Digital employee collaborative screening method and system

Abstract

The invention discloses a digital employee collaborative screening method and system, and belongs to the technical field of artificial intelligence, and the method comprises the steps: receiving multi-user concurrent requests, carrying out semantic isolation, generating an adjustment factor in combination with user portrait features, generating an initial semantic vector through two-channel semantic coding, and carrying out semantic conflict detection, thereby effectively avoiding semantic cross and confusion, and improving the user experience. The intention analysis of multi-user concurrent requests is realized; acquiring request features in real time, calculating computing power demand levels of different requests, sequencing the requests with the same computing power demand level, and performing resource allocation; a user adjustment instruction is obtained through a front end, a demand difference is identified, an incremental calculation scheme is generated, and state synchronization control is carried out; according to the method, digital employees are divided into analytic, computational and feedback intelligent agents, collaborative screening is carried out, feedback verification is carried out, it is ensured that a screening result better meets user requirements, and real-time capability iteration of a digital employee cluster is achieved.

Description

Technical Field

[0001] The present invention relates to a digital employee collaborative screening method and system, and belongs to the technical field of artificial intelligence. Background Art

[0002] In the digital age, information screening, as a core link in corporate decision-making, content management, and resource matching, has undergone a multi-stage transformation in its technological evolution from manual to intelligent-driven. Early information screening relied on manual work, with professionals reviewing and marking key information one by one. It could make judgments on implicit knowledge through human experience, but it had limitations, such as efficiency bottlenecks, lack of consistency, and limited scale. With the popularization of information technology, automated screening tools based on rule engines and keyword matching have become mainstream, and batch filtering of candidate information can be achieved through predefined screening rules or keyword dictionaries. In recent years, large language model technologies represented by Transformer have broken through the limitations of traditional rules and achieved semantic-level understanding and multimodal processing.

[0003] However, there are still many challenges in practical applications. In the scenario of multi-user concurrent requests, the adaptability problem is particularly prominent: due to the significant differences in the screening needs of different users, a large number of differentiated requests are simultaneously poured in. If these differentiated requests contain semantically overlapping or even ambiguous screening logic, for example, similar expressions correspond to completely different screening targets, it is very easy to cause rule confusion during the parallel parsing process, resulting in misjudgment or omission of screening conditions. At the same time, there are obvious hierarchical differences in the computational complexity of differentiated requests. The computing power requirements for simple keyword matching and cross-document semantic association analysis differ by several orders of magnitude. This difference makes it difficult to implement resource scheduling mechanisms. A precise balance is required. The high response speed requirements of lightweight requests conflict with the reliance of heavyweight requests on continuous computing power, often leading to unnecessary delays in lightweight requests due to resource preemption, or long waits for heavyweight requests due to insufficient computing power. What's more complicated is that if users dynamically adjust their requirements during the screening process, such as temporarily modifying the screening dimension or accuracy, and if other differentiated requests are processed at the same time, the resources required for dynamic adjustment may be occupied, which not only causes response delays to the adjustment instructions, but may also cause the screening process to be out of touch with the user's latest needs due to the disorder of the multi-request state synchronization mechanism, ultimately affecting the accuracy and timeliness of the screening results. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a digital employee collaborative screening method and system, which solves the adaptability problem in multi-user concurrent scenarios through the collaborative mechanism of multi-role digital employees and dynamic resource scheduling strategy.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Digital collaborative employee screening methods, including:

[0007] Receive concurrent requests from multiple users, perform semantic isolation, generate adjustment factors based on user profile features, generate initial semantic vectors through dual-channel semantic encoding, and perform semantic conflict detection;

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

[0009] The front-end obtains user adjustment instructions, identifies demand differences, generates incremental calculation plans, and performs state synchronization control;

[0010] Divide digital employees into categories, conduct collaborative 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 simultaneously adopt the LRU elimination strategy;

[0013] Call the user portrait API to obtain the three-dimensional feature vector, obtain the semantic adjustment coefficient through weighted summation, and calculate the adjustment factor;

[0014] Constructing a basic text vector and multiplying it by the adjustment factor to generate a weighted text vector;

[0015] Based on preset filtering rules, the weighted text vector is disassembled to remove semantic dimensions that do not meet the filtering rules;

[0016] Performing a softmax function on the filtered weighted text vector to generate an attention mask, extracting decision mode features from the context buffer, and generating a dedicated stop word list;

[0017] Start dual-channel semantic coding to generate a surface vector and a domain entity vector, and fuse the surface vector and the domain entity vector through residual connection 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 directly performed;

[0021] For conflicting requests, we encode the logical structure, calculate the nesting depth of modifiers, extract and combine the numerical range boundary features, and generate the structural feature fingerprint;

[0022] Call the decision tree generation algorithm, with the logical operator sequence as the root node, the modifier nesting depth as the intermediate node, and the user's historical preference as the leaf node to generate the screening logic;

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

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

[0025] Semantic depth levels are divided based on the breadth of associations between cross-domain entities 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 scope level is divided by the number of nodes. The computing power requirement level is generated through weighted calculation.

