Method for realizing lossless switching of agent task between basic large models

Abstract

The invention discloses a method for realizing lossless switching of agent tasks between basic large models, and relates to the technical field of artificial intelligence. Then analyzing and extracting information from the agent tasks of which the hierarchical categories are determined through an associated metadata template, and carrying out standardization processing to obtain a plurality of standardized agent tasks; processing the standardized agent tasks by a pre-constructed basic large model, recording a processing condition, evaluating and optimizing a basic large model sequence; when an agent task to be processed is processed, the state is processed and monitored according to the optimized basic large model sequence associated with the hierarchical category of the agent task, and the lossless processing flow of the agent task is realized in a mode of lossless switching of the model in the sequence when the state is abnormal.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically, it relates to a method for achieving lossless switching of intelligent agent tasks between basic large models. Background Technology

[0002] With the rapid development of artificial intelligence technology, the widespread application of intelligent agents in various fields has become increasingly important; the tasks that intelligent agents need to handle are diverse, and different tasks vary significantly in complexity, resource consumption and user interaction frequency.

[0003] Existing intelligent agent task processing technologies mainly rely on a single basic large model to handle all tasks. However, a single model is difficult to adapt to the diverse characteristics of tasks, resulting in low task completion efficiency. Secondly, existing technologies lack a lossless task switching mechanism, which can easily lead to task interruption when the model is overloaded or malfunctions, affecting user experience. Furthermore, existing standardized processes are insufficient in terms of semantic alignment and data conversion, and cannot effectively handle semantic differences and data format inconsistencies across domains. Finally, existing model evaluations are mostly based on a single performance indicator, lacking an evaluation system that comprehensively considers multiple dimensions of performance such as processing progress and recall. To address the above-mentioned problems, this application proposes a method for achieving lossless switching between basic large models for intelligent agent tasks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for seamless switching of agent tasks between basic large models, solving the problems of low task matching leading to low completion efficiency, easy task interruption, and difficulty in quick recovery in existing technologies.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for achieving lossless switching of agent tasks between basic large models, comprising the following steps: Step 1: Extract several agent tasks from the cloud database, classify the extracted agent tasks hierarchically, identify the hierarchical categories of the agent tasks, and display them. Step 2: Based on the agent task determined within any given hierarchical category, extract the data format, dimensional constraints, and semantic description of the agent task using the metadata template associated with the corresponding hierarchical category; Based on the metadata template associated with the determined hierarchical category of the agent task, the agent task is standardized to obtain a standardized agent task. Repeat the above steps to obtain several standardized intelligent agent tasks; Step 3: For the identified standardized agent tasks, each task is assigned to a number of pre-built basic large models for processing. The processing progress, response speed, and recall rate of each basic large model for different standardized agent tasks are recorded in real time, and the optimal basic large model sequence associated with different levels of categories is evaluated based on this. Then, extract the agent task to be processed and lock the hierarchical category of the agent task and the associated standardized agent task. The standardized agent task is then handed over to the preferred basic large model sequence associated with the hierarchical category of the agent task for processing, and the processing status is monitored in real time. If the processing status is normal, no processing is performed; If the processing state is abnormal, a lossless switch is performed in the preferred basic large model sequence.

[0006] As a further aspect of the present invention, the specific method for locking and displaying the hierarchical category of the intelligent agent task in step one is as follows: The total number of extracted agent tasks is denoted as... And record them as the agent task sequence according to the extraction order. ; extract Any intelligent agent task ,in This is a counting index, with values ​​ranging from 1 to... ; Get Task complexity Resource consumption User interaction frequency Combining The associated three-dimensional feature vector Calculate the three-dimensional feature vector The complexity of the agent task The associated complexity; Determined using the above method The three-dimensional feature vectors of all agent tasks are collected and summarized into a set of three-dimensional feature vectors. Calculate the complexity separately; The set of three-dimensional feature vectors is sorted in ascending order of complexity to obtain a sequence of three-dimensional feature vectors. Then, the total number of labels preset by the operator is used... Hierarchical classification is performed on the three-dimensional feature vector sequence to determine... There are 1 hierarchical category, denoted as the hierarchical category sequence. .

[0007] As a further aspect of the present invention, the specific method for standardizing the agent task in step two to obtain the standardized agent task is as follows: Extract the set of metadata templates preset by the operator. ,in, Include Individual metadata template; Sure Any metadata template Where m is a counting index, and The value range is 1 to , for The triplet in the middle, For data format, For dimensional constraints, For semantic specifications; Will and Perform mapping and binding, that is correspond ; Any level category in Zone 1 is intelligent agent tasks Using hierarchical categories Associated metadata template For agent tasks Standardization is performed to obtain standardized intelligent agent tasks. .

