Digital employee collaboration screening method and system

Abstract

This invention discloses a digital employee collaborative screening method and system, belonging to the field of artificial intelligence technology. The method includes: receiving concurrent requests from multiple users, performing semantic isolation, generating adjustment factors based on user profile features, generating initial semantic vectors through dual-channel semantic encoding, and performing semantic conflict detection to effectively avoid semantic cross-confusion and achieve intent parsing of concurrent user requests; real-time collection of request features, calculation of the computing power requirement levels of different requests, sorting requests with the same computing power requirement level, and allocating resources; obtaining user adjustment instructions through the front end, identifying requirement differences, generating incremental calculation schemes, and performing state synchronization control; dividing digital employees into three types of intelligent agents—analytical, computational, and feedback—for collaborative screening and feedback verification to ensure that the screening results better match user needs and achieve real-time capability iteration of the digital employee cluster.

Description

Technical Field

[0001] This invention relates to a digital employee collaboration screening method and system, belonging to the field of artificial intelligence technology. Background Technology

[0002] In the digital age, information filtering, as a core component of enterprise decision-making, content management, and resource allocation, has undergone a multi-stage technological evolution from being human-led to being driven by intelligence. Early information filtering relied on manual processes, with professionals reviewing and marking key information item by item. While this allowed for the assessment of tacit knowledge through human experience, it had limitations, such as efficiency bottlenecks, lack of consistency, and scalability constraints. With the popularization of information technology, automated filtering tools based on rule engines and keyword matching have become mainstream, enabling batch filtering of candidate information through predefined filtering rules or keyword dictionaries. In recent years, large language model technologies, represented by Transformer, have broken through the limitations of traditional rules, achieving semantic-level understanding and multimodal processing.

[0003] However, many challenges remain in practical applications, with adaptability issues being particularly prominent in scenarios with multiple concurrent user requests. Because different users have significantly different filtering needs, a large number of differentiated requests flood in simultaneously. If these differentiated requests contain semantically overlapping or even ambiguous filtering logic—for example, similar expressions corresponding to drastically different filtering targets—rule confusion can easily occur during parallel parsing, leading to misjudgments or omissions in filtering conditions. Furthermore, the computational complexity of differentiated requests varies significantly; simple keyword matching and cross-document semantic association analysis require several orders of magnitude more computing power, making resource scheduling mechanisms difficult to implement. Precise balance is needed. The high response speed requirements of lightweight requests conflict with the reliance on continuous computing power for heavyweight requests. This often leads to unnecessary delays for lightweight requests due to resource contention, or long waiting times for heavyweight requests due to insufficient computing power. More complicatedly, if users dynamically adjust their requirements during the filtering process, such as temporarily modifying filtering dimensions or precision, the resources required for dynamic adjustments may be occupied if other differentiated requests are processed simultaneously. This not only causes delays in the response to adjustment commands, but may also lead to a disconnect between the filtering process and the user's latest needs due to the disorder of the multi-request state synchronization mechanism, ultimately affecting the accuracy and timeliness of the filtering results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a digital employee collaboration screening method and system, which solves the adaptability problem in multi-user concurrent scenarios through a multi-role digital employee collaboration mechanism and dynamic resource scheduling strategy.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Digital employee collaboration screening methods include:

[0007] It receives concurrent requests from multiple users, performs semantic isolation, generates adjustment factors based on user profile features, generates an initial semantic vector through dual-channel semantic encoding, and performs semantic conflict detection.

[0008] Collect request characteristics in real time, calculate the computing power requirement level of different requests, and allocate resources accordingly;

[0009] By obtaining user adjustment instructions through the front end, identifying differences in requirements, generating incremental calculation plans, and performing state synchronization control;

[0010] Divide digital employees into categories, conduct collaboration screening, and perform feedback verification.

[0011] Specifically, the steps of semantic isolation include:

[0012] Assign a unique ID to each request, build a context cache, and synchronously adopt the LRU eviction policy;

[0013] The user profiling API is called to obtain a 3D feature vector, and the semantic adjustment coefficient is obtained by weighted summation. The adjustment factor is then calculated.

[0014] Construct a basic text vector, multiply it by the adjustment factor, and generate a weighted text vector;

[0015] Based on preset filtering rules, the weighted text vector is decomposed, and semantic dimensions that do not conform to the filtering rules are removed;

[0016] The filtered weighted text vectors are processed by the softmax function to generate an attention mask, and the decision pattern features of the context cache are extracted to generate a dedicated stop word list.

[0017] Dual-channel semantic encoding is initiated to generate surface vectors and domain entity vectors. The surface vectors and domain entity vectors are then fused through residual connections to generate an initial semantic vector.

[0018] Specifically, the steps of semantic conflict detection include:

[0019] For the initial semantic vectors of concurrent requests, calculate the semantic similarity between any two requests;

[0020] If the semantic similarity is greater than the semantic similarity threshold and the two requests are from different users, they are determined to be conflicting requests and included in the conflicting request list; otherwise, they are determined to be non-conflicting requests and the three-level intent calibration is performed directly.

[0021] For conflict requests, perform logical structure encoding, calculate the nesting depth of modifiers, extract and combine numerical range boundary features to generate structural feature fingerprints;

[0022] The decision tree generation algorithm is invoked, with the logical operator sequence as the root node, the nesting depth of modifiers as the intermediate nodes, and the user's historical preferences as the leaf nodes, to generate the filtering logic.

[0023] Perform three-level intent calibration to generate a three-dimensional confidence score, including semantic matching degree, domain consistency, and historical fit degree, and obtain the comprehensive confidence score through weighted calculation.

[0024] Specifically, the steps for resource allocation include:

[0025] The semantic depth level is divided according to the breadth of cross-domain entity associations in the intent vector and the degree of historical fit in the three-dimensional confidence score. A document association graph is generated based on entity link information and the association range level is divided according to the number of nodes. The computing power requirement level is generated by weighted calculation.

[0026] Extract behavioral features from real-time user operation logs and quantify them into sensitivity. Then, through weighted calculation, obtain the sensitivity of user operation behavior. Combine the decision speed preference features in the user profile to calculate the urgency of each request.

[0027] Based on the urgency and the overall confidence level, the priority of each request is obtained through weighted calculation. Requests with the same computing power requirement level are then sorted a second time based on the priority to generate a request queue.

