A human-machine collaborative task assignment method based on semantic vector portrait

Abstract

The application discloses a man-machine collaborative task distribution method based on semantic vector portrait, relates to the technical field of collaborative distribution, and comprises the following steps: determining a taking object from a candidate ranking, generating a distribution result, sending task content to the taking object according to the distribution result, collecting an execution state of the task content, and judging whether an artificial intelligence body satisfies a help-seeking condition; when the artificial intelligence body satisfies the help-seeking condition, reading the task content and the execution state, extracting a help-seeking matter from the execution state, and writing the task content, the execution state and the help-seeking matter in a preset help-seeking record format to generate a help-seeking record and update the distribution result according to the help-seeking record; and the taking object is determined again through the help-seeking record, the task content can be transferred to a more matched artificial intelligence body or operator, and the remediation efficiency, distribution accuracy and man-machine collaborative stability of an abnormal task are improved.

Description

Technical Field

[0001] This invention relates to the field of collaborative task assignment technology, and in particular to a human-computer collaborative task assignment method based on semantic vector profiling. Background Technology

[0002] With the development of large language models, semantic representation learning, vector retrieval, and knowledge graph technologies, task assignment methods have gradually evolved from manual rule-based assignment and keyword matching assignment to intelligent assignment based on semantic understanding and capability profiling. In recent years, AI agents have gained the ability to invoke tools, process text, analyze attachments, and generate intermediate results. Operators can also participate in complex task processing through online collaboration. Human-machine collaborative task processing is gradually becoming an important direction in intelligent office, crowdsourcing collaboration, and enterprise knowledge processing.

[0003] However, existing task assignment methods still have room for improvement. First, existing methods focus primarily on generating initial assignment results, neglecting the collection of execution status data and anomaly detection after task content is sent. This is especially true when AI agents encounter processing failures, incomplete outputs, insufficient permissions, unavailable tools, or timeouts during task execution, lacking a mechanism to determine whether the AI ​​agent meets the request-for-help criteria based on its execution status. Furthermore, existing methods typically save AI agent execution anomalies as failures or permanently transfer abnormal tasks to manual processing. This makes it difficult to extract request-for-help information that characterizes the cause of the anomaly from the task content and execution status, and it is also difficult to generate request-for-help records in a standardized format for reassignment. Consequently, the cause of the anomaly cannot be transformed into a basis for updating the assignment results, affecting the continuous flow of task content between the AI ​​agent and the operator. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a human-computer collaborative task assignment method based on semantic vector profiling to solve the problem of lacking a processing mechanism that determines whether an AI agent meets the request conditions based on its execution state.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a human-computer collaborative task assignment method based on semantic vector profiling, which includes,

[0008] Acquire AI agent access information and operator access information, register them as work object information, and construct a semantic network based on the work object information;

[0009] Semantic encoding of work object information is performed based on semantic networks to generate work profiles, and information on tasks to be assigned is obtained. Semantic parsing of the information on tasks to be assigned is performed based on semantic networks to generate task profiles.

[0010] Semantic matching of task profiles and work profiles is performed to obtain candidate objects. The candidate object ranking is formed by the capability coverage results, load status and historical quality of the candidate objects. The receiving object is determined from the candidate ranking and the assignment result is generated.

[0011] Based on the assignment results, the task content is sent to the receiving object, the execution status of the task content is collected, and it is determined whether the AI ​​agent meets the request conditions. When the AI ​​agent meets the request conditions, the task content and execution status are read, the request items are extracted from the execution status, and the task content, execution status and request items are written according to the preset request record format to generate a request record. The assignment results are updated based on the request record.

[0012] Based on the assignment results and the updated assignment results, the effective assignment results are determined. Based on the effective assignment results, the task content is sent to the effective recipients. The execution results returned by the effective recipients are received, and the execution status and execution results are verified. A completion record is generated, and the work profile and semantic network are updated using the completion record.

[0013] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the AI ​​agent access information includes AI agent identifier, interface address, calling permissions, declared skills, tool calling scope, historical calling records, availability status, response latency, quotation information, failure records, and help records.

[0014] The operator access information includes operator ID, login credentials, skill tags, permission tags, available time periods, calibration task results, historical completion quality, current load, response latency, quotation information, failure records, and help records.

[0015] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the step of constructing a semantic network based on work object information specifically includes:

[0016] Read the declared skills, skill tags, access permissions, access permission tags, tool access scope, availability status, historical access records, calibration task results, historical completion quality, current load, failure records, and help records from the work object information;

[0017] The declared skills and skill tags are grouped into ability tags, the call permissions and permission tags are grouped into permission tags, the tool call scope is grouped into tool tags, and the historical call records, calibration task results, historical completion quality, failure records and help records are grouped into feedback tags.

[0018] The capability label, permission label, tool label, availability status, current load, and feedback label are mapped to capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes in the semantic network, respectively.

[0019] Establish semantic relationships between work object information and capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes to form a semantic network.

[0020] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the generation of the work profile specifically includes:

[0021] Read the capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes corresponding to the working object information in the semantic network;

[0022] Vectorize the capability nodes, permission nodes, and tool nodes to obtain the capability vector;

[0023] The state nodes are encoded to obtain state vectors. The number of pending tasks, the number of tasks in processing, and the load occupancy in the load nodes are numerically organized to obtain a load value table.

[0024] The feedback nodes are quality encoded to obtain the feedback vector;

[0025] By associating capability vectors, state vectors, and feedback vectors with work object information, a work profile is generated.

[0026] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the generation of task profiling specifically includes:

[0027] Read the task text, task attachments, task objectives, task domain, output requirements, time limits, and permission requirements from the information of the task to be assigned;

[0028] The task text, task attachments, task objectives, and task domain are mapped to capability nodes and tool nodes in the semantic network to obtain the task semantics and capability requirements.

[0029] The output requirements are organized into fields to obtain result status restrictions, the time limit requirements are organized into time constraints to obtain time continuity restrictions, the permission requirements are mapped to permission nodes in the semantic network to obtain permission access restrictions, and the result status restrictions, time continuity restrictions, and permission access restrictions form task constraints.

[0030] The task semantics are vectorized to obtain the task vector, and the task constraints are constrained to obtain the constraint vector.

[0031] The task vector, capability requirements, and constraint vector are associated with the task information to be assigned to generate a task profile.

[0032] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the step of obtaining candidate objects specifically involves:

[0033] Read the task vector and constraint vector from the task profile, and read the capability vector, state vector and feedback vector from the job profile;

[0034] Semantic similarity matching is performed between the task vector and the capability vector to obtain the capability coverage result;

[0035] Read the time acceptance limit and access restriction from the task constraints corresponding to the constraint vector, compare the time acceptance limit with the state vector, available time period and response latency for time adaptation, compare the access restriction with the permission node for field comparison, and compare the load value table with the preset load conditions to obtain the load status.

[0036] The historical quality is obtained by comparing the feedback vector with the task domain in the task profile and performing a quality correlation comparison.

[0037] Based on the capacity coverage results, load status, and historical quality, candidate objects are obtained by filtering the work profiles.

[0038] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the generation of assignment results specifically includes:

[0039] Read the capability coverage results, load status, and historical quality of the candidate object;

[0040] The capability coverage results are compared with the capability requirements in the task profile to obtain the acceptance matching mark. The load status is compared with the preset load conditions to obtain the load occupancy mark. The historical quality is compared with the task domain in the task profile to obtain the quality stability mark.

[0041] The matching markers, load occupancy markers, and quality stability markers are collected and arranged according to a preset evaluation order to obtain candidate evaluation results;

[0042] The candidates are ranked according to the evaluation results to form a candidate ranking.

[0043] Read the candidate objects that meet the acceptance conditions from the candidate sort, determine them as acceptance objects, establish the correspondence between acceptance objects and task content, and generate the assignment results.

