification carrying structured instruction parameters and broadcasts it to relevant terminals when it detects a change in inventory status that triggers a preset threshold. When the bucket analysis identifies a serious quality risk pattern or a fatal error pattern, the platform agent directly broadcasts a risk notification carrying structured risk parameters to the agent cluster and triggers the approval engine middleware to create a risk approval process. The structured risk parameters can be directly parsed by the receiving agent and trigger a preset risk containment strategy. The risk notification carries the knowledge specification definition schema identifier that triggers the risk, the risk level, and the scope of the affected tasks.

2. The method of claim 1, wherein, The multidimensional process data includes process data, inspection data, and material consumption data. Each data record carries a task identifier, schema version identifier, and approval number.

3. The method of claim 1, wherein, When the external task execution unit retrieves the knowledge specification definition body through the standard protocol, it simultaneously retrieves standard document specifications from the structured knowledge base, atomic technical parameters from the dynamic technical knowledge nodes of the product lifecycle management platform, historical experience data from the structured knowledge base, and inventory and procurement data from the enterprise resource planning system, and aggregates them into a unified task initialization scheme under the constraints of the schema.

4. The method of claim 1, wherein, After the external task execution unit completes the structured initialization of the task, it retrieves the approval rule specification definition body from the structured knowledge base and sends the approval rule specification definition body and task instance data to the approval engine middleware, which is independent of the structured knowledge base, through a standard protocol. The approval engine middleware executes the approval process and generates a structured approval log with a tracking identifier. The approval log is archived to the structured knowledge base for end-to-end traceability.

5. The method of claim 1, wherein, After performing bucket analysis on the returned multi-dimensional process data, the platform's intelligent agent also includes a self-learning closed loop: broadcasting the bucketing results to relevant users and triggering a user feedback confirmation process through the approval engine middleware; receiving feedback from multiple users on the same bucketing result and performing self-reasoning and summarizing on the feedback content—integrating the feedback perspectives, feedback basis, and feedback conclusions of each user, identifying the consensus and disagreement parts in the feedback, and extracting factual conclusions; when there are contradictions in the feedback content, performing a weighted analysis on the focus of the contradiction based on the feedback basis—determining the source of the contradiction and the truth based on the historical reliability of each user's feedback, the objectivity of the feedback basis, and the consistency of the feedback conclusions; tagging the execution record based on the factual conclusions and storing the tagged execution record in the vector database of the knowledge base; periodically performing bucket analysis on the accumulated execution records, blacklisting low-quality execution patterns, and retaining high-quality execution patterns; and broadcasting the updated bucketing analysis results to relevant users again, forming a secondary feedback closed loop.

6. The method of claim 1, wherein, The data in the dynamic technical knowledge nodes of the product lifecycle management platform all come from the approved and archived tasks in the product lifecycle management platform. Each data record is associated with the source task identifier, approval number, archive time and version number, and is isolated by project identifier.

7. The method of claim 6, wherein, When the difference between the optimal parameters confirmed by the bucket and the atomic technology standard parameters of the current version in the dynamic technology knowledge node exceeds a preset threshold, the platform agent automatically creates a parameter upgrade change application and submits it to the approval engine middleware; after the upgrade change application is approved and archived, the platform agent is triggered to pull the updated parameters from the archived change tasks, write them into the dynamic technology knowledge node, and generate a new version.

8. The method of claim 1, wherein, The platform's intelligent agent obtains inventory snapshots from the enterprise resource planning system via a standard protocol at preset intervals, and dynamically calculates the safety stock threshold based on historical consumption rates. When the inventory level is lower than the safety stock threshold, a tiered early warning notification is generated and broadcast to relevant terminals, while simultaneously triggering the approval engine middleware to create a procurement approval process. When inventory levels fall below the emergency threshold, a purchase request is automatically created with material codes, suggested purchase quantities, and recommended supplier rankings, and submitted to the approval engine middleware.

9. The method of claim 1, wherein, The platform's intelligent agent obtains historical procurement data from the enterprise resource planning system through standard protocols, extracts supplier behavior quality indicators, and performs bin analysis according to preset intervals. When a supplier's key indicators continuously fall into the low-quality bin, a supplier behavior quality assessment report is automatically generated and broadcast to relevant users, while simultaneously triggering the approval engine middleware to create a supplier status adjustment approval process.