[0008] As a further aspect of the present invention, the specific method of standardizing the agent task to obtain a standardized agent task also includes: extract The data format, and Data format in The process involves matching; if a match is found, the process proceeds to the next step; otherwise, it fails. Convert the data format to the data format If the data format is correct, proceed to the next step; Analysis Dimensional constraints, and Dimensional constraints Perform verification and judgment Do the dimensional constraints conform to the metadata template? Dimensional constraints If it meets the requirements, proceed to the next step; if it does not meet the requirements, and Dimensional constraints are higher than dimensional constraints Then, through dimensionality reduction operations, Adjust the dimensional constraints to conform to the dimensional constraints. Similarly, if Dimensional constraints Then, through dimensional upgrading operations, Adjust the dimensional constraints to conform to the dimensional constraints. And proceed to the next step; extract Semantic description in, and through metadata template Chinese semantic specification For agent tasks Semantic alignment is performed on the semantic descriptions in the text, and generation is based on natural language processing techniques. The associated semantically compliant Semantic description.

[0009] As a further aspect of the present invention, in step two, according to obtaining the intelligent agent task... The associated standardized intelligent agent tasks The method is to process sequentially All agent tasks in the process, A standardized agent task is generated, and the agent task sequence is sorted according to its order to obtain the standardized agent task sequence. .

[0010] As a further aspect of the present invention, the specific method for evaluating the preferred basic large model sequence associated with different hierarchical categories in step three is as follows: Obtain the basic large model prepared by the operator, extract the total number of basic large models, and denote it as . Sort them according to the order of acquisition to obtain the basic large model sequence. ; From standardized intelligent agent task sequences Extract any one of the hierarchical categories as intelligent agent tasks The associated standardized intelligent agent tasks ,copy Divide and distribute to each Processing a basic large model; Extract any one of the basic large models from the basic large model sequence In processing Post-processing progress Response speed Recall rate ,in, This is a counting index, with values ​​ranging from 1 to... ; Obtain the calculation weights associated with the processing progress, response speed, and recall rate preset by the operator. ,in All are greater than 0, and ; Standardized intelligent agent tasks Processing progress Response speed Recall rate Normalization is performed to obtain the normalization processing progress. Normalized response speed Normalized recall rate and calculate deal with The overall score is recorded as ; Repeat the above steps to obtain Each basic large model handles its own processing. The overall score is determined and ranked according to the order of the basic large model sequence. The overall score is recorded as follows: The associated comprehensive score sequence ; extract The standardized agent tasks associated with all agent tasks in the process total One, get The comprehensive score sequence associated with each standardized intelligent agent task is obtained. A comprehensive scoring sequence, among which Preset values ​​for operators; from The first basic large model in Begin, summarize deal with The comprehensive score obtained from the standardized agent tasks associated with all agent tasks in the process is denoted as the hierarchical comprehensive score. And so on, to obtain Each of the basic large models is related to The comprehensive score at each level is denoted as the comprehensive score sequence at each level. ; Sort the hierarchical comprehensive score sequences in descending order, and then sort the basic large model sequences according to the order of the descending-sorted hierarchical comprehensive score sequences to obtain the sorted basic large model sequences. Recorded as a hierarchical category The associated preferred basic large model sequence; Then, in the hierarchical category sequence The above processing is performed on each hierarchical category to obtain... The preferred basic large model sequence associated with each hierarchical category.

[0011] As a further aspect of the present invention, in step three, the agent task to be processed is extracted and denoted as... ; locking The associated three-dimensional feature vectors are determined by calculating the complexity of the three-dimensional feature vectors. The level category it belongs to, and extract it. The preferred basic large model sequence associated with the hierarchical category; use Metadata templates associated with the hierarchical category Standardization is performed to obtain standardized intelligent agent tasks. ; Will Mix to The first basic large model in the associated preferred basic large model sequence is processed; If the first basic large model is processed Once completed, it is considered a task for processing the intelligent agent. The status is normal, no action is required; If the first basic large model is processed If an anomaly occurs, it is considered a task for handling the intelligent agent. The state is abnormal, and a lossless switch is performed.

[0012] As a further aspect of the present invention, the specific method for performing lossless switching in step three is as follows: When the first basic large model is determined to handle standardized intelligent agent tasks When the processing status is abnormal, re-extract. The next basic large model in the preferred basic large model sequence associated with the hierarchical category; agent tasks The associated standardized intelligent agent tasks The task is then handed over to the next basic large model for processing, and the processing status is monitored in real time until the agent task is processed. The processing status is normal.