[0028] Specifically, the steps for resource allocation include:

[0029] Real-time acquisition of multi-dimensional features, construction of a multi-dimensional feature state space, and design of a dual-objective reward function;

[0030] The system is trained using reinforcement learning algorithms to adjust resource allocation under different states and generate real-time scheduling strategies.

[0031] Initial resource allocation is performed according to the real-time scheduling strategy, matching corresponding computing nodes to requests with different computing power requirements, and allocating corresponding computing power proportions according to the priority of the requests.

[0032] The system acquires user interaction behavior with the filtering interface, generates a real-time behavior score, generates a historical habit score based on the user's past similar filtering operation records, and performs weighted merging to obtain the adjustment probability.

[0033] When the adjustment probability is greater than the adjustment threshold, the resource adjustment coefficient is calculated, and the reserved resources are calculated in combination with the basic coefficient and the computing power standard, and temporarily allocated from the current user's non-urgent requests.

[0034] When fluctuations in resource load are detected, resource allocation is dynamically adjusted.

[0035] Specifically, the steps to identify differences in demand include:

[0036] Acquire user interaction behavior on the filtering interface, capture dynamic adjustment commands, and classify user adjustment commands using a command classification model;

[0037] The number of new screening dimensions added in this adjustment is counted, the magnitude of the change in precision and the relevance of the target are calculated, and the adjustment impact factor is generated by weighted fusion.

[0038] Incremental computing schemes are generated for different types of adjustment instructions, including: generating a semantic correlation matrix based on dual-channel semantic encoding, calculating incremental coefficients and incremental computing power, and scheduling idle nodes in the reserved resource pool; differentiating precision modification instructions according to computing power requirement levels; re-analyzing new targets for target replacement instructions, reusing the basic resources of the original computing nodes, and allocating idle computing power.

[0039] The system acquires the priority of adjustment instructions and uses a token bucket mechanism to control the access frequency of concurrent adjustments, allowing only instructions holding tokens to enter the resource scheduling queue and generating resource preemption results.

[0040] Specifically, the steps for state synchronization include:

[0041] Before and after the adjustment, state snapshots are generated respectively. These state snapshots are backed up through a distributed storage system, associated with user ID and request ID, and synchronized to all associated nodes through a distributed protocol.

[0042] For the intermediate results of incremental calculation, extract the hash values ​​of the results before and after adjustment, and obtain the hash difference rate by subtracting the ratio of the number of identical hash values ​​to the total number of hash values ​​from the base coefficient. Set the difference threshold to perform difference hash comparison.

[0043] If the hash difference rate is less than the difference threshold, the incremental results are directly merged; otherwise, the subset of data points corresponding to the differences is located and a local recalculation is triggered.

[0044] The adjustment results are fed back and the model is iterated. User satisfaction with the adjustment results is collected. The adjustment effect is evaluated by combining the compatibility between the adjustment results and the original screening logic. The parameters of the intent model and scheduling model are updated through online learning.

[0045] Specifically, the collaborative screening process includes the following steps:

[0046] Based on the complexity of the intent vector and the level of computing power requirements, the agents are assigned tasks, the matching degree of the agents is calculated to allocate tasks, and the agent cooperation topology is generated.

[0047] The screening task is broken down into three levels, and the priority of the subtasks is calculated to generate a list of subtasks.

[0048] Each agent processes subtasks in parallel. When the outputs of two computational agents conflict, the weight of each conflicting computational agent's result is calculated, and the result of the computational agent with the highest weight is taken as the valid result.

[0049] Perform multi-level result aggregation to generate multi-level aggregated results;

[0050] The task completion rate is calculated based on the task completion data. At the same time, the performance score of the agent is obtained through weighted calculation, and a performance threshold is set.

[0051] When the agent's performance score is less than the performance threshold, parameter iteration is triggered.

[0052] Specifically, the feedback verification steps include:

[0053] Calculate feedback similarity, adjustment consistency, and matching ratio, and filter out semantic deviation, adjustment failure, and historical deviation;

[0054] The comprehensive verification score is generated by weighted calculation based on the combined results of the three levels of verification. If the comprehensive verification score is lower than the first verification threshold, the result is determined to be abnormal and the abnormality type is marked.

[0055] A second verification threshold is set. For abnormal results where the comprehensive verification score is lower than the second verification threshold, a deep feedback mode is adopted; otherwise, a lightweight feedback mode is adopted.

[0056] Based on the comprehensive verification score and feedback data, a three-level optimization is triggered, and all verification data is aggregated to generate a long-term performance score. If the long-term performance score is lower than the baseline threshold for three consecutive months, an architecture-level optimization is triggered.

[0057] The digital employee collaboration screening system includes: a parsing module, an allocation module, a dynamic adjustment module, and a feedback module;

[0058] The parsing module is used to isolate contextual data from different users and identify conflicting requests through semantic conflict detection.

[0059] The allocation module is used to collect request characteristics in real time to calculate the computing power demand level, combine the user's real-time operation behavior and decision speed preference to calculate the urgency and priority, and construct a real-time scheduling strategy.

[0060] The dynamic adjustment module is used to acquire and classify user adjustment instructions, generate incremental calculation schemes, and allocate resources to the adjustment instructions through priority scheduling.

[0061] The feedback module is used to assign roles to multiple agents and generate collaborative topologies, decompose subtasks and bind priorities, generate multi-level aggregated results through cross-agent collaborative computing and conflict arbitration, and perform multi-dimensional verification on the results.

[0062] The beneficial effects of this invention are:

[0063] Multi-user semantic isolation is achieved through an independent context cache, combined with dual-channel semantic encoding to integrate grammar and domain features. Semantic conflict detection and unique interpretation rule generation effectively avoid semantic cross-confusion between different user requests, improving intent parsing accuracy and resolving semantic interference issues under multi-user concurrency. Incremental computation reduces redundant computing power consumption, while priority scheduling and resource preemption mechanisms improve adjustment response speed. State snapshot synchronization and hash difference verification ensure consistency of results across multiple nodes, resolving slow adjustment response and chaotic state synchronization. Dynamic role allocation and conflict arbitration mechanisms ensure consistency in task decomposition and computation. A feedback verification system continuously optimizes model parameters and collaboration strategies through multi-dimensional verification and closed-loop iteration, improving the match between results and user needs, resolving inconsistencies in collaboration and insufficient result accuracy, and continuously correcting system deviations to further enhance the match between results and user needs. Attached Figure Description