[0044] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the step of collecting the execution status of the task content and determining whether the artificial intelligence agent meets the request-for-help conditions specifically involves:

[0045] Read the receiving object and task content from the assignment result;

[0046] Send the task content to the recipient and receive the processing progress, exception flags, completion flags and intermediate results returned by the recipient;

[0047] Save the processing progress, exception flags, completion flags, and intermediate results as the execution status;

[0048] When the target is an AI entity, the execution status will be compared with the preset request conditions;

[0049] When the execution status characterizes any of the following states: processing failure, incomplete output, insufficient permissions, unavailable tools, or timeout, the AI ​​agent is deemed to meet the request for assistance.

[0050] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the step of updating the assignment result according to the request record specifically includes:

[0051] When the AI ​​agent meets the request conditions, read the task content and execution status;

[0052] Extract the processing failure flag, incomplete output flag, insufficient permissions flag, tool unavailable flag, and timeout flag from the execution status to form a help request;

[0053] Organize the fields for task content, execution status, and requests for assistance to generate a request record;

[0054] Mapping the requests for help in the request records to a semantic network yields a request profile.

[0055] Semantic matching is performed between the request-for-help profile and the work profile to obtain updated candidate objects;

[0056] Based on the updated candidate objects, a new matching is performed to form an updated candidate ranking, and the updated recipient objects are determined from the updated candidate ranking to generate the updated assignment results.

[0057] As a preferred embodiment of the human-computer collaborative task assignment method based on semantic vector profiling described in this invention, the generation of completion records specifically includes:

[0058] Read the update recipient and task content from the update assignment results;

[0059] Based on the updated assignment results, the task content is sent to the newly determined recipient, and the execution results are received from the newly determined recipient.

[0060] Read the execution status and help records corresponding to the task content, extract the processing progress, exception markers and completion markers from the execution status, and extract the help requests from the help records;

[0061] The consistency of processing progress, exception markers, completion markers, and requests for help is verified with the execution results to obtain the verification results.

[0062] When the verification result meets the preset completion conditions, the execution status and execution result are written into the preset completion record format to generate a completion record.

[0063] The beneficial effects of this invention are as follows: By updating the assignment results through request records, the AI ​​agent can avoid directly interrupting the task when processing fails, outputs are incomplete, permissions are insufficient, tools are unavailable, or timeouts occur. Instead, the cause of the anomaly is transformed into the basis for subsequent reassignment, improving the continuous processing capability of the task content. At the same time, the request records can retain the correspondence between the task content, execution status, and request items, making the expression of the cause of the anomaly clearer and avoiding the rough assignment problem caused by fixed transfer to human. Furthermore, by re-determining the receiving object based on the request records, the task content can be transferred to a more suitable AI agent or operator, improving the efficiency of remediation of abnormal tasks, the accuracy of assignment, and the stability of human-machine collaboration. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 The flowchart shows a human-computer collaborative task assignment method based on semantic vector profiling.

[0066] Figure 2 A flowchart for forming a semantic network.

[0067] Figure 3 A flowchart for obtaining job profiles and task profiles.

[0068] Figure 4 This is a flowchart for generating the assignment results.

[0069] Figure 5 Flowchart for updating and completing rescue records. Detailed Implementation

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0071] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0072] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0073] Reference Figures 1-5 This is one embodiment of the present invention, which provides a human-computer collaborative task assignment method based on semantic vector profiling, including the following steps:

[0074] S1: Obtain AI agent access information and operator access information, register them as work object information, and construct a semantic network based on the work object information.

[0075] S1.1: Obtain AI Agent Access Information: Read the AI ​​agent identifier as the unique registration basis for the AI ​​agent; read the interface address as the communication entry point for the AI ​​agent to receive task content and return processing results; read the calling permissions and tool calling scope to determine the boundaries of task content that the AI ​​agent can access and the scope of tool use; read the declared skills to determine the task domain that the AI ​​agent can undertake; read historical call records, failure records, and help records to determine the AI ​​agent's past execution status; read the availability status, response latency, and quotation information to determine whether the AI ​​agent can currently undertake task content and the time and cost required to undertake the task content.

[0076] Obtain operator access information: Read the operator's identifier and login credentials as the basis for operator identity verification; read the skill tags and permission tags to determine the task areas and access boundaries that the operator can undertake; read the available time period to determine the time range within which the operator can receive task content; read the calibration task results, historical completion quality, failure records, and help records to determine the operator's past performance; read the current load, response latency, and quotation information to determine the operator's current workload, processing speed, and cost requirements.

[0077] The AI ​​agent access information and operator access information are organized according to the same registration fields, so that both AI agents and operators can form work object information that can be registered, compared and can participate in the subsequent semantic network construction.

[0078] S1.2: Read the declared skills, skill tags, calling permissions, permission tags, tool calling scope, availability status, historical calling records, calibration task results, historical completion quality, current load, failure records, and help records from the work object information.

[0079] Extract skill names, task domains, and skill descriptions from declared skills and skill tags. Merge skill names with consistent meanings into unified skill names, categorize unified skill names according to task domains, and remove duplicate unified skill names to form capability tags. Extract access objects, access scopes, and operation boundaries from call permissions and permission tags. Organize the access scopes of AI agents and operators according to the field order of access object, readable scope, processable scope, and submitable scope to form unified permission expressions. Arrange the unified permission expressions according to the operation boundaries of readable, processable, and submitable. When the same access object has readable scope in the call permissions and permission tags, the access scope is determined by the operation boundaries. When the scope, processable scope, or submittable scope are inconsistent, the inconsistent field content is marked as inconsistent permission content, forming permission tags. The callable tool name, callable tool type, and callable tool boundary are extracted from the tool call scope, and these are then organized to form tool tags. Completion results, quality evaluations, failure reasons, and request reasons are extracted from historical call records, calibration task results, historical completion quality, failure records, and request request records. Completion results are categorized as completion status, quality evaluations as quality status, failure reasons as failure status, and request reason as request request status, forming feedback tags. The availability status and current load are retained as the basis for subsequent mapping of status nodes and load nodes.

[0080] S1.3: Based on the capability tag name, task domain, and source work object information read from the capability tag, organize them according to the capability node field format to generate capability nodes, and record the capability tag name, task domain, and source work object information in the capability nodes; read the access scope, operation boundary, and source work object information read from the permission tag, organize them according to the permission node field format to generate permission nodes, and record the access scope, operation boundary, and source work object information in the permission nodes; read the callable tool name, callable tool type, callable tool boundary, and source work object information read from the tool tag, organize them according to the tool node field format to generate tool nodes, and record the callable tool name, callable tool type, callable tool boundary, and source work object information in the tool nodes; read the online status from the availability status. The system reads and organizes information on online status, receivable status, paused status, and source work object according to the status node field format, generating status nodes and recording the online status, receivable status, paused status, and source work object information in each status node. It also reads and organizes information on the number of pending tasks, the number of tasks in process, load occupancy, and source work object from the current load, according to the load node field format, generating load nodes and recording the number of pending tasks, the number of tasks in process, load occupancy, and source work object information in each load node. Finally, it reads and organizes information on completion status, quality status, failure status, request for help status, and source work object from the feedback tags, according to the feedback node field format, generating feedback nodes and recording the completion status, quality status, failure status, request for help status, and source work object information in each feedback node.