[0013] The beneficial effects of this invention are: This application constructs a standardized processing scheme for intelligent agent tasks, thereby obtaining a standardized processing method for intelligent agent tasks. This enables lossless switching of intelligent agent tasks between different basic large models, improves the flexibility and reliability of task processing, effectively solves the problems of low task completion efficiency and lack of fault response strategies in the prior art, and enhances the stability of intelligent agent tasks during processing. This application ensures the scientific and rational nature of agent task classification through precise three-dimensional feature vector extraction and hierarchical classification. Based on hierarchical classification, it utilizes metadata templates bound to hierarchical category mappings during the standardization stage of agent tasks to uniformly process the data format, dimensional constraints, and semantic specifications of agent tasks, effectively solving the adaptation problem caused by the diversity of agent tasks. Secondly, through data format matching and conversion, dimensional constraint verification and adjustment, and semantic alignment generation, agent tasks are presented in a standardized form, improving the universality and processing efficiency of agent tasks across different models. This application comprehensively evaluates and selects the optimal basic large model sequence, fully considering multiple dimensions such as processing progress, response speed, and recall rate. By combining normalization processing and weight calculation, it accurately selects the most suitable model sequence for each level category, optimizes agent task allocation, and improves processing efficiency and quality. Furthermore, it monitors the processing status in real time during task processing. If any anomaly occurs, it immediately and seamlessly switches to the next candidate model from the optimal basic large model sequence to continue processing. The entire process is seamless, ensuring uninterrupted progress of agent tasks, greatly enhancing the system's stability and reliability, effectively avoiding the risks and losses caused by agent task processing interruptions, and significantly improving the continuity and success rate of agent task processing. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a flowchart illustrating the method described in Embodiment 3 of the present invention; Figure 3 This is a flowchart illustrating the method described in Embodiment 4 of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1 A method for achieving lossless switching of agent tasks between basic large models, such as Figure 1 As shown, it specifically includes the following: First, before processing the agent task, the following steps are required: classify the agent task hierarchically, and then determine the metadata template for standardizing the agent task for different hierarchical categories. The metadata template contains three parameters for describing the characteristics of the agent task: data format, dimensional constraints, and semantic specifications. Standardizing the agent task with these three parameters will result in a standardized agent task. The standardized agent task can be adapted to several different basic models. That is, the basic models can interpret the original agent task through the standardized agent, and can ensure that different basic models have the same interpretation effect on the agent task. In addition to the above, by analyzing the data format, dimensional constraints, and semantic specifications in the agent task, we can further obtain the three-dimensional feature vector of the corresponding agent task, and determine several hierarchical categories by calculating the complexity of the three-dimensional feature vector (by analyzing the complexity of the three-dimensional feature vectors corresponding to several agent tasks, the amount of data of the several agent tasks is sufficient to cover the complexity range of the three-dimensional feature vectors associated with all agent tasks).

[0018] Then, several basic large models determined by the operators are used to process several standardized agent tasks in any given hierarchical category (each basic large model must process all agent tasks in any given hierarchical category). After processing, the processing progress, response speed, and recall rate of all basic large models in processing agent tasks are obtained. Based on the processing progress, response speed, and recall rate of the basic large models in processing agent tasks, the preferred basic large model sequence associated with the corresponding hierarchical category is evaluated. The first basic large model in the preferred basic large model sequence has the best effect in processing agent tasks in the corresponding hierarchical category, and the effect of subsequent basic large models in the preferred basic large model sequence in processing agent tasks in the corresponding hierarchical category increases sequentially. Similarly, the above method can be used to determine the preferred basic large model sequence associated with each of the several hierarchical categories; When a subsequent agent task that needs to be processed is received, the hierarchical category of the received agent task can be determined, and the preferred basic large model sequence associated with the hierarchical category of the agent task to be processed can be obtained. The agent task can be assigned to the first basic large model in the preferred basic large model sequence (the basic large model with the best performance in processing agent tasks of this hierarchical category) for processing, and the processing status of the first basic large model in processing the standardized agent task associated with the agent task can be monitored in real time. If the monitored processing status is normal, it means that the first basic large model can complete the agent's task without performing any other processing. If the monitored processing status is abnormal, continue to select a basic large model that is adjacent to the first basic large model from the preferred basic large model sequence, that is, the second basic large model; The system monitors the processing status of the second basic model in real time when handling standardized agent tasks, and repeats the above steps until the agent task is completed.