[0064] Figure 1 Flowchart for screening methods for digital employee collaboration;

[0065] Figure 2 This is a flowchart of semantic conflict detection in this invention;

[0066] Figure 3 This is a flowchart of resource allocation in this invention;

[0067] Figure 4 Structure diagram of a digital employee collaboration screening system. Detailed Implementation

[0068] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0069] Example 1:

[0070] refer to Figures 1 to 3 As shown in the figure, this embodiment introduces a digital employee collaboration screening method, including the following steps:

[0071] Step S1: Receive concurrent requests from multiple users, allocate an independent context buffer for each request, perform semantic isolation, generate adjustment factors based on user profile features, filter invalid dimensions after weighted processing of the request text vector, generate an initial semantic vector by fusing syntactic and domain features through dual-channel semantic encoding, perform semantic conflict detection to identify conflicting requests, generate structural feature fingerprints and generate unique interpretation rules, generate intent vectors and three-dimensional confidence scores through three-level calibration, effectively avoid semantic cross-confusion, improve parsing accuracy and personalized adaptation, and achieve accurate intent parsing of concurrent requests from multiple users;

[0072] Step S2: Real-time collection of request characteristics through the computing power monitoring nodes of the digital employee cluster, calculation of computing power demand levels for different requests, generation of urgency and priority based on real-time user operation behavior, sorting requests with the same computing power demand level to avoid delays caused by resource contention for lightweight requests, resource allocation based on reinforcement learning training of resource scheduling strategy with response latency and resource utilization as dual reward objectives, introduction of a dynamic adjustment mechanism for request priority, prediction of user adjustment intentions and reservation of resources, dynamic load balancing, improvement of resource allocation accuracy and flexibility, achieving a balance between response efficiency and utilization efficiency, and adapting to dynamic adjustment needs;

[0073] Step S3: Obtain user adjustment instructions through the front end, identify demand differences, calculate adjustment impact factors, quantify complexity, generate incremental calculation schemes for different adjustment types to reduce resource consumption, determine the priority of adjustment instructions by combining urgency and adjustment impact factors, ensure rapid adjustment response through resource scheduling, downgrade existing requested resources when necessary, and perform state synchronization control. Generate state snapshots before and after adjustment and synchronize them to related nodes, use hash difference verification to ensure consistency of results across multiple nodes, trigger local recalculation when differences are large, collect user feedback, inject adjustment data into model iteration and optimization, and improve the accuracy of subsequent adjustment prediction. The overall process effectively ensures the accuracy of demand difference identification and the consistency of state synchronization.

[0074] Step S4: Divide digital employees into three types of intelligent agents: analytical, computational, and feedback agents, conduct collaborative screening, allocate tasks based on historical performance and load adaptability to generate collaborative topologies, decompose tasks and bind priorities, ensure consistency of results through cross-agent collaborative computing and conflict arbitration, and dynamically migrate tasks to balance computing power when the load suddenly increases.

[0075] Step S5: Before outputting the filtering results, feedback verification is performed on the results generated by the multi-agent collaboration. The matching degree between the results and the user's initial needs, adjustment instructions and long-term preferences is evaluated through multi-dimensional verification. Anomalies are marked, and a hierarchical feedback mechanism is designed for anomalies. Feedback is collected and its effectiveness is quantified in combination with user profiles. Based on this, three-level optimization is triggered. Through long-term performance evaluation and architecture optimization, the accuracy and adaptability of the system are continuously improved to ensure that the filtering results are more in line with user needs and realize the real-time capability iteration of the digital employee cluster.

[0076] Specifically, the steps involved in semantic isolation include:

[0077] To avoid information confusion caused by memory address conflicts, contextual data from different users is physically isolated. When receiving concurrent requests from multiple users, such as user A and user B simultaneously initiating screening requests on the digital employee platform, a distributed lock mechanism assigns a unique ID to each request. A hash value is generated based on the ID, and this hash value is mapped to an independent memory partition as a context cache. The cache size is set to the length of the request text. The cache only stores the text of the current request, the user ID, and the initial parsing status. The parsing status includes unprocessed, processing, and completed. At the same time, an LRU eviction policy is adopted. When the memory usage exceeds the preset ratio, the cache that has not been operated on within the preset eviction interval is released first, so as to achieve physical isolation of data from different user contexts.

[0078] The user profiling API is called to obtain a three-dimensional feature vector containing knowledge density coefficients, domain-specific precision, and decision-making speed preferences. Deviation standardization is then used to compress the features in the three-dimensional feature vector to a smaller size. Within the interval, corresponding weight coefficients are assigned, and the semantic adjustment coefficient is calculated by weighted summation; among them, the knowledge density coefficient reflects the user's familiarity with professional terms, the domain precision reflects the user's focus on their domain, and the decision speed preference characterizes the user's sensitivity to response timeliness.

[0079] Based on the sum of the base coefficient and the semantic adjustment coefficient, the user's adjustment factor is calculated. The request text is converted into a Word2Vec vector and used as the base text vector. The base text vector is multiplied by the adjustment factor, which amplifies the semantic dimension weights that match the user's features and relatively suppresses the semantic dimension weights that do not match, thus obtaining a weighted text vector. In this embodiment, the base coefficient is set to 1 to ensure that even if the feature value in the three-dimensional feature vector is 0, the initial text vector still has a base weight.

[0080] After weighting, some low-value semantics will be amplified. At this point, the weighted text vector needs to be filtered. Based on the task scenario or domain characteristics, preset filtering rules are set, including numerical threshold filtering, domain relevance filtering, and outlier filtering. The weighted semantic vector is decomposed into multiple independent semantic dimensions, and each is compared with the above filtering rules. Semantic dimensions that do not meet the filtering rules are marked as to be filtered and processed. At this time, the filtered weighted semantic vector is generated. Among them, numerical threshold filtering sets a minimum effective threshold for the weight of semantic dimensions. Domain relevance filtering is based on the domain vocabulary or pre-trained domain classification model to identify and filter semantic dimensions that are not related to the current domain. Outlier filtering identifies extreme values ​​through statistical methods, judges them as calculation errors, and filters them. The semantic dimensions to be filtered include weakly relevant dimensions, irrelevant dimensions, and outlier dimensions. If the weight of a semantic dimension is lower than the minimum effective threshold, it is marked as a weakly relevant dimension. If a semantic dimension is judged by the domain classification model as not related to the current domain, it is marked as an irrelevant dimension. If the weight of a dimension exceeds the outlier range, it is marked as an outlier dimension.