[0081] Semantic relationships are established between capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes that share the same source work object information and the work object information. Specifically, a capability-enabled relationship is established between work object information and capability nodes; a permission-enabled relationship is established between work object information and permission nodes; a callable tool relationship is established between work object information and tool nodes; an available status relationship is established between work object information and status nodes; a load relationship is established between work object information and load nodes; and a historical feedback relationship is established between work object information and feedback nodes. A semantic network is formed by capability nodes, permission nodes, tool nodes, status nodes, load nodes, feedback nodes, and semantic relationships. The semantic network is stored using a graph structure, with the source work object information and node type serving as unique identifiers for each node. Semantic relationships record the relationship type, relationship weight, and update time. Relationship weights are determined based on capability coverage, historical completion quality, and call frequency, and the update time records the most recent update of the semantic relationship.

[0082] S2: Based on the semantic network, semantically encode the information of the work object to generate a work profile, and obtain the information of the tasks to be assigned. Based on the semantic network, semantically parse the information of the tasks to be assigned to generate a task profile.

[0083] S2.1: Extract declared skills, skill tags, calling permissions, permission tags, tool calling scope, availability status, historical calling records, calibration task results, historical completion quality, current load, failure records, and help records from historical AI agent access information and historical operator access information to form historical work object samples; extract task text, task attachment text content, task objectives, task domain, output requirements, time limit requirements, and permission requirements from historical pending task information and historical completed task information to form historical task samples; extract node names, node types, node content, and semantic relationships from capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes in the historical semantic network to form historical node samples; extract task profiles, work profiles, receiving objects, execution status, verification results, and completion quality from historical assignment results, historical updated assignment results, historical execution results, and historical completion records to form historical matching samples; use the historical work object samples, historical task samples, historical node samples, and historical matching samples together as the training set for the semantic coding model.

[0084] The semantic coding model uses a language representation network as the basic coding network, and is connected in the following order: input layer, word vector layer, positional coding layer, multi-level bidirectional attention coding layer, pooling layer, and output layer. The input layer receives node names, node types, node content, semantic relationships, task text, task objectives, and task constraint text. The word vector layer converts the text received from the input layer into word vectors. The positional coding layer adds positional order representations to the word vectors. The multi-level bidirectional attention coding layer performs contextual association calculations on the word vectors and positional order representations. The pooling layer summarizes the sequence vectors output by the multi-level bidirectional attention coding layer. The output layer outputs capability vectors, state vectors, feedback vectors, task vectors, and constraint vectors according to the type of the encoded object. The capability vector consists of capability sub-vectors corresponding to capability node fields, permission sub-vectors corresponding to permission node fields, and tool sub-vectors corresponding to tool node fields, arranged in order of field. The task vector is formed by concatenating the task capability sub-vectors corresponding to the capability requirement field, the task tool sub-vectors corresponding to the tool scope field, and the task domain sub-vectors corresponding to the task domain field in the order of the fields. The constraint vector is formed by concatenating the result constraint sub-vectors corresponding to the result state restriction, the time constraint sub-vectors corresponding to the time constraint, and the permission constraint sub-vectors corresponding to the access restriction in the order of the fields. The key parameters of the semantic coding model include word vector dimension, maximum input length, number of bidirectional attention coding layers, number of attention heads, hidden layer dimension, output vector dimension, learning rate, and training epochs. For example, the word vector dimension is set to 768, the maximum input length is set to 512, the number of bidirectional attention coding layers is set to 12, the number of attention heads is set to 12, the hidden layer dimension is set to 3072, the output vector dimension is set to 768, and the learning rate is set to 2×10⁻. 5 .

[0085] Historical work object samples, historical task samples, and historical node samples are organized into a unified text input. Historical matching samples are organized into supervision tags, including capability matching tags, permission matching tags, tool matching tags, state adaptation tags, quality matching tags, and task completion tags. When the capability node corresponding to the work object information covers the capability requirements in the historical task profile, the capability matching tag is marked as 1; otherwise, it is marked as 0. When the access object, access scope, and operation boundary in the permission node cover the permission access restrictions in the historical task profile, the permission matching tag is marked as 1; otherwise, it is marked as 0. When the callable tool type and callable tool boundary in the tool node cover the tool range in the historical task profile, the tool matching tag is marked as 1. The label is set to 1 if the work object information is in an acceptable state, the available time period meets the time acceptance limit, the response latency meets the response time limit, and the load value table meets the preset load conditions. Otherwise, it is set to 0. The quality matching label is set to 1 if the quality score in the same task domain is not lower than 80 points, the failure record does not exceed the preset number of failures, and the help record does not exceed the preset number of help requests. Otherwise, it is set to 0. When the historical completion mark indicates that the task content is completed and the historical verification result is verification passed, the task completion label is set to 1. Otherwise, it is set to 0. The positive samples of the corresponding training task are formed by the samples with a label value of 1, and the negative samples of the corresponding training task are formed by the samples with a label value of 0.

[0086] The unified text input is fed into the semantic coding model to obtain the training capability vector, training state vector, training feedback vector, training task vector, and training constraint vector.

[0087] The cosine similarity is calculated between the task capability subvector in the training task vector and the capability subvector in the training capability vector, and the result is normalized to the range of 0 to 1 to obtain the capability matching prediction result. Similarly, the cosine similarity is calculated between the permission constraint subvector in the training constraint vector and the permission subvector in the training capability vector, and the result is normalized to the range of 0 to 1 to obtain the permission matching prediction result. Likewise, the cosine similarity is calculated between the task tool subvector in the training task vector and the tool subvector in the training capability vector, and the result is normalized to the range of 0 to 1 to obtain the tool matching prediction result. Finally, the training state vector and available time period are considered. Footprint, normalized response latency, and normalized load occupancy are concatenated in field order and processed by fully connected computation and Sigmoid normalization to obtain the state adaptation prediction result. The task domain sub-vectors in the training task vector are compared with the training feedback vector by cosine similarity calculation, and the result is normalized to the range of 0 to 1 to obtain the quality matching prediction result. The capability matching prediction result, permission matching prediction result, tool matching prediction result, state adaptation prediction result, and quality matching prediction result are concatenated in a fixed order and processed by fully connected computation and Sigmoid normalization to obtain the task completion prediction result.

[0088] The cross-entropy loss of binary classification is used to calculate the loss between the prediction results of ability matching, permission matching, tool matching, state adaptation, quality matching, and task completion and their corresponding supervision labels. Each task loss is divided by the corresponding initial loss value in the initial training round to obtain the normalized loss of ability matching, permission matching, tool matching, state adaptation, quality matching, and task completion. The six normalized task losses are averaged with equal weights to obtain the total loss of the multi-task model, and the semantic encoding model parameters are updated only based on the total loss of the multi-task model.

[0089] To further explain, the convergence condition is set before the semantic coding model is trained. The loss of each task is normalized according to the initial loss value in the initial training round. When the normalized total loss of multiple tasks decreases by no more than 0.001 for three consecutive rounds, and the normalized ability matching loss, permission matching loss, tool matching loss, state adaptation loss, quality matching loss and task completion loss do not increase, the semantic coding model is determined to have reached the preset convergence condition.

[0090] S2.2: Based on the AI ​​agent identifier and operator identifier in the work object information, search for capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes that are consistent with the source work object information in the semantic network; read the capability, permission, and tool content recorded in the capability nodes, permission nodes, and tool nodes; read the available status and current load recorded in the status nodes and load nodes; read the completion status, quality status, failure status, and help status recorded in the feedback nodes to obtain the capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes corresponding to the work object information.

[0091] Read the capability tag name and task domain from the capability node, read the access scope and operation boundary from the permission node, and read the callable tool name, callable tool type, and callable tool boundary from the tool node. Concatenate the capability tag name, task domain, access scope, operation boundary, callable tool name, callable tool type, and callable tool boundary into capability encoded text according to a fixed field order. The fixed field order is capability node field, permission node field, and tool node field. Different fields in the capability encoded text are distinguished by field separators.