[0019] Example 2 This embodiment discloses a method for determining the hierarchical category of an agent's task, specifically including the following: This method requires pre-retrieving several processed agent tasks from cloud data. The amount of data in the extracted agent tasks must meet certain requirements, and the extracted agent tasks must be able to cover agent tasks under any circumstances. This is to ensure the comprehensiveness and accuracy of subsequent operations such as analyzing, classifying, standardizing, and determining the optimal basic large model sequence for agent tasks. The total number of extracted agent tasks is denoted as... , will the The tasks of each agent are sequentially extracted and denoted as an agent task sequence, as follows: ; Extract any one agent task from the agent task sequence and denote it as... ,in This is a counting index, with values ​​ranging from 1 to... This embodiment uses intelligent agent tasks. As an example, the remaining agent tasks are handled in the same way. Process it using the same method; Obtain agent tasks using log analysis tools The execution time and steps determine the agent's task. The task complexity (the calculation of complexity is based on existing technology and will not be elaborated upon in this solution) is denoted as... ; Then, use performance monitoring tools to track the agent's tasks in real time. During the processing Memory, disk Resources, and through and The relevant data tracked by the performance monitoring tool is indexed to obtain the agent task. Resource consumption, denoted as ; Recording agent tasks Calls during processing The interface is defined, and API interfaces requiring user participation are extracted. The total number of interfaces is used as the agent's task. The frequency of user interaction, denoted as ; Then, by combining operations, the task complexity is reduced. Resource consumption User interaction frequency Defined as an agent task The associated three-dimensional feature vector is denoted as: ; The three-dimensional feature vector Complexity and task complexity Resource consumption User interaction frequency All are directly proportional, and the task complexity is also calibrated. Resource consumption User interaction frequency Calculating three-dimensional feature vectors When the complexity is equal, the weights are calculated equally; By repeating the above steps, the agent's task sequence can be determined. The three-dimensional feature vectors of all agent tasks are collected and summarized into a set of three-dimensional feature vectors, represented as: ; The determined set of three-dimensional feature vectors is then sorted in ascending order of complexity, and the sorted result is recorded as a sequence of three-dimensional feature vectors. Then, based on the total number of labels preset by the operator. Hierarchical classification of the three-dimensional feature vector sequence, that is, hierarchical classification of the agent tasks associated with the three-dimensional feature vector sequence, can ultimately determine... Each level of category is determined according to the order of the hierarchical classification operations. Each hierarchical category is denoted as a hierarchical category sequence, represented as: ; The purpose of this embodiment is to systematically manage heterogeneous intelligent agent tasks by constructing a multi-dimensional feature quantification model, thereby achieving optimized resource allocation and improved agent performance in complex task scenarios. Its core lies in establishing an equally weighted three-dimensional feature vector model by collecting three key indicators from historical tasks in the cloud: execution time, resource consumption, and interaction frequency. This model quantifies the overall complexity of tasks and, based on a dynamic hierarchical classification mechanism with preset labels, reconstructs the original task sequence into a hierarchical system with a clear complexity gradient. This effectively overcomes the limitations of single-dimensional evaluation in traditional task scheduling, solves the problem of resource allocation imbalance caused by differences in task attributes, and forms an interpretable hierarchical mapping relationship of task complexity through total order permutation and adaptive clustering of the feature space. This provides a standardized decision-making basis for accurate routing of subsequent intelligent agent tasks, selection of basic large-scale models, and elastic scheduling of computing resources.

[0020] Example 3 This embodiment, based on embodiment 1, further discloses a method for obtaining standardized agent tasks by standardizing agent tasks, such as... Figure 2 As shown, it specifically includes the following: This embodiment obtains standardized agent tasks by standardizing the agent tasks, so as to ensure that the agent tasks can be seamlessly switched between different basic models in the form of standardized agent tasks, while ensuring that different basic models have the same parsing content for the same agent task.

[0021] Based on the content described in Example 2, a hierarchical category sequence can be obtained. Different hierarchical categories correspond to agent tasks of different complexities, therefore different metadata templates are needed to standardize agent tasks in different hierarchical categories. Extract the set of metadata templates preset by the operator. Metadata template collection It contains hierarchical category sequences Metadata templates with the same number of mid-level categories, i.e., a set of metadata templates. have Individual data template, represented as: ; Extract metadata template collection Any metadata template Example processing is performed on the metadata template. ,in, This is a counting index, with values ​​ranging from 1 to... Metadata template It represents any one of the metadata templates in the set; The metadata template Composed of triplets, for The data format in the middle, for Dimensional constraints in for Semantic specifications in; The data format clarifies the form in which the agent task exists during storage, transmission, and display, ensuring data accuracy and consistency; the dimensional constraints limit the agent task in terms of dimensions; the semantic specifications ensure that the semantics expressed by the agent task are clear and accurate, and have a clear meaning in a specific data environment. The metadata template set obtained in this embodiment The hierarchical category sequence obtained in Example 2 Perform mapping and binding, and lock any level category. Corresponding to any unique metadata template And any metadata template Applicable to any unique hierarchical category ; Next, use the metadata template. Metadata template The associated hierarchical category is The tasks of the agents within the system are standardized to determine any given agent task. The hierarchical category is the hierarchical category ; The standardization process mainly involves three steps, as described below: Step 1: From the defined agent task Extract the original data format (e.g., CSV, JSON, XML, etc.) and then process the intelligent agent task. The original data format and metadata template in Data format in Perform a match; If the match fails, the agent task will be... Convert the raw data format in the file into a metadata template. The data format is specified, and the next step is performed; If the match is successful, then the agent task is not performed. The original data format is not processed in any way, and you can proceed directly to the next step.