[0081] The filtered weighted text vector is processed using the softmax function to normalize the weights of all semantic dimensions. The probability values ​​within the interval are used to generate an attention mask to guide the weight allocation during dual-channel semantic encoding. At the same time, based on the user ID in the context cache, the user's last three filtering operation logs are queried to extract decision pattern features, including high-frequency filtering dimensions, adjustment frequency, and result correction type. The TF-IDF algorithm is used to generate a user-specific stop word list to filter irrelevant entities during dual-channel semantic encoding.

[0082] Dual-channel semantic encoding is initiated. The surface channel uses the BERT-base model to extract syntactic structure features and outputs a surface vector containing logical connectors and modifier positions. The deep channel uses the ERNIE-3.0 model to associate with the enterprise knowledge graph, maps the filtered weighted text vector to entity nodes, and outputs a domain entity vector. The surface vector and the domain entity vector are fused through residual connections to generate the initial semantic vector.

[0083] Specifically, the steps for semantic conflict detection include:

[0084] To extract and identify request groups with confusion and perform semantic conflict detection, the semantic similarity between any two requests is calculated using the cosine similarity formula for the initial semantic vectors of all concurrent requests, and a semantic conflict is determined based on the set semantic similarity threshold.

[0085] If the semantic similarity is greater than the semantic similarity threshold and the two requests are from different users, then there is a semantic conflict between the two concurrent requests, which is determined to be a conflicting request and included in the conflicting request list.

[0086] Otherwise, if there is no semantic conflict between the two concurrent requests, they are determined to be non-conflicting requests, and the Level 3 intent calibration is performed directly.

[0087] For conflict requests, structural feature fingerprints are generated first. The text of the conflict request is logically structured, logical connectors are converted into operator sequences, the nesting depth of modifiers is calculated, the boundary features of the numerical range are extracted, and the features are combined to generate structural feature fingerprints.

[0088] The rule generator is started based on structural feature fingerprints. It calls the decision tree generation algorithm, with the logical operator sequence as the root node, the nesting depth of modifiers as the intermediate node, and the user's historical preferences as the leaf node, to generate a unique rule for conflict requests. After being verified by the rule engine, it is converted into executable filtering logic.

[0089] The process involves three levels of intent calibration. The first level is historical behavior backtracking, which calculates the cosine similarity between the current request and the intent vectors of the user's last three filters. If the similarity is less than the historical threshold, the initial semantic vector is fine-tuned by increasing the weight of the high-frequency filter dimension. The second level is domain knowledge constraint, which calls the industry rule engine to correct the weight of the semantic dimension of the vector. The third level is real-time feedback prediction, which uses an LSTM model to predict the user's expected adjustments. The results of these three levels of intent calibration are then fused to generate the intent vector.

[0090] A three-dimensional confidence score is generated, including semantic matching degree, domain consistency, and historical fit. After weighted calculation, a comprehensive confidence score is obtained, and a scoring report is generated, which includes the intent vector, interpretation rules, confidence score, and parsing results of the three-dimensional features.

[0091] Specifically, the specific steps for resource allocation include:

[0092] The computational power requirement level is determined by quantifying the semantic complexity and relevance of the request. Based on the breadth of cross-domain entity relevance in the intent vector and the degree of historical fit in the three-dimensional confidence score, the semantic depth is divided into different levels. The higher the level, the deeper the semantic parsing required. A document relevance graph is generated based on entity link information, the number of nodes is counted, and the relevance range level is divided according to the number of nodes. The higher the level, the wider the range of information to be relevance. Based on the above quantification rules, the semantic depth level and relevance range level of each request are obtained. Through weighted calculation, a comprehensive computational power requirement level is generated. Different computational power requirement levels correspond to different computational complexities, and the corresponding computational power standards are called to provide a basis for resource allocation.

[0093] Three types of behavioral features are extracted from the user's real-time operation logs: the duration of cursor hovering over the filtering area, the number of refreshes within a sensitive interval, and the character input speed in the input box. These three features are quantified into sensitivity, and the user's operational behavior sensitivity is obtained through weighted calculation. Based on the decision speed preference feature in the user profile, the urgency of each request is calculated. The decision speed preference reflects the user's emphasis on response efficiency and, together with the operational behavior sensitivity, determines the level of urgency. It is obtained by multiplying the difference between the base coefficient and the decision speed preference with the operational behavior sensitivity.

[0094] Based on urgency and overall confidence, the priority of each request is obtained through weighted calculation. For requests with the same computing power requirement level, a secondary sorting is performed based on priority. Requests with higher priority have a greater advantage in resource allocation, thereby generating a sorted request queue.

[0095] A state space containing multi-dimensional features is constructed, and multi-dimensional features are collected in real time, including the proportion of requests of each computing power demand level in the current queue, the real-time CPU and GPU resource load, and the performance indicators of recent lightweight and heavyweight requests. These features together constitute the input state for policy training. With the minimum response latency and optimal resource utilization as the dual objectives, a dual-objective reward function is designed. For the response latency of a request, the closer the latency is to the preset reasonable range, the higher the reward; otherwise, it will be penalized. For the resource utilization, the closer the resource utilization is to the ideal range, the higher the reward; the further it deviates, the heavier the penalty.

[0096] The scheduling strategy is trained by reinforcement learning algorithm, and the resource allocation method is tried to be adjusted under different states. The scheduling strategy is continuously optimized based on the rewards obtained until the scheduling strategy converges, and a real-time scheduling strategy is generated to dynamically adjust the resource allocation logic according to the real-time state.

[0097] Initial resource allocation is performed based on real-time scheduling strategies, matching corresponding computing nodes to requests with different computing power requirements. Lightweight requests are allocated to edge computing nodes, while heavyweight requests are allocated to the central GPU resource pool. The corresponding computing power ratio is also allocated according to the priority of the request, with higher priority requests receiving more resources of the same level.