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

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

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

[0029] Collect multi-dimensional features in real time, construct a multi-dimensional feature state space, and design a dual-objective reward function;

[0030] Through training with reinforcement learning algorithms, resource allocation is adjusted under different conditions to generate real-time scheduling strategies;

[0031] Perform initial resource allocation according to the real-time scheduling strategy, match corresponding computing nodes to requests with different computing power requirements, and allocate corresponding computing power proportions according to the priority of the requests;

[0032] Obtain the user's interaction behavior with the screening interface to generate a real-time behavior comprehensive score. Based on the user's past similar screening operation records, generate a historical habit comprehensive score, and perform 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 by combining the basic coefficient and the computing power standard, and temporarily allocated from the current user's non-urgent request;

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

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

[0036] Obtain user interaction behaviors on the screening interface, capture dynamic adjustment instructions, and classify user adjustment instructions using the instruction classification model;

[0037] Count the number of newly added screening dimensions in this adjustment, calculate the accuracy change range and target correlation, and generate the adjustment impact factor through weighted fusion;

[0038] Generate incremental computing solutions for different adjustment instruction types, including: generating a semantic relevance matrix based on dual-channel semantic encoding, calculating incremental coefficients and incremental computing volume, and scheduling idle nodes in the reserved resource pool; performing differentiated processing on precision modification instructions based on computing power requirements; re-parsing new targets for target replacement instructions, reusing the original computing node's basic resources, and allocating idle computing power;

[0039] Obtain the adjustment instruction priority and use the token bucket mechanism to control the access frequency of concurrent adjustments. Only instructions holding tokens are allowed to enter the resource scheduling queue to generate resource preemption results.

[0040] Specifically, the state synchronization steps include:

[0041] Before and after the adjustment, a state snapshot is generated. The state snapshot is backed up by a distributed storage system, associated with the user ID and the request ID, and synchronized to all related nodes through a distributed protocol.

[0042] For the intermediate results of the incremental calculation, extract the result hash values ​​before and after the 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, and set a difference threshold to perform a difference hash comparison;

[0043] If the hash difference rate is less than the difference threshold, the incremental results are directly merged; otherwise, the sub-dataset corresponding to the difference is located and a local recalculation is triggered;

[0044] Provide feedback on adjustment results and model iteration, collect user satisfaction evaluations on the adjustment results, evaluate the adjustment effect based on the compatibility between the adjustment results and the original screening logic, and update the intent model and scheduling model parameters through online learning.

[0045] Specifically, the steps of collaborative screening include:

[0046] The agents are divided into different groups based on the complexity of the intent vector and the level of computing power required. The matching degree of the agents is calculated to assign tasks and generate the agent collaboration topology.

[0047] Decompose the screening task into three levels, calculate the subtask priority, and generate a subtask list;

[0048] Each agent processes subtasks in parallel. When the output results of two computing agents conflict, the weights of the results of the conflicting computing agents are calculated, and the result of the computing agent with the highest weight is taken as the valid result.

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

[0050] Calculate the task completion rate based on the task completion data, obtain the agent performance score through weighted calculation, and set the performance threshold;

[0051] When the agent 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 deviations, adjustment failures, and historical deviations;

[0054] The three-level verification results are combined to generate a comprehensive verification score through weighted calculation. Once 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] Setting a second verification threshold, and for abnormal results where the comprehensive verification score is lower than the second verification threshold, adopting a deep feedback mode; otherwise, adopting a light feedback mode;

[0056] Based on the comprehensive verification score and feedback data, three-level optimization is triggered, and the full 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, architecture-level optimization is triggered.

[0057] Digital employee collaborative screening system, including: parsing module, allocation module, dynamic adjustment module and feedback module;

[0058] The parsing module is used to isolate contextual data of 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, calculate the urgency and priority based on the user's real-time operation behavior and decision-making speed preference, and build a real-time scheduling strategy;

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

[0061] The feedback module is used to assign multi-agent roles and generate a collaborative topology, decompose subtasks and bind priorities, generate multi-level aggregation results through cross-agent collaborative calculation and conflict arbitration, and perform multi-dimensional verification on the results.

[0062] Beneficial effects of the present invention:

[0063] Multi-user semantic isolation is achieved through independent context cache areas, and dual-channel semantic encoding is combined with grammatical and domain features. At the same time, with the help of semantic conflict detection and unique interpretation rule generation, semantic cross-confusion of requests from different users is effectively avoided, the accuracy of intent parsing is improved, and the problem of semantic interference under multi-user concurrency is solved; redundant computing power consumption is reduced through incremental computing solutions, and the adjustment response speed is improved by combining priority scheduling and resource preemption mechanisms. With the help of state snapshot synchronization and hash difference verification, the consistency of multi-node results is guaranteed, and the problems of slow adjustment response and chaotic state synchronization are solved; and through dynamic role allocation and conflict arbitration mechanisms, the consistency of task decomposition and calculation is ensured. The feedback verification system continuously optimizes model parameters and collaboration strategies through multi-dimensional verification and closed-loop iteration, improves the matching degree between results and user needs, solves the problems of inconsistent collaboration and insufficient result accuracy, and continuously corrects system deviations to improve the matching degree between results and user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Flowchart of collaborative screening methods for digital employees;

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

[0066] Figure 3 A flowchart of resource allocation in the present invention;

[0067] Figure 4 Screening system architecture diagram for digital workforce collaboration. DETAILED DESCRIPTION