[0022] Step 2: Analyze the intelligent agent Dimensional constraints in the data, and simultaneously extract metadata templates. Dimensional constraints What needs to be explained is the dimensional constraint. This refers to a range of constraints within a single dimension, not a single dimension. The dimensional constraints are determined by the operator based on the actual situation; agent tasks Dimensional constraints and metadata templates The dimensional constraints defined in the code are compared to determine the agent's task. Do the dimensional constraints conform to the metadata template? Dimensional constraints defined in the document; If the agent task The dimensional constraints do not conform to the metadata template If the dimensional constraints defined in the code are not considered, then the agent's task will be further evaluated. Are the dimensional constraints higher than the metadata template? Dimensional constraints in the context of agent tasks Dimensional constraints are higher than those of metadata templates. The dimensionality constraint in the equation is addressed by reducing the dimensionality of the agent's task through dimensionality reduction operations. Adjust the dimensional constraints to conform to the metadata template. Dimensional constraints in the context; if the agent's task Dimensional constraints are lower than those of metadata templates The dimensional constraints in the equation are addressed by using dimensionality-up operations to transform the agent's task. Adjust the dimensional constraints to conform to the metadata template. The dimensional constraints in the data will affect the agent's task. Adjust the dimensional constraints to conform to the metadata template. After defining the dimensional constraints, proceed to the next step; If the agent task The dimensional constraints conform to the metadata template The dimensional constraints defined in the code do not apply to the agent's task. Do not process the dimensional constraints and proceed directly to the next step.

[0023] Step 3: Implement the intelligent agent task using vocabulary matching, semantic structure matching, and natural language processing technologies. Semantic descriptions and metadata templates in Semantic alignment is performed according to the semantic specifications, and intelligent agent tasks are generated. The associated conforming metadata template The standard semantic description of the semantic specification.

[0024] By summarizing the processing results from Step 1 to Step 3, we obtain the agent task. The associated standardized intelligent agent tasks According to the task of processing intelligent agents The method for agent task sequences All agent tasks are processed to obtain standardized agent tasks associated with all agent tasks. It's important to note that agent tasks belonging to different hierarchical categories are processed using the metadata template associated with their respective hierarchical category, rather than using the agent task itself. Process the associated metadata template; Then, the obtained standardized agent tasks are arranged according to the agent task sequence. Sort the data in the correct order to obtain a standardized sequence of agent tasks, denoted as: .

[0025] This embodiment describes a method for standardizing agent tasks, aiming to ensure that agent tasks can be seamlessly switched between different basic large models and that different large models can consistently resolve the same task. The core lies in selecting an appropriate metadata template based on the hierarchical category sequence. The template includes three parts: data format, dimensional constraints, and semantic specifications. By extracting the original data format, dimensional constraints, and semantic description of the agent task, it is matched, adjusted, and aligned with the corresponding template to finally generate a standardized agent task. Key steps include data format matching and conversion, dimensional constraint comparison and adjustment (dimensionality reduction or expansion), semantic alignment, and standard semantic description generation. The purpose of this embodiment is to ensure the compatibility of agent tasks in different models and environments.