[0098] By analyzing user behavior in the filter parameter area, we can analyze user intentions, predict user adjustment intentions, and reserve resources to meet dynamic needs. We can acquire user interaction behavior with the filter interface in real time, including the user cursor's stay in the filter parameter area, the frequency of clicking filter conditions, and the hovering time on the adjustment button, and convert this into feature scores. These three types of feature scores are integrated to form a real-time behavior comprehensive score. We can retrieve the user's past filter operation records to obtain the proportion of times the user ultimately performed an adjustment operation when initiating the same type of filter request in the past. The higher the proportion, the more likely the user is to make an adjustment in similar scenarios, and this is converted into a feature score. We can also obtain the user's average number of adjustment operations per month. The more times, the more accustomed the user is to adjust the filter results, and this is converted into a feature score. We can integrate these to generate a historical habit comprehensive score. The real-time behavior comprehensive score and the historical habit comprehensive score are combined according to a specific weighting logic to obtain the final adjustment probability.

[0099] When the adjustment probability is greater than the adjustment threshold, reserve backup resources. The resource adjustment coefficient is obtained by multiplying the monthly average number of adjustments by the reservation coefficient. The reserve resources are calculated by summing the resource adjustment coefficient and the base coefficient, combined with the computing power standard, so as to reserve a certain amount of backup resources in advance. These resources are temporarily allocated from the current user's non-urgent requests to ensure that the user can quickly obtain computing power support when triggering an adjustment, and avoid adjustment response delays caused by resource occupation.

[0100] When fluctuations in resource load are detected, resource allocation is dynamically adjusted. If light requests experience delays, the resource ratio of heavy requests in the same batch is appropriately reduced. If GPU resource utilization is low, some medium-level computing power requests are scheduled to GPU acceleration. The balance of resource load is maintained through real-time adjustments.

[0101] Specifically, the steps for identifying differences in demand include:

[0102] This system dynamically captures and classifies adjustment commands. It acquires user interactions on the filtering interface via the front end, such as clicking the edit button, adjusting parameter sliders, and entering new filtering conditions. The text descriptions of these operations are extracted and associated with user IDs to match the context cache. Intent vectors and historical user filtering operation logs are retrieved, and a command classification model is used to categorize user adjustment commands into four types: dimension addition commands, precision modification commands, target replacement commands, and condition deletion commands. Dimension addition commands add new filtering dimensions beyond the existing conditions; precision modification commands adjust numerical thresholds; target replacement commands completely change the filtering object; and condition deletion commands remove an existing condition. The command classification model takes text descriptions of user interactions, associated intent vectors, and historical filtering logs as input. By fusing multi-source data features, including semantic dimension weights, user profile features, and domain knowledge graphs, it outputs classification results for the four types of adjustment commands: dimension addition, precision modification, target replacement, and condition deletion. A transfer learning strategy based on the BERT model is used, combined with industry-specific labeled data for fine-tuning to enhance classification accuracy in professional scenarios.

[0103] The number of new filtering dimensions added in this adjustment is counted. Based on the target value and the current value, the change in precision is calculated. The cosine similarity between the calculated new target semantic vector and the original target semantic vector is used as the target relevance. The number of new filtering dimensions, the change in precision, and the target relevance are combined and weighted to generate an adjustment impact factor, thereby quantifying the complexity of the adjustment.

[0104] An incremental computation scheme is generated, employing differentiated strategies for different adjustment instruction types. These include: using dual-channel semantic encoding to record the correlation strength between each screening dimension, generating a semantic correlation matrix, determining the correlation between the added / deleted dimension and the original screening conditions, obtaining the incremental coefficient based on the product of the correlation and the adjustment influence factor, plus the base coefficient, and calculating the incremental computation volume in conjunction with the original computation volume. Idle nodes in the reserved resource pool are prioritized for computation to avoid consuming the core resources of the original request. For precision modification instructions, based on the computing power requirement level, lightweight adjustments are fine-tuned directly on the current node, while heavyweight adjustments are broken down into computation steps to gradually improve precision, with intermediate results updated synchronously in each iteration. For target replacement instructions, the interpretation rules are invoked to re-parse the new target, reusing the basic resources of the original computing node through a federated learning framework, replacing only the computation modules related to the target, and allocating idle computing power based on resource load to minimize resource consumption while ensuring computational accuracy.

[0105] Resource priority scheduling for adjustment instructions is performed. The priority of the adjustment instructions is determined by combining the urgency and the current adjustment impact factor. The urgency reflects the user's need for response speed, and the adjustment impact factor reflects the complexity of the adjustment. After the two are weighted and combined, the adjustment instructions with higher priority will be given priority to obtain resources.

[0106] A token bucket mechanism is used to control the access frequency of concurrent adjustments, allowing only instructions holding tokens to enter the resource scheduling queue to avoid node overload. High-priority adjustments use reserved backup resources first. If resources are still insufficient, the resources of the original requests are gradually downgraded in the order of non-urgent requests, lightweight requests, and heavyweight requests. The released computing power is prioritized to ensure adjustment needs, ensuring that adjustment instructions can quickly obtain the necessary computing support, thereby generating resource preemption results, which include the ID of the downgraded / suspended original request and the released computing power value, as well as the final allocation node and computing power ratio of the adjustment instructions.

[0107] Specifically, the steps for state synchronization include:

[0108] Before and after the adjustment, state snapshots containing filtering results, computing node identifiers, and processing progress are generated respectively. The hash value of the intermediate result of incremental computation is calculated using the SHA-256 algorithm. The state snapshot is backed up through a distributed storage system, associated with the user ID and request ID, and synchronized to all associated nodes through a distributed protocol, including the intelligent agent responsible for semantic parsing and data processing. The state snapshot is synchronized to all associated nodes through the Raft protocol. The Leader node is responsible for verifying the consistency of the state snapshots of each Follower node. If they are inconsistent, incremental synchronization is triggered.

[0109] For the intermediate results of incremental calculation, extract the hash values ​​of the results before and after adjustment. Subtract the ratio of the number of identical hash values ​​to the total number of hash values ​​from the base coefficient to obtain the hash difference rate. Set a difference threshold for hash comparison. If the hash difference rate is less than the difference threshold, directly merge the incremental results. Otherwise, locate the subset of data points corresponding to the differences and trigger local recalculation to ensure consistency of results across nodes.

[0110] Finally, the adjustment results are fed back and the model is iterated. The final adjusted results are pushed to the user, along with an explanation of the impact of the adjustment, such as the change in the number of results after adding new conditions and the impact of precision adjustment on the range of results. User satisfaction with the adjustment results is collected, and the adjustment effect is evaluated in combination with the compatibility of the adjustment results with the original screening logic.