[0068] The technical solution of the present invention is described in detail below through 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 on the technical solution of the present invention. 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, this embodiment introduces a digital employee collaborative screening method, which includes 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 portrait features, perform weighted processing on the request text vector and filter out invalid dimensions, generate an initial semantic vector by fusing grammar and domain features through dual-channel semantic encoding, perform semantic conflict detection to identify conflicting requests, generate structural feature fingerprints and unique interpretation rules, and generate intent vectors and three-dimensional confidence scores through three-level calibration, effectively avoiding semantic cross-confusion, improving parsing accuracy and personalized adaptability, and achieving accurate intent parsing of concurrent requests from multiple users;

[0072] Step S2: Request features are collected in real time through the computing power monitoring nodes of the digital employee cluster, and the computing power requirement levels of different requests are calculated. The urgency and priority levels are generated based on the real-time user operation behavior, and requests with the same computing power requirement level are sorted to avoid delays caused by resource preemption for lightweight requests. The resource scheduling strategy is trained based on reinforcement learning, with response delay and resource utilization as dual reward goals. Resources are allocated and a dynamic adjustment mechanism for request priority is introduced. At the same time, user adjustment intentions are predicted and resources are reserved to dynamically balance the load. This improves the accuracy and flexibility of resource allocation, achieves a balance between response efficiency and utilization efficiency, and adapts to dynamic adjustment needs.

[0073] Step S3: The front-end obtains user adjustment instructions, identifies demand differences, calculates adjustment impact factors, quantifies complexity, generates incremental calculation plans for different adjustment types to reduce resource consumption, and determines the priority of adjustment instructions based on urgency and adjustment impact factors. Resource scheduling ensures rapid response to adjustments, downgrades original requested resources when necessary, and performs state synchronization control. By generating state snapshots before and after the adjustment and synchronizing them to related nodes, hash difference verification is used to ensure consistency of multi-node results. Large differences trigger local recalculation. At the same time, user feedback is collected, and adjustment data is injected into the model for iterative optimization to improve the accuracy of subsequent adjustment predictions. The overall process effectively ensures the accuracy of demand difference identification and the consistency of state synchronization.

[0074] Step S4: Divide digital workers into three types of agents: analytical, computational, and feedback agents. Collaborative screening is performed. Tasks are assigned based on historical performance and load adaptability to generate a collaborative topology. Tasks are broken down and prioritized. Cross-agent collaborative computing and conflict arbitration ensure consistent results. Dynamically migrate tasks to balance computing power when load spikes occur.

[0075] Step S5: Before outputting the screening results, feedback verification is performed on the results generated by multi-agent collaboration. The matching degree between the evaluation results and the user's initial needs, adjustment instructions and long-term preferences is evaluated through multi-dimensional verification. Anomalies are marked, and a layered feedback mechanism is designed for anomalies. Feedback is collected and effectiveness is quantified based on user portraits. 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 screening results are more in line with user needs and realize real-time capability iteration of the digital employee cluster.

[0076] Specifically, the steps of semantic isolation include:

[0077] To avoid information confusion caused by memory address conflicts, the context data of different users are isolated through physical isolation methods. When receiving concurrent requests from multiple users, such as screening requests initiated by user A and user B on the digital employee platform at the same time, a unique ID is assigned to each request through a distributed lock mechanism. A hash value is generated based on the ID and mapped to an independent memory partition through the hash value as a context cache area. The cache area size is set to the length of the request text. Times, only the text of the current request, user ID and initial parsing status are stored, where the parsing status includes unprocessed, processing, and completed. At the same time, the LRU elimination strategy is adopted. When the memory usage exceeds the preset ratio, the cache area with no operations within the preset elimination interval is released first, so as to achieve physical isolation of different user context data;

[0078] Call the user portrait API to obtain a three-dimensional feature vector containing knowledge density coefficient, domain specific accuracy, and decision-making speed preference, and compress the features in the three-dimensional feature vector into The semantic adjustment coefficient is calculated by weighted summation. The knowledge density coefficient reflects the user's familiarity with professional terms, the domain expertise reflects the user's focus on the field to which they belong, and the decision speed preference represents 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, and the request text is converted into a Word2Vec vector. This is used as the base text vector, and the base text vector is multiplied by the adjustment factor. This amplifies the weights of the semantic dimensions in the text that match the user's features, while relatively suppresses the weights of the semantic dimensions that do not match the user's features, thereby obtaining a weighted text vector. In this embodiment, the base coefficient is set to 1 to ensure that even if the eigenvalue 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 time, the weighted text vector needs to be filtered. According to the task scenario or domain characteristics, preset filtering rules, including numerical threshold filtering, domain relevance filtering and outlier filtering, will decompose the weighted semantic vector into multiple independent semantic dimensions, compare the above filtering rules one by one, and mark the semantic dimensions that do not meet the filtering rules as to be filtered and processed. At this time, the filtered weighted semantic vector is generated; among them, the numerical threshold filtering sets the minimum effective threshold of the semantic dimension weight, the 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, and the outlier filtering uses statistical methods to identify extreme values, determine them as calculation errors and filter them. The semantic dimensions to be filtered include weakly relevant dimensions, irrelevant dimensions and abnormal 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 determined by the domain classification model to be 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 abnormal dimension.