[0026] Example 4 This embodiment, based on Embodiments 1, 2, and 3, further discloses a method for processing standardized agent tasks using a preferred large model and achieving lossless switching, such as... Figure 3 As shown, it specifically includes the following: Before processing the standardized intelligent agent task obtained in Example 3, it is necessary to extract all the basic large models pre-built (prepared) by the operator. The understanding (computation or processing) capabilities of the basic large models are different, and the range of tasks they are good at are also different. Determine the total number of the extracted basic large models, denoted as . Following the order in which the basic large models were extracted, the basic large models were initially sorted (serialized) to obtain a sequence of basic large models, represented as follows: ; Determine any level category The example is the object being processed, and other hierarchical categories are all handled according to the hierarchical category. The processing method is to handle it according to the hierarchy and category. Determine any one located in the standardized agent task sequence Standardized intelligent agent tasks within and standardize intelligent agent tasks copy The copy Standardized intelligent agent task Assigned sequentially to the basic large model sequence Each of the basic large models in the process is processed; Real-time monitoring of any one of the basic large models in the basic large model sequence (The processing method for the remaining basic large models is the same as that for the basic large models.) The processing method will be used (details will not be elaborated in this solution) for processing standardized intelligent agent tasks. Post-processing progress Response speed Recall rate ,in, This is a counting index, with values ​​ranging from 1 to... Processing progress Basic large model Processing standardized intelligent agent tasks The completion percentage (which can be obtained through the logs built into the basic large model or an external monitoring system), and Response speed Basic large model From receiving the standardized intelligent agent task The time interval between the first and subsequent feedback (this can be obtained by examining the logs built into the basic large model for standardized agent tasks) The timestamp of reception and the timestamp of initial feedback can be used to calculate the difference between these two timestamps, which yields the response time and recall rate. Basic large model The proportion of valid outputs in the task results (calculated by analyzing the log files of the basic large model output, counting the number of valid outputs and total outputs, and calculating the recall rate), and .

[0027] The calculation weights associated with processing progress, response speed, and recall rate are obtained based on the operator's preset calculations according to the actual situation. The calculation weight for processing progress is... The weighting of response speed is The weighting for recall rate calculation is: and All are greater than 0. ; Standardized intelligent agent tasks Processing progress Response speed Recall rate Normalization is performed to obtain the normalized processing progress. Normalized response speed Normalized recall rate Among them, regarding response speed Obtain the operator's preset maximum response speed , to maximize response speed After normalization, the normalized maximum response speed is obtained. The normalized maximum response speed ; For the normalized progress after normalization Normalized response speed Normalized recall rate The comprehensive scoring formula is as follows: The basic large model was calculated. Processing standardized intelligent agent tasks Overall rating ; Following the method described above, the same process can be applied to the remaining basic large models, ultimately yielding... Each of the basic large models handles standardized agent tasks. The overall score is then calculated, and the results are processed according to the order of the basic large model sequence. Each of the basic large models handles standardized agent tasks. The comprehensive scores are sorted to obtain the standardized agent task. The associated comprehensive score sequence is represented as follows: .

[0028] Re-extracting standardized intelligent agents task The category of the level The standardized agent tasks associated with all agent tasks in the system (including standardized agent tasks) ), determine the total number as It should be explained that each hierarchical category contains standardized agent tasks. indivual; For hierarchical categories Extracted from A standardized intelligent agent task, repeating the above steps to obtain... The comprehensive score sequence associated with each standardized intelligent agent task is obtained. A comprehensive scoring sequence, and from the basic large model sequence The first basic large model in Let's begin by summarizing the basic large model. In processing hierarchical categories The comprehensive score obtained from the standardized agent tasks associated with all agent tasks in the process is denoted as the hierarchical comprehensive score. and the basic large model sequence The same steps are performed on the remaining basic large models to obtain... A basic large model The corresponding hierarchical comprehensive scores, arranged in the order of the basic large model sequence, yield a hierarchical comprehensive score sequence, represented as follows: ; Next, sort the hierarchical comprehensive score sequence according to the hierarchical comprehensive score values ​​in descending order, and then sort the basic large model sequence according to the order of the sorted hierarchical comprehensive score sequence. The sorted basic large model sequence is represented as follows: And recorded as the hierarchical category. The associated preferred basic large model sequence, wherein the preferred basic large model sequence processes hierarchical categories from the first basic large model to the last basic large model. The performance of standardized intelligent agent tasks decreases from high to low for each type.

[0029] According to processing level category Get the hierarchy category The associated method of selecting the best basic large model sequence for hierarchical category sequences In Continue processing the same way for each level of category, and you will eventually get... The preferred basic large model sequence associated with each hierarchical category.

[0030] When a new agent task is detected that needs to be processed, the corresponding agent task is extracted and recorded as an agent task. ; Locking the agent task using the method described in Example 2 The associated three-dimensional feature vectors are used to calculate the agent's task. The complexity of the associated three-dimensional feature vectors determines the agent task. The hierarchical category it belongs to, and extract the agent's task. The preferred basic large model sequence associated with the hierarchical category; Then based on the agent's task The metadata template associated with the hierarchical category, and the use of this metadata template for agent tasks. After standardization, the agent task is obtained. The associated standardized intelligent agent task is denoted as Then, the defined standardized intelligent agent tasks Assigned to agent tasks The first basic large model in the associated preferred basic large model sequence performs the processing steps, and the processing of the standardized agent task by the first basic large model is monitored in real time. The processing status; If the first basic large model handles standardized agent tasks Completed, and the standardized intelligent agent task has been accomplished. All processing steps in this process are considered as the first basic large model for processing standardized intelligent agent tasks. The processing status is normal, and no processing is required; If the first basic large model handles standardized agent tasks An anomaly occurred; the standardized intelligent agent task was not completed. If all processing steps or processing steps are interrupted, it is considered that the first basic large model is used to process the standardized intelligent agent task. If the processing status is abnormal, then the next candidate basic model in the preferred basic model sequence is extracted again. For the first basic model, the next candidate basic model is the second basic model in the preferred basic model sequence. The standardized agent task corresponding to the agent task is handed over to the next candidate base model for processing, and the processing status is monitored in real time until the processing status of the standardized agent task is normal.