[0111] Data such as adjustment type, adjustment influencing factors, and adjustment effect evaluation are injected into the intent model and scheduling model. The model parameters are updated through online learning, such as correcting the user's adjustment probability for specific dimensions and optimizing resource reservation strategies, so as to make subsequent adjustment predictions more accurate.

[0112] Specifically, the steps involved in collaborative screening include:

[0113] Dynamic allocation of multi-agent roles is implemented. Based on the complexity of the intent vector and the level of computing power requirements, three types of agents are determined: analytical agents, computational agents, and feedback agents. Among them, analytical agents, based on dynamic attention masks, are responsible for transforming fuzzy requirements into precise conditions; computational agents, according to the resource allocation results, undertake specific data operations, with lightweight tasks handled by the CPU and complex tasks by the GPU; feedback agents correlate and adjust influencing factors and monitor user operations in real time to predict fine-tuning needs.

[0114] Based on the success rate of the agent in handling similar tasks in the past, the intent complexity matching is obtained. Based on the fit between the agent's current load and the task's computing power requirements, the computing power adaptability is obtained. Combining the intent complexity matching and computing power adaptability, the agent matching degree is calculated to allocate tasks. Agents with high matching degrees take priority in accepting tasks. The agent collaboration topology is generated, including role type, number, and communication link.

[0115] Subtasks are decomposed and prioritized. The overall screening task is decomposed into three levels: semantic dimension, data range, and operation type. The first-level task is split according to the intent vector dimension, and the second-level data range is determined by the incremental computing scheme. The data is then assigned to the corresponding computational agents according to the operation type.

[0116] Based on the current computing power usage ratio, the agent load rate is obtained. The lower the load, the higher the priority correction value. The priority of subtasks is obtained by weighted calculation based on the urgency, agent load rate, and adjustment influencing factors. Subtasks with higher priority will get resources first. A subtask list is generated based on the priority, including dimension ID, data range, operation type, and priority label.

[0117] Cross-agent collaborative computing and conflict arbitration are carried out. Each agent processes sub-tasks in parallel and transmits data through a real-time communication protocol. The secondary parsing results of the analytical agent are pushed to the computational agent in real time with the latency controlled at an extremely low level. Once the output results of two computational agents conflict, a three-dimensional confidence score is invoked, including semantic matching degree and domain consistency. The result weight of each conflicting computational agent is obtained through weighted calculation, and the result of the computational agent with the highest weight is taken as the valid result.

[0118] When there is a sudden increase in the load of an agent, the feedback agent automatically migrates some tasks to the same type of low-load agent based on the resource pool data, so as to achieve dynamic balance of computing power.

[0119] Multi-level result aggregation is performed. In the basic aggregation stage, the sub-task results of each computational agent are reorganized according to the intent vector dimension and consistency is verified by hash difference rate. In the enhanced aggregation stage, combined with the quality score adjustment, the dimension with the number of user adjustments reaching the adjustment threshold is given higher weight, strengthening its proportion in the result. After the aggregation is completed, the aggregation result is converted into a user-friendly form, the mapping rules are associated with the user profile, and multi-level aggregation results are generated, including raw data and visualization charts.

[0120] Collect task completion data from each agent, calculate the task completion rate, and obtain the agent performance score by weighted calculation based on the difference between the actual computing power consumed and the theoretical requirement, combined with the difference between the base coefficient and this difference ratio and the task completion rate. Set a performance threshold to determine whether to optimize the agent.

[0121] Once the agent's performance score falls below the performance threshold, parameter iteration is triggered. The analytical agent updates the attention mask weights to improve the parsing accuracy, the computational agent optimizes the task splitting granularity to balance the load, and the feedback agent corrects the prediction threshold to respond and adjust more quickly, thereby generating the agent's iteration parameters.

[0122] Specifically, the feedback verification steps include:

[0123] The feedback similarity between the multi-level aggregation results and the intent vector is calculated using cosine similarity to assess the degree of matching between the results and the user's initial needs. Once the feedback similarity is less than the semantic matching threshold, it is marked as semantic deviation.

[0124] The consistency of adjustment is quantified by the degree of deviation between the actual difference rate and the expected difference rate. This includes: setting the expected difference rate based on the adjustment impact factor; obtaining the difference rate of the results before and after adjustment based on the overlap between the adjusted results and the original results; comparing the actual difference rate of the results before and after adjustment with the expected difference rate; and calculating the consistency of adjustment by combining the base coefficient. Once the consistency of adjustment is less than the consistency threshold, it is marked as an adjustment failure.

[0125] Extract historical features of the user's last three filtering results, calculate the matching ratio between the current result and the historical features to reflect the consistency between the result and the user's long-term preferences. Once the matching ratio is less than the historical matching threshold, it is marked as historical deviation.

[0126] Based on the combined results of the three levels of verification, a comprehensive verification score is generated through weighted calculation. In terms of weight allocation, similarity accounts for the highest proportion, followed by consistency and matching ratios. Once the comprehensive verification score is lower than the first verification threshold, the result is judged as abnormal and the abnormality type is marked.

[0127] For the marked abnormal results, a hierarchical feedback mechanism is designed, and a second verification threshold is set. For abnormal results with a comprehensive verification score lower than the second verification threshold, a deep feedback mode is adopted, which guides users to clarify the abnormality type through a structured form, including missed detection, false detection or dimensional deviation, and provides specific correction suggestions. Otherwise, a lightweight feedback mode is adopted, which collects users' binary evaluation of the results through a simple interactive interface and associates it with user profiles.

[0128] User feedback is categorized and coded. Items that do not meet the criteria are considered missed detections, while items that do not meet the criteria are considered false detections. Imbalance in the weights of core screening dimensions is considered a dimension bias. After user feedback is categorized and coded, the effectiveness of the feedback is calculated by integrating user expertise and the level of detail in the feedback.

[0129] Based on the comprehensive verification score and feedback data, a three-level optimization is triggered. When the comprehensive verification score is lower than the second verification threshold and the feedback validity is greater than the first feedback threshold, an emergency correction is performed. The incremental calculation scheme is immediately invoked, the relevant sub-tasks are re-executed, and the second result is pushed after correction. When the comprehensive verification score is between the second and first verification thresholds or the feedback validity is between the first and second feedback thresholds, iterative module calculations are performed. For semantic deviations, the dual-channel semantic encoding model is updated, and the training weights of abnormal cases are increased. For adjustments that fail, the incremental calculation formula is optimized. Otherwise, the strategy is upgraded, and the agent collaboration topology is fine-tuned based on implicit user feedback.