[0081] The filtered weighted text vector is processed by the softmax function to normalize the weights of all semantic dimensions to The probability value within the interval is finally generated into 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 buffer, the user's three most recent screening operation logs are queried to extract decision pattern features, including high-frequency screening dimensions, adjustment frequency, and result correction type. The TF-IDF algorithm is used to generate a user-specific stop word list to facilitate filtering of irrelevant entities during dual-channel semantic encoding.

[0082] Start dual-channel semantic encoding. The surface channel uses the BERT-base model to extract grammatical structure features and outputs surface vectors containing logical connectives and modifier positions. The deep channel uses the ERNIE-3.0 model to associate the enterprise knowledge graph, maps the filtered weighted text vector to the entity node, outputs the domain entity vector, and fuses the surface vector and the domain entity vector through residual connection to generate the initial semantic vector.

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

[0084] To extract and identify confusing request groups, semantic conflict detection is performed. For the initial semantic vectors of all concurrent requests, the cosine similarity formula is used to calculate the semantic similarity between any two requests, and whether there is 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, there is a semantic conflict between the two concurrent requests, which are determined to be conflicting requests and included in the conflicting request list;

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

[0087] For conflicting requests, we first generate structural feature fingerprints, encode the logical structure of the conflicting request text, convert logical connectives into operator sequences, calculate the modifier nesting depth, extract the boundary features of the numerical range, and combine them to generate structural feature fingerprints.

[0088] Based on the structural feature fingerprint, the interpretation rule generator is activated and the decision tree generation algorithm is called. With the logical operator sequence as the root node, the modifier nesting depth as the intermediate node, and the user's historical preference as the leaf node, a unique interpretation rule is generated for the conflicting request. After verification by the rule engine, it is converted into executable screening logic.

[0089] Perform three-level intent calibration. The first level is historical behavior backtracking, calculating the cosine similarity between the current request and the intent vectors of the user's last three screens. If the cosine similarity is less than the historical similarity threshold, the initial semantic vector is fine-tuned to increase the weight of the high-frequency screening dimension. The second level is domain knowledge constraint, calling the industry rule engine to correct the weight of the vector semantic dimension. The third level is real-time feedback prediction, using the LSTM model to predict the user's expected adjustments. The results of these three levels of intent calibration are integrated to generate an intent vector.

[0090] Generate a three-dimensional confidence score, including semantic matching, 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 three-dimensional feature analysis result package.

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

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

[0093] Extract three types of behavioral features from real-time user operation logs, including the length of time the cursor stays in the filter area, the number of refreshes within the sensitive interval, and the speed of character input in the input box. These three features are quantified into sensitivities and weighted to obtain the user's operation behavior sensitivity. Based on the decision-making speed preference characteristics in the user profile, the urgency of each request is calculated. The decision-making speed preference reflects the user's emphasis on response efficiency and, together with the operation behavior sensitivity, determines the level of urgency. The urgency is calculated by multiplying the difference between the base coefficient and the decision-making speed preference by the operation behavior sensitivity.

[0094] Based on urgency and comprehensive confidence, the priority of each request is obtained through weighted calculation. Requests with the same computing power requirement level are sorted again based on priority. Requests with higher priorities have more advantages in resource allocation, thus generating a sorted request queue.

[0095] Construct a state space containing multi-dimensional features and collect multi-dimensional features in real time, including the proportion of requests at each computing power requirement 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 dual objectives of minimizing response latency and optimizing resource utilization. A dual-objective reward function is designed. For request response latency, the closer the latency is to the preset reasonable range, the higher the reward; otherwise, the penalty will be incurred. For 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 through reinforcement learning algorithms, and resource allocation methods are adjusted under different conditions. The scheduling strategy is continuously optimized based on the rewards obtained until the scheduling strategy converges. A real-time scheduling strategy is generated to dynamically adjust the resource allocation logic according to the real-time status.

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

[0098] By analyzing the user's dwelling behavior in the filtering parameter area and the user's operational intention, the user's adjustment intention is predicted and resources are reserved to cope with dynamic needs. The user's interaction behavior with the filtering interface is obtained in real time, including the user's cursor dwelling in the filtering parameter area, the frequency of clicking on the filtering conditions, and the hovering duration on the adjustment button. These are converted into feature scores and integrated into these three feature scores to form a real-time behavior comprehensive score. The user's past filtering operation records are retrieved to obtain the proportion of users who ultimately performed adjustment operations when initiating filtering requests of the same type as the current one in the past. The higher the proportion, the easier it is for users to adjust in similar scenarios. This is converted into a feature score. The average number of adjustment operations per month is obtained. The higher the number, the more accustomed users are to adjusting the filtering results as a whole. This is converted into a feature score for integration 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 average monthly adjustment times by the reservation coefficient. The reserved resources are calculated by summing the resource adjustment coefficient and the basic coefficient, combined with the computing power standard. This allows a certain amount of backup resources to be reserved in advance and temporarily allocated from the current user's non-urgent requests. This ensures that computing power support is quickly obtained when the user triggers an adjustment, avoiding adjustment response delays caused by resource occupation.