[0092] After the capability-encoded text is input into the semantic encoding model, the word vector layer converts the field names, field values, and field separators in the capability-encoded text into a word vector sequence. The positional encoding layer adds the field order to the word vector sequence to obtain the encoded input sequence. The multi-layer bidirectional attention encoding layer uses the encoded input sequence as the first layer input. In each layer, a query vector, key vector, and value vector are generated. The query vector and key vector are multiplied by a dot product, and then divided by the square root of the vector dimension. After normalization, attention weights are obtained. The attention weights are then used to sum the value vectors to obtain the context vector corresponding to each field position. The attention weight of the capability label name corresponding to the field position to the task domain corresponding to the field position represents the association between the capability label name and the task domain. The attention weight of the access scope corresponding to the field position represents the association between the capability label name and the task domain. The attention weights corresponding to the field positions of the operation boundary represent the association between the access range and the operation boundary. The attention weights corresponding to the field positions of the callable tool name and the callable tool type and the callable tool boundary represent the association between the callable tool name, the callable tool type, and the callable tool boundary. The overall association between capability nodes, permission nodes, and tool nodes is determined by the attention weights corresponding to the field positions of capability nodes, permission nodes, and tool nodes. The multi-layer bidirectional attention encoding layer outputs a sequence vector containing the context vectors of all field positions. The pooling layer performs average pooling on the sequence vector to obtain the capability summary vector. The output layer performs linear calculations on the capability summary vector with the trained output weight matrix and output bias vector to obtain the capability vector.

[0093] The system reads the online, acceptable, and paused states from the status nodes. These states are then concatenated in field order to form a status-encoded text. This text is input into the semantic coding model, where the input layer, word vector layer, positional encoding layer, multi-layer bidirectional attention encoding layer, pooling layer, and output layer complete the status encoding in the same order, resulting in a status vector. The system also reads the number of pending tasks, the number of tasks in progress, and the load occupancy from the load nodes. These information is saved as a load value table. This load value table is not used in the semantic coding model's text vector calculation; instead, it is used in subsequent constraint adaptation comparisons based on preset load conditions to determine if the work object information meets the load-bearing requirements.

[0094] The completion status, quality status, failure status, and request for help status are read from the feedback node. The completion status, quality status, failure status, and request for help status are concatenated into feedback encoded text according to a fixed field order. After the feedback encoded text is input into the semantic encoding model, the input layer, word vector layer, position encoding layer, multi-layer bidirectional attention encoding layer, pooling layer, and output layer complete the quality encoding in the same encoding order to obtain the feedback vector.

[0095] S2.3: Using the AI ​​agent identifier and operator identifier corresponding to the work object information as the association basis, write the capability vector into the capability representation position corresponding to the work object information, write the state vector into the state representation position corresponding to the work object information, and write the feedback vector into the feedback representation position corresponding to the work object information; bind the capability vector, state vector, and feedback vector sequentially to form a vector combination corresponding to the work object information; maintain the correspondence between the vector combination and the AI ​​agent access information and operator access information in the work object information to generate a work profile that can characterize the capability characteristics, state characteristics, and feedback characteristics of the work object information.

[0096] S2.4: Read the task text, task attachments, task objectives, and task domain from the task information to be assigned, and extract the processing object and processing action from the task text, the attachment content and attachment type from the task attachments, the completion purpose and output direction from the task objective, and the domain name from the task domain; perform semantic comparison between the processing object, processing action, completion purpose, and domain name and the capability nodes in the semantic network, specifically: match the processing object with the capability tag name in the capability node, calculate the semantic similarity between the processing action and completion purpose and the capability description in the capability node, and compare the domain name with the task domain in the capability node for field consistency; when the domain name matches the task domain, and the semantic similarity between the processing action and completion purpose reaches the preset capability matching threshold, record the corresponding capability node as a matching capability node, and then proceed according to the matching capability node. The capability scope corresponding to the task information to be assigned is determined. The attachment content, attachment type, processing action, and output direction are semantically compared with tool nodes in the semantic network. Specifically, the attachment type is compared for field consistency with the callable tool types in the tool node; the attachment content and processing action are matched with the callable tool names in the tool node using keywords; and the output direction is calculated for semantic similarity with the callable tool boundaries in the tool node. When the attachment type matches the callable tool type, and the semantic similarity between the output direction and the callable tool boundary reaches a preset tool matching threshold, the corresponding tool node is recorded as a matching tool node. Based on the matching tool node, the tool scope corresponding to the task information to be assigned is determined. The capability scope and tool scope are organized into task semantics, and the capability content required to undertake the task information to be assigned is extracted from the capability scope to form capability requirements.

[0097] To further explain, the capability matching threshold is based on historical task information to be assigned, historical capability nodes, and historical completion quality settings. When the historical completion quality meets the qualified standard, the lower limit of the semantic similarity between the processing action and the completion purpose and the capability description in the capability node is used as the capability matching threshold, which is set to 0.75. The tool matching threshold is based on historical task information to be assigned, historical tool nodes, and historical execution results. When the historical execution results meet the output requirements, the lower limit of the semantic similarity between the output direction and the callable tool boundary in the tool node is used as the tool matching threshold, which is set to 0.70.

[0098] S2.5: Extract the result format, format requirements, and quality requirements from the output requirements. Organize these requirements into output constraints, and record them in the order of the result format field, format requirement field, and quality requirement field to form result status constraints. Do not match output constraints with status nodes again. Extract the completion deadline, response deadline, and stage deadline from the time limit requirements. Organize these requirements into time constraints, and record them in the order of the completion deadline field, response deadline field, and stage deadline field to form time continuity constraints. Next, the time constraints are compared with the status nodes corresponding to the online, acceptable, and paused states. The access object, access scope, and operation boundary are extracted from the permission requirements and organized into permission constraints. The permission nodes with the same access object are searched in the semantic network. The access scope and operation boundary in the permission constraints are matched with the access scope and operation boundary in the permission nodes. The access object, required access scope, and required operation boundary are recorded as permission access restrictions. The result status restrictions, time continuity restrictions, and permission access restrictions are merged and organized to obtain the task constraints.

[0099] S2.6: The capability scope and tool scope are read from the task semantics, and the capability content required for receiving and assigning tasks is read from the capability requirements. The capability scope, tool scope, and capability content are concatenated into task-encoded text according to the order of task semantic fields, tool scope fields, and capability requirement fields. Different fields in the task-encoded text are distinguished by field separators. After the task-encoded text is input into the semantic encoding model, the word vector layer converts the field names, field values, and field separators in the task-encoded text into a word vector sequence. The positional encoding layer adds the field order to the word vector sequence. The multi-layer bidirectional attention encoding layer takes the word vector sequence as input, calculates the attention weights between field positions through query vectors, key vectors, and value vectors, and obtains the context vector corresponding to each field position based on the attention weights. The pooling layer performs average pooling on the sequence vector containing the context vectors of all field positions to obtain the task summary vector. The output layer performs linear calculations on the task summary vector with the trained output weight matrix and output bias vector to obtain the task vector. The result status constraint, time continuity constraint, and access restriction are read from the task constraints. The result status constraint, time continuity constraint, and access restriction are concatenated into constraint encoded text in the order of result status field, time continuity field, and access restriction field. After the constraint encoded text is input into the semantic encoding model, the word vector layer, position encoding layer, multi-layer bidirectional attention encoding layer, pooling layer, and output layer complete the constraint encoding according to the encoding method of the task encoded text to obtain the constraint vector.

[0100] S2.7: Using the task information to be assigned as the associated object, write the task vector into the task semantic representation position corresponding to the task information to be assigned, write the capability requirements into the acceptance capability representation position corresponding to the task information to be assigned, and write the constraint vector into the task constraint representation position corresponding to the task information to be assigned; bind the task vector, capability requirements, and constraint vector sequentially to form the task representation combination corresponding to the task information to be assigned; maintain the correspondence between the task representation combination and the task text, task attachments, task objectives, task domain, output requirements, time limits, and permission requirements in the task information to be assigned, and generate a task profile that can represent the semantic features, acceptance capability requirements, and constraints of the task information to be assigned.