[0031] This embodiment achieves improved task processing efficiency and accuracy by efficiently processing standardized agent tasks and enabling lossless switching through the selection of optimal large models. First, all basic large models are extracted, and their total number and sequence are determined. Then, for standardized agent tasks in each hierarchical category, they are assigned to various basic large models for processing, while key indicators such as processing progress, response speed, and recall rate are monitored in real time. By obtaining weights preset by the operator, these indicators are normalized and a comprehensive score is calculated to determine the optimal order of the basic large models, forming a sequence of optimal basic large models related to the hierarchical category. When processing a new agent task, its associated hierarchical category and metadata template are locked. After standardization processing, the task is assigned to the first basic large model in the optimal sequence for processing, and the processing status is monitored in real time. If an anomaly occurs, the system automatically switches to the next candidate large model, ensuring the stability and continuity of task processing.

[0032] All data in the formulas described above are numerical calculations performed with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0033] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0034] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. A method for achieving lossless switching of agent tasks between basic large models, characterized in that, This method includes the following steps: Step 1: Extract several agent tasks from the cloud database, classify the extracted agent tasks hierarchically, identify the hierarchical categories of the agent tasks, and display them. Step 2: Based on the agent task determined within any given hierarchical category, extract the data format, dimensional constraints, and semantic description of the agent task using the metadata template associated with the corresponding hierarchical category; Based on the metadata template associated with the determined hierarchical category of the agent task, the agent task is standardized to obtain a standardized agent task. Repeat the above steps to obtain several standardized intelligent agent tasks; Step 3: For the identified standardized agent tasks, each task is assigned to a number of pre-built basic large models for processing. The processing progress, response speed, and recall rate of each basic large model for different standardized agent tasks are recorded in real time, and the optimal basic large model sequence associated with different levels of categories is evaluated based on this. Then, extract the agent task to be processed and lock the hierarchical category of the agent task and the associated standardized agent task. The standardized agent task is then handed over to the preferred basic large model sequence associated with the hierarchical category of the agent task for processing, and the processing status is monitored in real time. If the processing status is normal, no processing is performed; If the processing state is abnormal, a lossless switch is performed in the preferred basic large model sequence.

2. The method for achieving lossless switching of agent tasks between basic large models according to claim 1, characterized in that, In step one, the specific method for locking and displaying the hierarchical category of the agent's task is as follows: The total number of extracted agent tasks is denoted as... And record them as the agent task sequence according to the extraction order. ; extract Any intelligent agent task ,in This is a counting index, with values ​​ranging from 1 to... ; Get Task complexity Resource consumption User interaction frequency Combining The associated three-dimensional feature vector Calculate the three-dimensional feature vector The complexity of the agent task The associated complexity; Determined using the above method The three-dimensional feature vectors of all agent tasks are collected and summarized into a set of three-dimensional feature vectors. Calculate the complexity separately; The set of three-dimensional feature vectors is sorted in ascending order of complexity to obtain a sequence of three-dimensional feature vectors. Then, the total number of labels preset by the operator is used... Hierarchical classification is performed on the three-dimensional feature vector sequence to determine... There are 1 hierarchical category, denoted as the hierarchical category sequence. .

3. The method for achieving lossless switching of agent tasks between basic large models according to claim 2, characterized in that, In step two, the standardized agent task is processed to obtain the standardized agent task in the following way: Extract the set of metadata templates preset by the operator. ,in, Include Individual metadata template; Sure Any metadata template Where m is a counting index, and The value range is 1 to , for The triplet in the middle, For data format, For dimensional constraints, For semantic specifications; Will and Perform mapping and binding, that is correspond ; Any level category in Zone 1 is intelligent agent tasks Using hierarchical categories Associated metadata template For agent tasks Standardization is performed to obtain standardized intelligent agent tasks. .