[0130] The optimization effect is evaluated by matching the corrected results with user feedback, and the process is iterated until the preset standard is reached.

[0131] Every month, all verification data is aggregated, including comprehensive verification score, feedback effectiveness, and correction accuracy, to establish a performance baseline. The evaluation indicators cover three dimensions: stability, adaptability, and efficiency. The stability indicator is quantified by the volatility of the monthly abnormal result ratio. The adaptability indicator is grouped by user profile to calculate the average feedback effectiveness of each group. The efficiency indicator focuses on the average time and resource consumption of emergency corrections. A long-term performance score is generated by integrating the above three indicators, with stability and adaptability having higher weights and efficiency having lower weights.

[0132] If the performance score remains below the baseline threshold for three consecutive months, an architecture-level optimization is triggered, reconstructing the reinforcement learning state space to improve the adaptability of resource scheduling, ultimately achieving continuous improvement in system performance.

[0133] Example 2:

[0134] Please see Figure 4 Another embodiment of the present invention provides a digital employee collaboration screening system, comprising: a parsing module, an allocation module, a dynamic adjustment module, and a feedback module;

[0135] The parsing module is used to avoid information confusion by physically isolating the contextual data of different users, receiving concurrent requests from multiple users and assigning unique IDs to map to independent memory partitions, calling user profiles to generate adjustment factors, generating attention masks and exclusive stop word lists after weighted filtering of request text vectors, starting dual-channel semantic encoding to generate initial semantic vectors, identifying conflicting requests through semantic conflict detection and generating interpretation rules, and generating intent vectors and three-dimensional confidence scores through three-level intent calibration.

[0136] The allocation module is used to collect request features in real time through the computing power monitoring nodes of the digital employee cluster, quantify the semantic complexity and correlation range of the requests to determine the computing power demand level, calculate the urgency and priority by combining the user's real-time operation behavior and decision speed preference, sort requests of the same computing power demand level, build a dual-objective reinforcement learning model to train the scheduling strategy to obtain the real-time scheduling strategy, allocate corresponding computing nodes and computing power proportions to requests of different computing power demand levels according to the real-time scheduling strategy, predict the user's adjustment intention and reserve spare resources, and dynamically balance the resource load.

[0137] The dynamic adjustment module is used to acquire and classify user adjustment instructions, calculate adjustment impact factors, generate incremental calculation schemes for different types of adjustment instructions, allocate resources to adjustment instructions through priority scheduling, downgrade existing requested resources when necessary, synchronize state snapshots before and after adjustment, ensure consistency of results across multiple nodes through hash difference verification, collect user feedback, and iteratively optimize the intent model and scheduling model.

[0138] The feedback module is used to assign roles to multiple agents and generate collaborative topologies, break down subtasks and bind priorities, generate multi-level aggregated results through cross-agent collaborative computing and conflict arbitration, perform multi-dimensional verification of the results, collect user feedback and trigger emergency corrections, module iterations or policy upgrades, regularly aggregate data to update the performance baseline, and continuously optimize system performance to improve the accuracy and efficiency of screening.

[0139] Working principle and effects:

[0140] Upon receiving multiple user requests, semantic isolation is achieved through an independent context cache. Adjustment factors are generated to optimize text vectors based on user profiles, and initial semantic vectors are generated through dual-channel semantic encoding. Conflicting requests are identified through semantic conflict detection, generating specific interpretation rules and intent vectors. Computational power requirements are quantified based on request complexity, and urgency and priority are generated based on user behavior. Resources are dynamically allocated through reinforcement learning strategies, and computational power is reserved based on predicted adjustment intentions. When users make adjustments, instructions are captured and categorized to generate incremental computation schemes. Efficient response and multi-node consistency are ensured through resource priority scheduling and state synchronization. Multiple agents collaborate according to their roles, decompose tasks, and dynamically balance the load. Finally, the system undergoes multi-dimensional verification and feedback iteration to effectively avoid semantic cross-confusion, improve resource allocation accuracy and adjustment response speed, ensure the accuracy and adaptability of collaborative filtering results, significantly reduce user operation costs, and meet the efficient filtering needs in multi-user concurrent scenarios.

[0141] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A digital employee collaboration screening method, characterized in that, include: It receives concurrent requests from multiple users, performs semantic isolation, generates adjustment factors based on user profile features, generates an initial semantic vector through dual-channel semantic encoding, and performs semantic conflict detection. Collect request characteristics in real time, calculate the computing power requirement level of different requests, and allocate resources accordingly; By obtaining user adjustment instructions through the front end, identifying differences in requirements, generating incremental calculation plans, and performing state synchronization control; Classify digital employees, conduct collaboration screening, and perform feedback verification; The steps of semantic isolation include: Assign a unique ID to each request, build a context cache, and synchronously adopt the LRU eviction policy; The user profiling API is called to obtain a 3D feature vector, and the semantic adjustment coefficient is obtained by weighted summation. The adjustment factor is then calculated. Construct a basic text vector, multiply it by the adjustment factor, and generate a weighted text vector; Based on preset filtering rules, the weighted text vector is decomposed, and semantic dimensions that do not conform to the filtering rules are removed; The filtered weighted text vectors are processed by the softmax function to generate an attention mask, and the decision pattern features of the context cache are extracted to generate a dedicated stop word list. Dual-channel semantic encoding is initiated to generate surface vectors and domain entity vectors. The surface vectors and domain entity vectors are then fused through residual connections to generate an initial semantic vector.

2. The digital employee collaboration screening method according to claim 1, characterized in that, The steps of semantic conflict detection include: For the initial semantic vectors of concurrent requests, calculate the semantic similarity between any two requests; If the semantic similarity is greater than the semantic similarity threshold and the two requests are from different users, they are determined to be conflicting requests and included in the conflicting request list; otherwise, they are determined to be non-conflicting requests and the three-level intent calibration is performed directly. For conflict requests, perform logical structure encoding, calculate the nesting depth of modifiers, extract and combine numerical range boundary features to generate structural feature fingerprints; The decision tree generation algorithm is invoked, with the logical operator sequence as the root node, the nesting depth of modifiers as the intermediate nodes, and the user's historical preferences as the leaf nodes, to generate the filtering logic. Perform three-level intent calibration, fuse the results of the three-level intent calibration, generate an intent vector, and generate a three-dimensional confidence score, including semantic matching degree, domain consistency, and historical fit degree, and obtain the comprehensive confidence score through weighted calculation.