10. The method according to claim 1, characterized in that, The bucketing analysis is grouped by the schema version identifier, and the return data within the same schema version is bucketed according to the time decay weight. When the schema version changes, the bucketed data of the old and new versions are recalibrated based on the change record, so that the bucketing analysis results continue to converge as the data accumulates.

11. The method according to claim 1, characterized in that, The data cleaning closed loop specifically includes: the platform agent verifies each piece of returned data according to the value range constraints and data type constraints defined in the Schema, and removes abnormal data that exceeds the value range or does not match the data type; the removed abnormal data is stored in association with the cleaned data, carrying the removal reason identifier and verification timestamp for subsequent auditing and traceability.

12. The method according to claim 1, characterized in that, The risk notification includes: a knowledge specification definition schema identifier that triggers the risk, a risk level, a risk pattern description, the scope of affected tasks, and structured instruction parameters that can directly trigger the receiving agent to execute a preset risk containment strategy.

13. The method according to claim 1, characterized in that, The standard protocol is the MCP protocol; the external task execution unit initiates a request to the knowledge consumption service through the MCP protocol to obtain the knowledge specification definition body.

14. The method according to claim 1, characterized in that, The schema also includes its own meta-description information, which includes at least one of the following: schema version identifier, effective time, failure condition, and applicable boundaries. When the schema reaches the failure condition, the system automatically triggers the schema migration or deactivation process. The schema lifecycle management follows the automatic registration, version tracking, and failure detection mechanism of the standard template self-learning system.

15. The method according to claim 1, characterized in that, The end-to-end traceability is recorded in the form of a structured traceability tree. Each node in the traceability tree includes an operation step identifier, an execution entity identifier, an application schema version identifier, an operation timestamp, and a hash signature of the operation result. Each node is associated through a hash chain to form an immutable audit chain. The hash signature and the vector storage encryption scheme share the same key management system.

16. The method according to claim 1, characterized in that, Each agent in the agent cluster carries an agent capability schema, which defines the agent's capability type, permission level, and responsive instruction schema. When the platform's intelligent agent broadcasts a risk notification, it should include the target capability requirements in the risk notification. The receiving agent matches the target capability requirements in the risk notification with its own agent capability schema. If the match is successful, it automatically parses and executes the corresponding risk containment strategy.

17. The method according to claim 1, characterized in that, The schema adopts a hierarchical inheritance structure, including at least three levels: global schema, domain schema, and project schema. Lower-level schemas automatically inherit all constraint rules, bucketing thresholds, and risk levels from upper-level schemas. Lower-level schemas may tighten constraints within the limits allowed by upper-level schemas, but may not loosen constraints defined by upper-level schemas.

18. A multi-agent collaborative system driven by a knowledge specification definition schema, characterized in that, include: A structured knowledge base is used to store knowledge specification definition bodies carrying complete schema definitions. The schema defines the knowledge item name, knowledge item type, the relationship between knowledge items, constraint rules, bucketing thresholds and risk levels, and includes the schema's own meta-description information. The standard protocol service module is used to encapsulate the knowledge specification definition into a knowledge consumption service that can be called externally; The knowledge consumption module is used to receive knowledge call requests from external task execution units, retrieve and return the corresponding knowledge specification definition body; The multidimensional process data feedback module is used to receive multidimensional process data carrying schema version identifiers back from external task execution units; The data cleaning module is used to perform legality verification and integrity checks on the returned data according to the constraint rules defined in the Schema, remove abnormal data, and form a data cleaning closed loop. The schema-driven bucketing engine is used to perform bucketing analysis on the cleaned data according to the bucketing thresholds and risk levels defined in the schema. The multi-path controlled distribution module is used to archive the optimal data confirmed by bucketing to the knowledge base historical experience database, compare it with the atomic standards of the product lifecycle management platform, and update the inventory status of the enterprise resource planning system. The risk broadcast module is used to directly broadcast risk notifications carrying structured instruction parameters to the intelligent agent cluster when the bucket analysis identifies a serious quality risk pattern or a fatal error pattern. The receiving intelligent agents can directly parse and trigger the preset risk containment strategy, and at the same time trigger the approval engine middleware to create a risk approval process. The approval engine middleware, independent of the structured knowledge base, is used to receive approval requests, execute approval processes, and generate structured approval logs.

19. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 17.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 17.

Images