[0100] When resource load fluctuations are detected, resource allocation is adjusted dynamically. If lightweight requests are delayed, the resource share of heavyweight requests in the same batch is appropriately reduced. If GPU resource utilization is low, some requests with medium computing power requirements are scheduled to GPU acceleration, maintaining resource load balance through real-time adjustments.

[0101] Specifically, the steps to identify demand differences include:

[0102] Dynamic adjustment instruction capture and type classification are performed. The front-end obtains the user's interactive behavior in the filtering interface, such as clicking the modify button, adjusting the parameter slider, and entering new filtering conditions. The text descriptions of these operations are extracted and associated with the user ID to match the context cache. The intent vector and the user's historical filtering operation log are called, and the instruction classification model is used to classify the user adjustment instructions, distinguishing four types of adjustment instructions: dimension addition instructions, precision modification instructions, target replacement instructions, and condition deletion instructions. Among them, dimension addition instructions add new filtering dimensions in addition to the original conditions, precision modification instructions adjust the numerical threshold, target replacement instructions completely change the filtering object, and condition deletion instructions remove a certain original condition. The instruction classification model takes the text description of the user's interactive behavior, the associated intent vector, and the historical filtering log as input. By fusing multi-source data features, including semantic dimension weights, user portrait features, and domain knowledge graphs, it outputs the classification results of the four types of adjustment instructions: dimension addition, precision modification, target replacement, and condition deletion. The transfer learning strategy based on the BERT model is fine-tuned with industry-specific labeled data to enhance the classification accuracy in professional scenarios.

[0103] Count the number of newly added filtering dimensions in this adjustment, calculate the precision change based on the target value and the current value, and use the cosine similarity between the calculated new target semantic vector and the original target semantic vector as the target relevance. Combine the number of newly added filtering dimensions, the precision change, and the target relevance to generate an adjustment impact factor through weighted fusion to quantify the complexity of the adjustment.

[0104] Generate incremental computing plans and adopt differentiated strategies for different types of adjustment instructions, including: based on dual-channel semantic encoding, record the correlation strength between each screening dimension, generate a semantic correlation matrix, determine the correlation between the newly added / deleted dimension and the original screening condition, and obtain the incremental coefficient based on the product of the correlation and the adjustment impact factor plus the basic coefficient. Combined with the original computing amount, the incremental computing amount is calculated, and the idle nodes in the reserved resource pool are preferentially scheduled to perform the calculation to avoid occupying the core resources of the original request; for precision modification instructions, based on the computing power requirement level, lightweight adjustments are directly fine-tuned on the current node, and heavyweight adjustments split the calculation steps to gradually improve the accuracy, and the intermediate results are updated synchronously with each iteration; for target replacement instructions, the interpretation rules are called to re-parse the new target, and the basic resources of the original computing node are reused through the federated learning framework. Only the computing modules related to the target are replaced, and the idle computing power is allocated according to the resource load situation to minimize resource consumption while ensuring the accuracy of the calculation;

[0105] Prioritize resource scheduling for adjustment instructions. The priority of adjustment instructions is determined by combining urgency and the current adjustment impact factor. Urgency reflects the user's demand for response speed, while the adjustment impact factor reflects the complexity of the adjustment. After weighted integration of the two, adjustment instructions with higher priorities are given priority in obtaining 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 give priority to using reserved backup resources. 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 guarantee the adjustment needs, ensuring that the adjustment instructions can quickly obtain the necessary computing support, thereby generating a resource preemption result, which includes the original request ID that was downgraded / suspended, the released computing power value, the final allocation node of the adjustment instruction, and the computing power ratio.

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

[0108] Before and after the adjustment, a state snapshot is generated, including the screening results, computing node identification, and processing progress. The hash value of the intermediate result of the incremental calculation is calculated using the SHA-256 algorithm. The state snapshot is backed up by a distributed storage system, associated with the user ID and request ID, and synchronized to all related nodes through a distributed protocol, including the intelligent agent responsible for semantic parsing and data processing. The state snapshot is synchronized to all related nodes through the Raft protocol. The leader node is responsible for verifying the consistency of the state snapshot of each follower node. If there is inconsistency, incremental synchronization is triggered.

[0109] For the intermediate results of the incremental calculation, extract the result hash values ​​before and after the 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. Then set a difference threshold for performing a difference hash comparison. If the hash difference rate is less than the difference threshold, directly merge the incremental results. Otherwise, locate the sub-dataset corresponding to the difference and trigger a local recalculation to ensure consistent results at each node.

[0110] Finally, feedback on the adjustment results and model iteration are conducted, and the final results after adjustment are pushed to users, along with an explanation of the impact of the adjustment, such as the change in the number of results after the new conditions are added, and the impact of the precision adjustment on the result range. User satisfaction evaluations on the adjustment results are collected, and the adjustment effect is evaluated based on the compatibility of the adjustment results with the original screening logic.

[0111] Inject data such as adjustment type, adjustment influencing factor, and adjustment effect evaluation into the intention model and scheduling model, and update model parameters through online learning, such as correcting the user's adjustment probability for specific dimensions and optimizing resource reservation strategies, to make subsequent adjustment predictions more accurate.