4. The method for achieving lossless switching of agent tasks between basic large models according to claim 3, characterized in that, Standardizing agent tasks to obtain standardized agent tasks also includes the following specific methods: extract The data format, and Data format in The process involves matching; if a match is found, the process proceeds to the next step; otherwise, it fails. Convert the data format to the data format If the data format is correct, proceed to the next step; Analysis Dimensional constraints, and Dimensional constraints Perform verification and judgment Do the dimensional constraints conform to the metadata template? Dimensional constraints If it meets the requirements, proceed to the next step; if it does not meet the requirements, and Dimensional constraints are higher than dimensional constraints Then, through dimensionality reduction operations, Adjust the dimensional constraints to conform to the dimensional constraints. Similarly, if Dimensional constraints Then, through dimensional upgrading operations, Adjust the dimensional constraints to conform to the dimensional constraints. And proceed to the next step; extract Semantic description in, and through metadata template Chinese semantic specification For agent tasks Semantic alignment is performed on the semantic descriptions in the text, and generation is based on natural language processing techniques. The associated semantically compliant Semantic description.

5. The method for achieving lossless switching of agent tasks between basic large models according to claim 4, characterized in that, In step two, according to obtaining the agent task The associated standardized intelligent agent tasks The method is to process sequentially All agent tasks in the process, A standardized agent task is generated, and the agent task sequence is sorted according to its order to obtain the standardized agent task sequence. .

6. The method for achieving lossless switching of agent tasks between basic large models according to claim 1, characterized in that, In step three, the specific method for evaluating the preferred basic large model sequence associated with different hierarchical categories is as follows: Obtain the basic large model prepared by the operator, extract the total number of basic large models, and denote it as . Sort them according to the order of acquisition to obtain the basic large model sequence. ; From standardized intelligent agent task sequences Extract any one of the hierarchical categories as intelligent agent tasks The associated standardized intelligent agent tasks ,copy Divide and distribute to each Processing a basic large model; Extract any one of the basic large models from the basic large model sequence In processing Post-processing progress Response speed Recall rate ,in, This is a counting index, with values ​​ranging from 1 to... ; Obtain the calculation weights associated with the processing progress, response speed, and recall rate preset by the operator. ,in All are greater than 0, and ; Standardized intelligent agent tasks Processing progress Response speed Recall rate Normalization is performed to obtain the normalization processing progress. Normalized response speed Normalized recall rate and calculate deal with The overall score is recorded as ; Repeat the above steps to obtain Each basic large model handles its own processing. The overall score is determined and ranked according to the order of the basic large model sequence. The overall score is recorded as follows: The associated comprehensive score sequence ; extract The standardized agent tasks associated with all agent tasks in the process total One, get The comprehensive score sequence associated with each standardized intelligent agent task is obtained. A comprehensive scoring sequence, among which Preset values ​​for operators; from The first basic large model in Begin, summarize deal with The comprehensive score obtained from the standardized agent tasks associated with all agent tasks in the process is denoted as the hierarchical comprehensive score. And so on, to obtain Each of the basic large models is related to The comprehensive score at each level is denoted as the comprehensive score sequence at each level. ; Sort the hierarchical comprehensive score sequences in descending order, and then sort the basic large model sequences according to the order of the descending-sorted hierarchical comprehensive score sequences to obtain the sorted basic large model sequences. Recorded as a hierarchical category The associated preferred basic large model sequence; Then, in the hierarchical category sequence The above processing is performed on each hierarchical category to obtain... The preferred basic large model sequence associated with each hierarchical category.

7. The method for achieving lossless switching of agent tasks between basic large models according to claim 6, characterized in that, In step three, the agent task to be processed is extracted and denoted as... ; locking The associated three-dimensional feature vectors are determined by calculating the complexity of the three-dimensional feature vectors. The level category it belongs to, and extract it. The preferred basic large model sequence associated with the hierarchical category; use Metadata templates associated with the hierarchical category Standardization is performed to obtain standardized intelligent agent tasks. ; Will Mix to The first basic large model in the associated preferred basic large model sequence is processed; If the first basic large model is processed Once completed, it is considered a task for processing the intelligent agent. The status is normal, no action is required; If the first basic large model is processed If an anomaly occurs, it is considered a task for handling the intelligent agent. The state is abnormal, and a lossless switch is performed.

8. The method for achieving lossless switching of agent tasks between basic large models according to claim 7, characterized in that, In step three, the specific method for performing lossless switching is as follows: When the first basic large model is determined to handle standardized intelligent agent tasks When the processing status is abnormal, re-extract. The next basic large model in the preferred basic large model sequence associated with the hierarchical category; agent tasks The associated standardized intelligent agent tasks The task is then handed over to the next basic large model for processing, and the processing status is monitored in real time until the agent task is processed. The processing status is normal.

Images