[0101] S3: Perform semantic matching between the task profile and the work profile to obtain candidate objects. Form a candidate ranking based on the capability coverage results, load status and historical quality of the candidate objects. Determine the receiving object from the candidate ranking and generate the assignment result.

[0102] S3.1: Read the task vector and constraint vector from the task profile, and the capability vector, state vector, and feedback vector from the work profile; perform a dimensionality consistency check on the task vector and capability vector. If the dimensions are consistent, normalize the task vector and capability vector, calculate the cosine similarity between the normalized task vector and capability vector, and use the cosine similarity as the semantic matching value between the task profile and the work profile; compare the semantic matching value with a preset capability coverage threshold. If the semantic matching value reaches the preset capability coverage threshold, mark the work object information corresponding to the work profile as capability covered; if the semantic matching value does not reach the preset capability coverage threshold, mark the work object information corresponding to the work profile as capability uncovered, and obtain the capability coverage result.

[0103] To further explain, the capability coverage threshold is set based on historical assignment results, historical execution results, and historical completion quality. The lower limit of the cosine similarity between the task vector and the capability vector when the historical completion quality meets the qualified standard is used as the capability coverage threshold, for example, it is set to 0.75.

[0104] Read the result status constraints, time constraints, and access restrictions from the task constraints, and read the status vector, permission nodes, and load value table corresponding to the work profile. The status vector is only used to determine whether the work object information is in an acceptable state and is no longer used to override access restrictions and result status constraints. Read the access object, access scope, and operation boundary from the access restrictions, and compare the fields for consistency with the access object, access scope, and operation boundary in the permission nodes. If the fields match, record that the permission is satisfied; if the fields do not match, record that the permission is not satisfied. Read the completion deadline, response time limit, and stage time limit from the time constraints, and compare the time conditions with the available time period and response delay in the work object information. If the completion deadline, response time limit, and stage time limit are all satisfied by the available time period and response delay, record that the time is satisfied; if any time condition is not satisfied, record that the time is not satisfied. Read the load... The system loads the number of pending tasks, the number of tasks in progress, and the load status from the numerical table, and compares these values ​​with preset load conditions. If the load does not exceed the preset load conditions, it is recorded as load satisfied; otherwise, it is recorded as load unsatisfied. When the state vector represents an acceptable state, satisfied permissions, satisfied time, and satisfied load, the work object information corresponding to the work profile is marked as load acceptable. When the state vector represents a paused state, unsatisfied permissions, unsatisfied time, or unsatisfied load, the work object information corresponding to the work profile is marked as load unacceptable, thus obtaining the load status. These status restrictions are retained for use during execution result verification.

[0105] To further explain, the load condition is based on the current load, historical call records, and historical completion quality settings in the historical work object information. It is used to compare the number of pending tasks, the number of tasks in process, and the load occupancy in the load value table. For example, if the number of pending tasks does not exceed 5, the number of tasks in process does not exceed 3, and the load occupancy does not exceed 80%, the work object information corresponding to the work profile is determined to meet the load condition.

[0106] S3.2: Read the task domain from the task profile, organize the task domain into domain comparison text according to domain name and domain keywords, concatenate the domain comparison text according to the order of the domain name field and the domain keyword field, and input the domain comparison text into the semantic encoding model. The word vector layer, position encoding layer, multi-layer bidirectional attention encoding layer, pooling layer and output layer encode it according to the encoding method of the task vector to obtain the domain vector; calculate the cosine similarity between the domain vector and the feedback vector to obtain the quality association value; read the completion status, quality status, failure status and request for help status from the feedback node corresponding to the feedback vector, and read the information of the same work object in the same... The system reads the output content, output format, completion time, time limit, and anomaly flags from the historical completion records within a task domain. Each historical completion record is then scored for quality: complete output content scores 40 points, conforming output format scores 20 points, meeting the time limit score scores 20 points, and anomaly flags indicating new anomalies score 20 points. Any unmet score item scores 0 points. The scores from these four items are added together to obtain the single-item quality score. The average of these single-item quality scores within the same task domain is then calculated to obtain the task domain quality score. A task domain quality score of 80 or higher is recorded as satisfactory, while a score below 80 is recorded as unsatisfactory.

[0107] When the quality correlation value reaches the preset quality correlation threshold, and the completion status of the feedback vector is "completed," the quality status is "qualified," the failure count does not exceed the preset number of failures, and the number of requests for help does not exceed the preset number of requests for help, the work object information corresponding to the work profile is marked as historically qualified. When the quality correlation value does not reach the preset quality correlation threshold, or when there are contents in the completion status, quality status, failure status, and request for help corresponding to the feedback vector that do not meet the preset quality conditions, the work object information corresponding to the work profile is marked as historically unqualified, and historical quality is obtained. The work object information with the capability coverage result being capability covered, the load status being load-bearing, and the historical quality being historically qualified is selected to obtain candidate objects.

[0108] To further explain, the quality association threshold is based on historical task domains, historical feedback vectors, and historical completion quality settings. The lower limit of the cosine similarity between the domain vector and the feedback vector corresponding to the historical completion quality meeting the qualified standard is used as the preset quality association threshold, for example, set to 0.70. The historical completion quality meeting the qualified standard means that the verification result in the historical completion record is that the verification passed and the quality score is not lower than 80 points. The number of failures is based on historical failure records and historical completion quality settings. The upper limit of the number of failure records allowed for the same work object information in the same task domain while maintaining the historical completion quality meeting the qualified standard is used as the preset number of failures, for example, set to 2 times. The number of requests for help is based on historical request for help records and historical completion quality settings. The upper limit of the number of request for help records allowed for the same work object information in the same task domain while maintaining the historical completion quality meeting the qualified standard is used as the preset number of requests for help, for example, set to 3 times.

[0109] S3.3: Retrieve capability coverage results, load status, and historical quality from candidate objects, and retrieve capability requirements and task domains from task profiles. Simultaneously, retrieve capability nodes from the work profiles corresponding to candidate objects. Compare the capability tag names and task domains in the capability nodes corresponding to the work profiles with the capability requirements item by item. When the capability tag names and task domains in the capability nodes cover the capability content in the capability requirements, and the capability coverage result is capability coverage, generate a matching mark as a match. When the capability tag names and task domains in the capability nodes do not cover the capability content in the capability requirements, generate a mismatch mark as a mismatch. Analyze the number of pending tasks in the load status... The number of tasks and load occupancy during processing are compared with preset load conditions. When the load status is that the load can be handled and does not exceed the preset load conditions, a load occupancy mark is generated as "handling". When the load status is that the load cannot be handled or exceeds the preset load conditions, a load occupancy mark is generated as "unhandling". The quality records of task domains in the historical quality are compared with the quality records of task domains in the task profile. When the historical quality is qualified and the quality records of task domains are consistent with the task domains, a quality stability mark is generated as "stable". When the historical quality is unqualified or the quality records of task domains are inconsistent with the task domains, a quality stability mark is generated as "unstable".