[0112] Specifically, the steps of collaborative screening include:

[0113] Dynamically allocate roles among multiple agents, and determine the division of labor among three types of agents based on the complexity of the intent vector and the level of computing power required. These agents include analytical agents, computational agents, and feedback agents. Analytical agents, based on dynamic attention masks, are responsible for converting fuzzy requirements into precise conditions. Computational agents, based on resource allocation results, undertake specific data operations, with lightweight tasks handled by the CPU and complex tasks by the GPU. Feedback agents adjust the impact factors in a correlated manner, monitoring user operations in real time to predict fine-tuning needs.

[0114] Based on the success rate of the agent's historical processing of similar tasks, the intention 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 intention complexity matching and computing power adaptability, the agent matching degree is calculated to assign tasks. Agents with high matching degrees are given priority to undertake tasks. The agent collaboration topology is generated, including role type, number, and communication links.

[0115] Perform subtask decomposition and priority binding, breaking down the overall screening task into three levels: semantic dimension, data range, and operation type. Split the first-level tasks based on the intent vector dimension, determine the second-level data range based on the incremental computing solution, and assign them to the corresponding computing agents based on the operation type.

[0116] Based on the inverse ratio of the current computing power usage, the agent load rate is obtained. The lower the load, the higher the priority correction value. The urgency, agent load rate, and adjustment impact factor are comprehensively calculated and weighted to obtain the subtask priority. Subtasks with higher priority are given priority to obtain resources. Based on the priority, a subtask list is generated, including dimension ID, data range, operation type, and priority label.

[0117] Conduct cross-agent collaborative computing and conflict arbitration. Each agent processes subtasks in parallel and transmits data through a real-time communication protocol. The secondary analysis results of the parsing agent are pushed to the computing agent in real time, with extremely low latency. If the output results of two computing agents conflict, a three-dimensional confidence score is used, including semantic matching and domain consistency. The weight of the results of each conflicting computing agent is obtained through weighted calculation, and the result of the computing agent with the highest weight is taken as the valid result.

[0118] When there is a sudden increase in agent load, the feedback agent automatically migrates some tasks to similar low-load agents based on resource pool data to achieve dynamic balance of computing power;

[0119] Perform multi-level result aggregation. In the basic aggregation phase, the subtask results of each computing agent are reorganized according to the intent vector dimension. The consistency is verified through the hash difference rate. In the enhanced aggregation phase, combined with the adjustment quality score, higher weights are assigned to dimensions whose user adjustment times reach the adjustment threshold, increasing their proportion in the results. After the aggregation is complete, the aggregation results are converted into a user-friendly format, and the mapping rules are associated with the user profile to generate multi-level aggregation results, including raw data and visual charts.

[0120] Collect task completion data for each agent and calculate the task completion rate. Based on the difference between actual computing power consumption and theoretical demand, the difference between the base coefficient and this difference ratio, and the task completion rate, a weighted calculation is performed to obtain the agent performance score. A performance threshold is then set to determine whether to optimize the agent.

[0121] Once the agent's performance score is lower than the performance threshold, parameter iteration is triggered. The parsing agent updates the attention mask weight to improve the parsing accuracy, the computing agent optimizes the task splitting granularity to balance the load, and the feedback agent corrects the prediction threshold to respond to adjustments faster, thereby generating the agent's iterative parameters.

[0122] Specifically, the steps of feedback verification 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 match between the results and the user's initial needs. Once the feedback similarity is less than the semantic matching threshold, it is marked as a semantic deviation.

[0124] Adjustment consistency is quantified by the degree of deviation between the actual difference rate and the expected difference rate, including: presetting the expected difference rate based on the adjustment influencing factor, obtaining the difference rate of the results before and after the adjustment based on the overlap between the adjusted result and the original result, comparing the actual difference rate of the results before and after the adjustment with the expected difference rate, and calculating the adjustment consistency in combination with the basic coefficient. Once the adjustment consistency is less than the consistency threshold, it is marked as adjustment failure;

[0125] Extract the historical features of the user's last three screening results and 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 a historical deviation.

[0126] The three-level verification results are combined to generate a comprehensive verification score through weighted calculation. The feedback similarity accounts for the highest proportion in the weight distribution, followed by the adjustment consistency and matching ratio. Once the comprehensive verification score falls below the first verification threshold, the result is judged to be abnormal and the abnormal type is marked.

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

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

[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 plan is immediately called, the relevant subtasks are re-executed, and the secondary results are pushed after correction. When the comprehensive verification score is between the second verification threshold and the first verification threshold or the feedback validity is between the first feedback threshold and the second feedback threshold, module iterative operations are performed. For semantic deviations, the dual-channel semantic encoding model is updated, the training weight of abnormal cases is increased, and the incremental calculation formula is optimized for adjustment failures. Otherwise, a strategy upgrade is performed, and the agent collaboration topology is fine-tuned based on implicit user feedback.

[0130] The optimization effect is evaluated by comparing the corrected results with user feedback, and the optimization process is iterated continuously until the preset standards are met.

[0131] We aggregate all validation data monthly, including comprehensive validation scores, feedback effectiveness, and correction accuracy, to establish a performance baseline. Evaluation metrics cover three dimensions: stability, adaptability, and efficiency. The stability metric is quantified by the monthly volatility of the proportion of abnormal results. The adaptability metric calculates the average feedback effectiveness of each group by user profile. The efficiency metric focuses on the average time and resource consumption for emergency corrections. By integrating these three metrics, we generate a long-term performance score, with stability and adaptability given higher weights, followed by efficiency.