3. The digital employee collaboration screening method according to claim 2, characterized in that, The steps involved in resource allocation include: The semantic depth level is divided according to the breadth of cross-domain entity associations in the intent vector and the degree of historical fit in the three-dimensional confidence score. A document association graph is generated based on entity link information and the association range level is divided according to the number of nodes. The computing power requirement level is generated by weighted calculation. Extract behavioral features from real-time user operation logs and quantify them into sensitivity. Then, through weighted calculation, obtain the sensitivity of user operation behavior. Combine the decision speed preference features in the user profile to calculate the urgency of each request. Based on the urgency and the overall confidence level, the priority of each request is obtained through weighted calculation. Requests with the same computing power requirement level are then sorted a second time based on the priority to generate a request queue. Real-time collection of multi-dimensional features, including the proportion of requests at each computing power demand level in the current queue, real-time CPU and GPU resource load, and performance indicators of recent lightweight and heavyweight requests, to construct a multi-dimensional feature state space and design a dual-objective reward function. The system is trained using reinforcement learning algorithms to adjust resource allocation under different states and generate real-time scheduling strategies. Initial resource allocation is performed according to the real-time scheduling strategy, matching corresponding computing nodes to requests with different computing power requirements, and allocating corresponding computing power proportions according to the priority of the requests. The system acquires user interaction behavior with the filtering interface, generates a real-time behavior score, generates a historical habit score based on the user's past similar filtering operation records, and performs weighted merging to obtain the adjustment probability. When the adjustment probability is greater than the adjustment threshold, the resource adjustment coefficient is calculated, and the reserved resources are calculated in combination with the basic coefficient and computing power standard, and temporarily allocated from the current user's non-urgent requests. When fluctuations in resource load are detected, resource allocation is dynamically adjusted.

4. The digital employee collaboration screening method according to claim 3, characterized in that, The steps to identify differences in requirements include: Acquire user interaction behavior on the filtering interface, capture dynamic adjustment commands, and classify user adjustment commands using a command classification model; The number of new screening dimensions added in this adjustment is counted, the magnitude of the change in precision and the relevance of the target are calculated, and the adjustment impact factor is generated by weighted fusion. Incremental computing schemes are generated for different types of adjustment instructions, including: generating a semantic correlation matrix based on dual-channel semantic encoding, calculating incremental coefficients and incremental computing power, and scheduling idle nodes in the reserved resource pool; differentiating precision modification instructions according to computing power requirement levels; re-analyzing new targets for target replacement instructions, reusing the basic resources of the original computing nodes, and allocating idle computing power. The system acquires the priority of adjustment instructions and uses a token bucket mechanism to control the access frequency of concurrent adjustments, allowing only instructions holding tokens to enter the resource scheduling queue and generating resource preemption results.

5. The digital employee collaboration screening method according to claim 4, characterized in that, The steps for state synchronization include: Before and after the adjustment, state snapshots are generated respectively. These state snapshots are backed up through a distributed storage system, associated with user ID and request ID, and synchronized to all associated nodes through a distributed protocol. For the intermediate results of incremental calculation, extract the hash values ​​of the results before and after adjustment, and obtain the hash difference rate by subtracting the ratio of the number of identical hash values ​​to the total number of hash values ​​from the base coefficient. Set the difference threshold to perform difference hash comparison. If the hash difference rate is less than the difference threshold, the incremental results are directly merged; otherwise, the subset of data points corresponding to the differences is located and a local recalculation is triggered. The adjustment results are fed back and the model is iterated. User satisfaction with the adjustment results is collected. The adjustment effect is evaluated by combining the compatibility between the adjustment results and the original screening logic. The parameters of the intent model and scheduling model are updated through online learning.

6. The digital employee collaboration screening method according to claim 5, characterized in that, The collaborative screening process includes the following steps: Based on the complexity of the intent vector and the level of computing power requirements, the agents are assigned tasks, the matching degree of the agents is calculated, and the agent cooperation topology is generated. The screening task is broken down into three levels, and the priority of the subtasks is calculated to generate a list of subtasks. Each agent processes subtasks in parallel. When the outputs of two computational agents conflict, the weight of each conflicting computational agent's result is calculated, and the result of the computational agent with the highest weight is taken as the valid result. Perform multi-level result aggregation to generate multi-level aggregated results; The task completion rate is calculated based on the task completion data. At the same time, the performance score of the agent is obtained through weighted calculation, and a performance threshold is set. When the agent's performance score is less than the performance threshold, parameter iteration is triggered.

7. The digital employee collaboration screening method according to claim 6, characterized in that, The feedback verification steps include: Calculate feedback similarity, adjustment consistency, and matching ratio, and filter out semantic deviation, adjustment failure, and historical deviation; The comprehensive verification score is generated by weighted calculation based on the combined results of the three levels of verification. If the comprehensive verification score is lower than the first verification threshold, the result is determined to be abnormal and the abnormality type is marked. A second verification threshold is set. For abnormal results where the comprehensive verification score is lower than the second verification threshold, a deep feedback mode is adopted; otherwise, a lightweight feedback mode is adopted. Based on the comprehensive verification score and feedback data, a three-level optimization is triggered, and all verification data is aggregated to generate a long-term performance score. If the long-term performance score is lower than the baseline threshold for three consecutive months, an architecture-level optimization is triggered.

8. A digital employee collaboration screening system, used to implement the digital employee collaboration screening method as described in any one of claims 1-7, characterized in that, include: The module includes a parsing module, an allocation module, a dynamic adjustment module, and a feedback module. The parsing module is used to isolate contextual data from different users and identify conflicting requests through semantic conflict detection. The allocation module is used to collect request characteristics in real time to calculate the computing power demand level, combine the user's real-time operation behavior and decision speed preference to calculate the urgency and priority, and construct a real-time scheduling strategy. The dynamic adjustment module is used to acquire and classify user adjustment instructions, generate incremental calculation schemes, and allocate resources to the adjustment instructions through priority scheduling. The feedback module is used to assign roles to multiple agents and generate collaborative topologies, decompose subtasks and bind priorities, generate multi-level aggregated results through cross-agent collaborative computing and conflict arbitration, and perform multi-dimensional verification on the results.

Images