[0110] S3.4: Write the acceptance matching flag, load occupancy flag, and quality stability flag into the candidate evaluation record corresponding to the same candidate object, and arrange them according to the preset evaluation order of the acceptance matching flag, load occupancy flag, and quality stability flag; when the acceptance matching flag is "matched," the load occupancy flag is "acceptable," and the quality stability flag is "stable," the candidate evaluation record is determined as a candidate evaluation result that meets the acceptance conditions; when any one of the following flags exists: acceptance matching flag is "mismatched," load occupancy flag is "unacceptable," and quality stability flag is "unstable," the candidate evaluation record is determined as a candidate evaluation result that does not meet the acceptance conditions; for candidate objects that meet the acceptance conditions, continue reading... The candidate objects are ranked according to semantic matching value, load utilization rate corresponding to load occupancy, quality correlation value, response latency, and pricing information, in the following order: semantic matching value from high to low, load utilization rate from low to high, quality correlation value from high to low, response latency from low to high, and pricing information from low to high. When multiple candidate objects have the same semantic matching value, load utilization rate is compared first. When load utilization rate is the same, quality correlation value is compared first. When quality correlation value is the same, response latency is compared first. When response latency is the same, pricing information is compared first. When pricing information is the same, the candidate object that was registered earlier is ranked first.

[0111] Read the candidate object with the highest ranking position and whose candidate evaluation result meets the acceptance condition from the candidate sort, and determine the candidate object as the acceptance object; read the task content corresponding to the task information to be assigned, and establish a correspondence between the work object information of the acceptance object and the task content; record the acceptance object, task content, candidate evaluation result and assignment time in the correspondence, and generate the assignment result.

[0112] S4: Send the task content to the recipient according to the assignment result, collect the execution status of the task content, and determine whether the AI ​​agent meets the request conditions. When the AI ​​agent meets the request conditions, read the task content and execution status, extract the request items from the execution status, and write the task content, execution status and request items according to the preset request record format to generate a request record. Update the assignment result according to the request record.

[0113] S4.1: Read the receiving object and task content from the dispatch results, and verify the correspondence between the receiving object and the task content; when the receiving object comes from the AI ​​agent access information, determine the interface address based on the AI ​​agent identifier, and send the task content to the interface address; when the receiving object comes from the operator access information, confirm that the operator is in a receiving state based on the operator identifier and login credentials, and send the task content to the operator; after receiving the task content, the receiving object returns the processing progress, exception flag, completion flag, and intermediate result. The processing progress is used to indicate the current processing stage of the task content, the exception flag is used to indicate whether there is a processing failure, insufficient permissions, tool unavailability, or timeout during the processing of the task content, the completion flag is used to indicate whether the task content has been processed and completed, and the intermediate result is used to save the content that has been generated during the processing of the task content but has not yet been submitted as the final execution result.

[0114] S4.2: Bind the processing progress, exception marker, completion marker, and intermediate results according to the task content identifier, the receiving object identifier, and the return time, and save them as the execution status corresponding to the task content; when the receiving object is an AI agent, read the processing progress, exception marker, completion marker, and intermediate results in the execution status, compare the processing progress with the time limit requirement corresponding to the task content, compare the exception marker with processing failure, insufficient permissions, and tool unavailability, and compare the completion marker and intermediate results with the output completeness requirement; when the processing progress exceeds the time limit requirement and the completion marker does not indicate completion, it is determined that the task content has timed out and has not been completed; when the exception marker includes at least one of processing failure, insufficient permissions, and tool unavailability, it is determined that the task content has a corresponding exception; when the completion marker indicates completion but the intermediate results lack the output content required by the task content, it is determined that the task content has incomplete output; when the execution status indicates that the task content has at least one of processing failure, incomplete output, insufficient permissions, tool unavailability, and timed out and has not been completed, it is determined that the AI ​​agent meets the request for help conditions.

[0115] S4.3: When the AI ​​agent meets the request conditions, it reads the task content according to the task content identifier recorded in the assignment result, and simultaneously reads the execution status corresponding to the task content; in the execution status, it searches for exception markers, completion markers, processing progress, and intermediate results. When the exception marker includes processing failure, a processing failure marker is generated; when the completion marker indicates completion but the intermediate result lacks the necessary content corresponding to the task text, task attachments, task objectives, and output requirements, an output incomplete marker is generated; when the exception marker includes insufficient permissions, an insufficient permissions marker is generated; when the exception marker includes tool unavailable, a tool unavailable marker is generated; when the processing progress exceeds the time limit and the completion marker does not indicate completion, a timeout incomplete marker is generated; the processing failure marker, output incomplete marker, insufficient permissions marker, tool unavailable marker, and timeout incomplete marker are bound to the corresponding task fields in the task content and the completed content in the intermediate results to form a request for help. This allows the execution exception of the AI ​​agent to be broken down into specific request reasons that can be reassigned, preventing the task content from being directly terminated due to the AI ​​agent's processing failure, and enabling subsequent request records to carry the task content, execution status, and request for help backflow to update the assignment result.

[0116] S4.4: Write the task text, task attachments, task objectives, task domain, output requirements, time limits, and permission requirements from the task content into the task field, write the processing progress, exception flag, completion flag, and intermediate results from the execution status into the status field, write the processing failure flag, incomplete output flag, insufficient permissions flag, tool unavailable flag, and timeout flag from the help request into the help request field, and establish a correspondence between the help request field and the corresponding task field and the completed content in the intermediate results to generate a help request record.

[0117] When the request field contains a processing failure flag, determine the processing action corresponding to the stoppage of the task content from the processing progress, read the processing failure reason from the exception flag, and read the capability requirements from the task profile; compare the capability requirements with the capability nodes corresponding to the current receiving object's work profile item by item, and identify the capability content in the capability requirements that is not covered by capability nodes and is consistent with the processing action corresponding to the stoppage of processing as capability gaps; when all capability requirements are covered by the current receiving object's work profile, do not forcibly generate capability gaps, and retain the processing failure reason in the request item.

[0118] When a request field contains an incomplete output flag, the consistency of the result format in the output requirements is compared with the return content type of the intermediate results. The missing return content type is identified as missing result format content. The format requirements in the output requirements are broken down into required fields, required sections, and order, and compared with the intermediate results item by item. The missing required fields, required sections, and order are identified as missing format content. The quality requirements in the output requirements are compared with the corresponding verification results of the intermediate results. Content that does not meet the quality requirements is identified as missing quality content. The missing result format content, missing format content, and missing quality content are grouped into missing output content, and the capability content required to process the missing output content but not covered by the capability node of the current receiving object is identified as capability gap.

[0119] When the request field contains a tool unavailable flag, the tool scope is read from the task profile. The tool nodes corresponding to the tool scope are compared with the tool nodes corresponding to the current receiving object's working profile. Tool nodes required by the task profile but not currently available to the receiving object, or those marked as uncallable, are identified as tool gaps. When the request field contains a permission insufficient flag, the access objects, access scopes, and operation boundaries in the permission requirements are compared item by item with the permission nodes corresponding to the current receiving object's working profile. Access objects, access scopes, and operation boundaries not covered by permission nodes are identified as permission restrictions. When the request field contains a timeout flag, the remaining completion time, response time limit, and stage time limit are read from the time limit requirements and combined with the current processing progress to form status restrictions. Status restrictions include acceptable status requirements, remaining completion time requirements, and response latency requirements. Based on the capability gaps, tool gaps, status restrictions, and permission restrictions, a request profile is formed.

[0120] S4.5: Re-match the capability gaps, tool gaps, permission restrictions, and status restrictions in the request profile with the work profile; read the capability node name, task domain, and missing output content from the capability gap, and concatenate them into capability gap text according to the field order of capability node name, task domain, and missing output content; input the capability gap text into a semantic encoding model for vectorization encoding to obtain the capability gap vector; calculate the cosine similarity between the capability gap vector and the capability vector in the work profile to obtain the capability matching result; compare the consistency of the tool node corresponding to the tool gap with the tool node corresponding to the work profile to obtain the tool completion result; compare the access scope and operation boundary of the permission node corresponding to the permission restriction with the permission node corresponding to the work profile to obtain the permission satisfaction result; compare the acceptable status requirements in the status restrictions with the status vector in the work profile; compare the remaining completion time requirements and response latency requirements with the available time period and response latency in the work object information to obtain the status acceptance result; read the load value table corresponding to the work profile, and compare the number of tasks to be processed, the number of tasks in processing, and the load occupancy with the preset load conditions to obtain the load satisfaction result.