[0132] If the long-term performance score is lower than the baseline threshold for three consecutive months, architecture-level optimization will be triggered, and the reinforcement learning state space will be reconstructed to improve the adaptability of resource scheduling, ultimately achieving continuous improvement in system performance.

[0133] Example 2:

[0134] See also Figure 4 ,Another embodiment provided by the present invention: a digital employee collaborative 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. It receives concurrent requests from multiple users and assigns unique IDs to independent memory partitions. It uses user profiles to generate adjustment factors, performs weighted filtering on the request text vector, and generates an attention mask and a dedicated stop word list. It initiates dual-channel semantic encoding to generate an initial semantic vector. It uses semantic conflict detection to identify conflicting requests and generate interpretation rules. It then performs three-level intent calibration to generate an intent vector and a three-dimensional confidence score.

[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 requirement level, calculate the urgency and priority based on the user's real-time operation behavior and decision-making speed preference, sort the requests with the same computing power requirement level, build a dual-objective reinforcement learning model to train the scheduling strategy, and obtain the real-time scheduling strategy. Based on the real-time scheduling strategy, the module allocates corresponding computing nodes and computing power proportions to requests with different computing power requirement levels, predicts the user's adjustment intentions, reserves spare resources, and dynamically balances the resource load.

[0137] The dynamic adjustment module is used to obtain and classify user adjustment instructions, calculate adjustment impact factors, generate incremental computing plans for different adjustment instruction types, allocate resources to adjustment instructions through priority scheduling, downgrade the original requested resources when necessary, synchronize state snapshots before and after the adjustment, ensure consistent 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 multi-agent roles and generate collaborative topology, break down subtasks and bind priorities, generate multi-level aggregation results through cross-agent collaborative calculation and conflict arbitration, and perform multi-dimensional verification on the results. It collects user feedback and triggers emergency corrections, module iterations or strategy upgrades, regularly aggregates data to update performance baselines, and continuously optimizes system performance to improve screening accuracy and efficiency.

[0139] Working principle and effect:

[0140] After receiving requests from multiple users, semantic isolation is achieved through an independent context cache, and the text vector is optimized by generating adjustment factors based on user portraits, and the initial semantic vector is generated through dual-channel semantic encoding; conflicting requests are identified through semantic conflict detection, and exclusive interpretation rules and intent vectors are generated; the computing power requirements are quantified based on the complexity of the request, and the urgency and priority are generated based on user behavior, and resources are dynamically allocated through reinforcement learning strategies, and computing power is reserved for prediction and adjustment intentions; when users make adjustments, instructions are captured and classified to generate incremental computing solutions, and efficient response and multi-node consistency are ensured through resource priority scheduling and status synchronization; multiple agents work together according to their roles, disassemble tasks and dynamically balance the load, and finally optimize the system through multi-dimensional verification and feedback iteration, effectively avoiding semantic cross-confusion, improving resource allocation accuracy and adjustment response speed, ensuring the accuracy and adaptability of collaborative screening results, significantly reducing user operating costs, and meeting efficient screening 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 embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart 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 collaborative screening method, characterized in that: include: Receive concurrent requests from multiple users, perform semantic isolation, generate adjustment factors based on user profile features, generate initial semantic vectors through dual-channel semantic encoding, and perform semantic conflict detection; Collect request characteristics in real time, calculate the computing power requirements of different requests, and allocate resources; The front-end obtains user adjustment instructions, identifies demand differences, generates incremental calculation plans, and performs state synchronization control; Classify digital employees into types, conduct collaborative screening, and conduct feedback verification.

2. The digital employee collaborative screening method according to claim 1, characterized in that: The steps of semantic isolation include: Assign a unique ID to each request, build a context cache, and simultaneously adopt the LRU elimination strategy; Call the user portrait API to obtain the three-dimensional feature vector, obtain the semantic adjustment coefficient through weighted summation, and calculate the adjustment factor; Constructing a basic text vector and multiplying it by the adjustment factor to generate a weighted text vector; Based on preset filtering rules, the weighted text vector is disassembled to remove semantic dimensions that do not meet the filtering rules; Performing a softmax function on the filtered weighted text vector to generate an attention mask, extracting decision mode features from the context buffer, and generating a dedicated stop word list; Start dual-channel semantic coding to generate a surface vector and a domain entity vector, and fuse the surface vector and the domain entity vector through residual connection to generate an initial semantic vector.

3. The digital employee collaborative screening method according to claim 2, 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 directly performed; For conflicting requests, we encode the logical structure, calculate the nesting depth of modifiers, extract and combine the numerical range boundary features, and generate the structural feature fingerprint; Call the decision tree generation algorithm, with the logical operator sequence as the root node, the modifier nesting depth as the intermediate node, and the user's historical preference as the leaf node to generate the screening logic; Perform three-level intent calibration to generate a three-dimensional confidence score, including semantic matching, domain consistency, and historical fit, and obtain a comprehensive confidence score through weighted calculation.