[0121] The system reads the AI ​​entity identifier that triggered the request from the request record and excludes the corresponding work object information from the update candidates. If the work object information has already formed any of the following flags for the same task content and the same request, such as processing failure, incomplete output, insufficient permissions, unavailable tools, or timeout, the work object information is excluded from the update candidates. When the capability matching result reaches the capability coverage threshold, the tool completion result is consistent, the permission satisfaction result is satisfied, and the status acceptance result is acceptable, the work object information corresponding to the work profile is determined as the update candidate.

[0122] The update assignment count is set in the request record. When the request is reassigned for the first time based on the request record, the update assignment count is recorded as 1. Each time an update assignment result is generated again, the update assignment count is incremented by 1. When there are no update candidate objects, candidates with sufficient permissions, an acceptable status, and an operator status are selected from the work object information. The update recipient is determined in the order of load utilization from low to high, quality correlation value from high to low, and response latency from low to high. If there are no operators in an acceptable status, the request record is retained, and the matching is re-executed after the operator status is updated to an acceptable status. When the update assignment count reaches 3 times, the update recipient is no longer determined from the AI ​​agents. Instead, the update recipient is directly determined from the operators with sufficient permissions and an acceptable status to avoid continuous cyclical assignment of task content among AI agents.

[0123] The update candidate objects are recalculated based on their capability coverage, load status, and historical quality. They are then ranked according to semantic matching value (high to low), load occupancy (low to high), quality correlation value (high to low), response latency (low to high), and pricing information (low to high) to form an update candidate ranking. The update candidate object with the highest ranking that meets the acceptance criteria is selected as the update acceptance object. The update acceptance object, task content, request record, request profile, and update candidate ranking are then written into the update assignment result. Through rematching the request profile and the work profile, execution anomalies of the AI ​​agent can be transformed from failure results into a basis for reassignment, allowing the task content to continue flowing even in cases of processing failure, incomplete output, insufficient permissions, tool unavailability, and timeout.

[0124] S5: Determine the effective assignment result based on the assignment result and the updated assignment result, send the task content to the effective recipient based on the effective assignment result, receive the execution result returned by the effective recipient, verify the execution status and execution result, generate a completion record, and use the completion record to update the working profile and semantic network.

[0125] S5.1: Read the update recipient and task content from the update assignment result, and verify the correspondence between the update recipient, the request record, and the task content; determine the receiving method of the update recipient based on the update assignment result, and send the task content and the request items in the request record to the update recipient, so that the update recipient can continue to process the unfinished task content of the AI ​​agent; after the update recipient completes the task content, return the execution result, and retain the processing content corresponding to the task content, request items, and completed content in the execution result.

[0126] S5.2: Read the execution status corresponding to the task content, and extract the processing progress, exception marker, and completion marker from the execution status; extract the request item from the request record; compare the processing progress with the processing stage in the execution result. When the processing progress indicates that the task content has entered the completion stage and the execution result has been returned, record the progress as consistent; compare the exception marker with the processing description in the execution result. When the exception marker does not indicate a new exception, and the execution result has already processed the corresponding content such as processing failure, incomplete output, insufficient permissions, tool unavailable, and timeout failure, record the exception as consistent; compare the completion marker with the processing description in the execution result. The submission status is compared. When the completion mark indicates that the task content is completed and the execution result contains the output content required by the task content, the completion is recorded as consistent. The request for help is compared with the supplementary processing content in the execution result. When the execution result covers the processing failure reason, missing output content, permission processing content, tool processing content, and timeout remediation content corresponding to the request for help, the request for help is recorded as consistent. When progress consistency, exception consistency, completion consistency, and request for help consistency are all true, a verification result that passes the verification is generated. When any of the progress consistency, exception consistency, completion consistency, and request for help consistency is false, a verification result that fails the verification is generated, and the inconsistency content is recorded in the verification result.

[0127] S5.3: When the verification result is "verification passed", the completion mark indicates that the task content is completed, the exception mark does not indicate a new exception, and the execution result meets the output requirements in the task content, the verification result is determined to meet the preset completion conditions; the processing progress, exception mark and completion mark in the execution status, the request for help in the request record, and the processing content, submission content and completion time in the execution result are written into the preset completion record format in the order of the fields of task content identifier, recipient, execution status, execution result and verification result to generate a completion record.

[0128] Extract completion time, exception marker, request for help, and execution result from the completion record; read the task content sending time, initial processing progress return time, and the load node corresponding to the receiving object; remove the processing task record corresponding to the task content identifier from the load node, decrement the number of processing tasks by 1, and record 0 if the number of processing tasks is less than 0; recalculate the load occupancy based on the updated number of pending tasks and processing tasks, where the load occupancy is the sum of the number of pending tasks and processing tasks divided by the preset maximum number of tasks to be received; determine the time difference between the initial processing progress return time and the task content sending time as the current response latency; weight the historical response latency and the current response latency with a 4:1 ratio to obtain the updated response latency, and write the updated response latency into the work object information.

[0129] The status vector is regenerated based on the updated online status, acceptable status, and paused status. The feedback node is updated based on the quality score, anomaly marker, help requests, and execution results in the completion record, and the feedback vector is regenerated. The regenerated status vector and feedback vector are associated with the work object information to complete the work profile update. When the completion record does not represent changes in declared skills, access permissions, and tool access scope, the original capability vector is retained to avoid changing declared capabilities, permissions, and tool boundaries based solely on the results of a single task.

[0130] Write the execution result, quality score, anomaly marker, and request for help from the completion record into the feedback node in the semantic network, and increment the call count corresponding to the work object information by 1; normalize the capability coverage result, quality score, and call frequency to the range of 0 to 1, and average them with equal weights to obtain the update relationship weight; write the update relationship weight into the semantic relationship between the work object information and the corresponding capability node, tool node, and feedback node, and write the completion record generation time into the update time in the semantic relationship to complete the semantic network update.

[0131] To further clarify, the completion record format includes task content identifier, recipient, processing progress, exception marker, completion marker, request for help, execution result, verification result, task content sending time, first processing progress return time, completion time, response latency, and completion time. The maximum number of tasks that can be accepted is set based on the current load, historical call records, historical completion quality, and response latency in the historical work object information. The maximum number of tasks that can be accepted simultaneously for the same work object information, provided that the historical completion quality is satisfactory and the response latency does not exceed the time limit requirement, is used as the basis for setting the maximum number of tasks to be accepted and the number of tasks in process. When the work object information is an AI agent, the maximum number of tasks that can be accepted is set to 10; when the work object information is an operator, the maximum number of tasks that can be accepted is set to 6.

[0132] S5.4: When the AI ​​agent does not meet the request conditions, and the execution status does not indicate that the task content has failed, the output is incomplete, the permissions are insufficient, the tool is unavailable, or the timeout has not been completed, the assignment result remains valid, no request record is generated, and no updated assignment result is generated; the assignment result is determined as the effective assignment result, and the receiving object in the assignment result is determined as the effective receiving object, and the effective receiving object continues to process the task content; after the effective receiving object completes the task content, the execution result returned by the effective receiving object is received, and the execution status corresponding to the task content is read. The processing progress, exception marker, completion marker, and intermediate results in the execution status are checked for consistency with the execution result; when the execution status indicates that the task content has been completed, the exception marker does not indicate an exception, the intermediate results are consistent with the execution result, and the execution result meets the output requirements in the task content, the execution status and execution result are written into the completion record format, a completion record is generated, and the working profile and semantic network are updated using the completion record.