4. The digital employee collaborative screening method according to claim 3, characterized in that: The steps in resource allocation include: Semantic depth levels are divided based on the breadth of associations between cross-domain entities 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 scope level is divided by the number of nodes. The computing power requirement level is generated through weighted calculation. Extract behavioral features from real-time user operation logs and quantify them into sensitivity. Then, through weighted calculation, we obtain the user's operation behavior sensitivity. Combined with the decision-making speed preference features in the user profile, we calculate the urgency of each request. Based on the urgency and the comprehensive confidence, the priority of each request is obtained through weighted calculation, and requests with the same computing power requirement level are secondary sorted based on the priority to generate a request queue.

5. The digital employee collaborative screening method according to claim 4, characterized in that: The resource allocation steps also include: Collect multi-dimensional features in real time, construct a multi-dimensional feature state space, and design a dual-objective reward function; Through training with reinforcement learning algorithms, resource allocation is adjusted under different conditions to generate real-time scheduling strategies; Perform initial resource allocation according to the real-time scheduling strategy, match corresponding computing nodes to requests with different computing power requirements, and allocate corresponding computing power proportions according to the priority of the requests; Obtain the user's interaction behavior with the screening interface to generate a real-time behavior comprehensive score. Based on the user's past similar screening operation records, generate a historical habit comprehensive score, and perform 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 by combining the basic coefficient and the computing power standard, and temporarily allocated from the current user's non-urgent request; When resource load fluctuations are detected, resource allocation is dynamically adjusted.

6. The digital employee collaborative screening method according to claim 5, characterized in that: Steps to identify requirements gaps include: Obtain user interaction behaviors on the screening interface, capture dynamic adjustment instructions, and classify user adjustment instructions using the instruction classification model; Count the number of newly added screening dimensions in this adjustment, calculate the accuracy change range and target correlation, and generate the adjustment impact factor through weighted fusion; Generate incremental computing solutions for different adjustment instruction types, including: generating a semantic relevance matrix based on dual-channel semantic encoding, calculating incremental coefficients and incremental computing volume, and scheduling idle nodes in the reserved resource pool; performing differentiated processing on precision modification instructions based on computing power requirements; re-parsing new targets for target replacement instructions, reusing the original computing node's basic resources, and allocating idle computing power; Obtain the adjustment instruction priority and use the token bucket mechanism to control the access frequency of concurrent adjustments. Only instructions holding tokens are allowed to enter the resource scheduling queue to generate resource preemption results.

7. The digital employee collaborative screening method according to claim 6, characterized in that: The steps of state synchronization include: Before and after the adjustment, a state snapshot is generated. The state snapshot is backed up by a distributed storage system, associated with the user ID and the request ID, and synchronized to all related nodes through a distributed protocol. For the intermediate results of the incremental calculation, extract the result hash values ​​before and after the 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, and set a difference threshold to perform a difference hash comparison; If the hash difference rate is less than the difference threshold, the incremental results are directly merged; otherwise, the sub-dataset corresponding to the difference is located and a local recalculation is triggered; Provide feedback on adjustment results and model iteration, collect user satisfaction evaluations on the adjustment results, evaluate the adjustment effect based on the compatibility between the adjustment results and the original screening logic, and update the intent model and scheduling model parameters through online learning.

8. The digital employee collaborative screening method according to claim 7, characterized in that: The steps of collaborative screening include: Based on the complexity of the intent vector and the level of computing power required, the agents are divided into different roles, the matching degree of the agents is calculated, and the agent collaboration topology is generated; Decompose the screening task into three levels, calculate the subtask priority, and generate a subtask list; Each agent processes subtasks in parallel. When the output results of two computing agents conflict, the weights of the results of the conflicting computing agents are calculated, and the result of the computing agent with the highest weight is taken as the valid result. Perform multi-level result aggregation to generate multi-level aggregation results; Calculate the task completion rate based on the task completion data, obtain the agent performance score through weighted calculation, and set the performance threshold; When the agent performance score is less than the performance threshold, parameter iteration is triggered.

9. The digital employee collaborative screening method according to claim 8, characterized in that: The steps of feedback verification include: Calculate feedback similarity, adjustment consistency, and matching ratio, and filter out semantic deviations, adjustment failures, and historical deviations; The three-level verification results are combined to generate a comprehensive verification score through weighted calculation. Once the comprehensive verification score is lower than the first verification threshold, the result is determined to be abnormal and the abnormality type is marked; Setting a second verification threshold, and for abnormal results where the comprehensive verification score is lower than the second verification threshold, adopting a deep feedback mode; otherwise, adopting a light feedback mode; Based on the comprehensive verification score and feedback data, three-level optimization is triggered, and the full 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, architecture-level optimization is triggered.

10. A digital employee collaborative screening system, which is used to implement the digital employee collaborative screening method according to any one of claims 1 to 9, characterized in that: include: Parsing module, allocation module, dynamic adjustment module and feedback module; The parsing module is used to isolate contextual data of 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, calculate the urgency and priority based on the user's real-time operation behavior and decision-making speed preference, and build a real-time scheduling strategy; The dynamic adjustment module is used to obtain and classify user adjustment instructions, generate incremental calculation plans, and allocate resources to the adjustment instructions through priority scheduling; The feedback module is used to assign multi-agent roles and generate a collaborative topology, decompose subtasks and bind priorities, generate multi-level aggregation results through cross-agent collaborative calculation and conflict arbitration, and perform multi-dimensional verification on the results.

Images