[0133] In summary, this invention improves task continuity by updating assignment results through request records. This prevents task interruptions caused by AI agents failing to process tasks, providing incomplete output, lacking sufficient permissions, having unavailable tools, or failing to complete tasks within timeouts. Instead, it transforms the cause of the anomaly into a basis for subsequent reassignment. Furthermore, the request records preserve the correspondence between task content, execution status, and requests for assistance, making the cause of anomalies clearer and avoiding the rough assignment issues caused by consistently transferring tasks to human agents. Finally, by re-determining the recipient based on the request records, the task content can be transferred to a more suitable AI agent or operator, improving the efficiency of remediating abnormal tasks, the accuracy of assignment, and the stability of human-machine collaboration.

[0134] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A human-computer collaborative task assignment method based on semantic vector profiling, characterized in that: include, Acquire AI agent access information and operator access information, register them as work object information, and construct a semantic network based on the work object information; Semantic encoding of work object information is performed based on semantic networks to generate work profiles, and information on tasks to be assigned is obtained. Semantic parsing of the information on tasks to be assigned is performed based on semantic networks to generate task profiles. Semantic matching of task profiles and work profiles is performed to obtain candidate objects. The candidate object ranking is formed by the capability coverage results, load status and historical quality of the candidate objects. The receiving object is determined from the candidate ranking and the assignment result is generated. Based on the assignment results, the task content is sent to the receiving object, the execution status of the task content is collected, and it is determined whether the AI ​​agent meets the request conditions. When the AI ​​agent meets the request conditions, the task content and execution status are read, the request items are extracted from the execution status, and the task content, execution status and request items are written according to the preset request record format to generate a request record. The assignment results are updated based on the request record. Based on the assignment results and the updated assignment results, the effective assignment results are determined. Based on the effective assignment results, the task content is sent to the effective recipients. The execution results returned by the effective recipients are received, and the execution status and execution results are verified. A completion record is generated, and the work profile and semantic network are updated using the completion record.

2. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 1, characterized in that: The AI ​​agent access information includes AI agent identifier, interface address, calling permissions, declared skills, tool calling scope, historical calling records, availability status, response latency, quotation information, failure records, and help records; The operator access information includes operator ID, login credentials, skill tags, permission tags, available time periods, calibration task results, historical completion quality, current load, response latency, quotation information, failure records, and help records.

3. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 2, characterized in that: The construction of a semantic network based on work object information specifically includes: Read the declared skills, skill tags, access permissions, access permission tags, tool access scope, availability status, historical access records, calibration task results, historical completion quality, current load, failure records, and help records from the work object information; The declared skills and skill tags are grouped into ability tags, the call permissions and permission tags are grouped into permission tags, the tool call scope is grouped into tool tags, and the historical call records, calibration task results, historical completion quality, failure records and help records are grouped into feedback tags; The capability label, permission label, tool label, availability status, current load, and feedback label are mapped to capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes in the semantic network, respectively. Establish semantic relationships between work object information and capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes to form a semantic network.

4. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 3, characterized in that: The generation of the working profile specifically involves: Read the capability nodes, permission nodes, tool nodes, status nodes, load nodes, and feedback nodes corresponding to the working object information in the semantic network; Vectorize the capability nodes, permission nodes, and tool nodes to obtain the capability vector; The state nodes are encoded to obtain state vectors. The number of pending tasks, the number of tasks in processing, and the load occupancy in the load nodes are numerically organized to obtain a load value table. The feedback nodes are quality encoded to obtain the feedback vector; By associating capability vectors, state vectors, and feedback vectors with work object information, a work profile is generated.

5. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 4, characterized in that: The generated task profile is specifically as follows: Read the task text, task attachments, task objectives, task domain, output requirements, time limits, and permission requirements from the information of the task to be assigned; The task text, task attachments, task objectives, and task domain are mapped to capability nodes and tool nodes in the semantic network to obtain the task semantics and capability requirements. The output requirements are organized into fields to obtain result status restrictions, the time limit requirements are organized into time constraints to obtain time continuity restrictions, the permission requirements are mapped to permission nodes in the semantic network to obtain permission access restrictions, and the result status restrictions, time continuity restrictions, and permission access restrictions form task constraints. The task semantics are vectorized to obtain the task vector, and the task constraints are constrained to obtain the constraint vector. The task vector, capability requirements, and constraint vector are associated with the task information to be assigned to generate a task profile.

6. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 5, characterized in that: The process of obtaining candidate objects specifically involves: Read the task vector and constraint vector from the task profile, and read the capability vector, state vector and feedback vector from the job profile; Semantic similarity matching is performed between the task vector and the capability vector to obtain the capability coverage result; Read the time acceptance limit and access restriction from the task constraints corresponding to the constraint vector, compare the time acceptance limit with the state vector, available time period and response latency for time adaptation, compare the access restriction with the permission node for field comparison, and compare the load value table with the preset load conditions to obtain the load status. The historical quality is obtained by comparing the feedback vector with the task domain in the task profile and performing a quality correlation comparison. Based on the capacity coverage results, load status, and historical quality, candidate objects are obtained by filtering the work profiles.

7. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 6, characterized in that: The specific details of generating the assignment result are as follows: Read the capability coverage results, load status, and historical quality of the candidate object; The capability coverage results are compared with the capability requirements in the task profile to obtain the acceptance matching mark. The load status is compared with the preset load conditions to obtain the load occupancy mark. The historical quality is compared with the task domain in the task profile to obtain the quality stability mark. The matching markers, load occupancy markers, and quality stability markers are collected and arranged according to a preset evaluation order to obtain candidate evaluation results; The candidates are ranked according to the evaluation results to form a candidate ranking. Read the candidate objects that meet the acceptance conditions from the candidate sort, determine them as acceptance objects, establish the correspondence between acceptance objects and task content, and generate the assignment results.

8. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 7, characterized in that: The execution status of the data collection task is monitored, and it is determined whether the AI ​​agent meets the request for assistance conditions, specifically as follows: Read the receiving object and task content from the assignment result; Send the task content to the recipient and receive the processing progress, exception flags, completion flags and intermediate results returned by the recipient; Save the processing progress, exception flags, completion flags, and intermediate results as the execution status; When the target is an AI entity, the execution status will be compared with the preset request conditions; When the execution status characterizes any of the following states: processing failure, incomplete output, insufficient permissions, unavailable tools, or timeout, the AI ​​agent is deemed to meet the request for assistance.

9. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 8, characterized in that: The process of updating the assignment results based on the request record is as follows: When the AI ​​agent meets the request conditions, read the task content and execution status; Extract the processing failure flag, incomplete output flag, insufficient permissions flag, tool unavailable flag, and timeout flag from the execution status to form a help request; Organize the fields for task content, execution status, and requests for assistance to generate a request record; Mapping the requests for help in the request records to a semantic network yields a request profile. Semantic matching is performed between the request-for-help profile and the work profile to obtain updated candidate objects; Based on the updated candidate objects, a new matching is performed to form an updated candidate ranking, and the updated recipient objects are determined from the updated candidate ranking to generate the updated assignment results.

10. The human-computer collaborative task assignment method based on semantic vector profiling as described in claim 9, characterized in that: The generated completion record is specifically as follows: Read the update recipient and task content from the update assignment results; Based on the updated assignment results, the task content is sent to the newly determined recipient, and the execution results are received from the newly determined recipient. Read the execution status and help records corresponding to the task content, extract the processing progress, exception markers and completion markers from the execution status, and extract the help requests from the help records; The consistency of processing progress, exception markers, completion markers, and requests for help is verified with the execution results to obtain the verification results. When the verification result meets the preset completion conditions, the execution status and execution result are written into the preset completion record format to generate a completion record.

Images