A model service method, system, and electronic device

By selecting heterogeneous agents and adjudicating agents within the agent application system, and utilizing input agents and adjudicating agents for model invocation and result adjudication, the problems of limited knowledge coverage and poor cross-domain adaptability in existing agent services are solved, thereby improving the accuracy and cross-domain adaptability of generated content.

CN123114664APending Publication Date: 2026-11-17PURPLE MOUNTAIN LAB
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611307515.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent agent services suffer from limitations in knowledge coverage and are prone to factual errors when handling tasks in specialized domains. Furthermore, they exhibit poor adaptability to complex cross-domain problems, making it difficult to achieve dynamic adjustment and cross-domain adaptation.

Method used

By selecting heterogeneous agents and adjudicating agents within the agent application system, and utilizing input agents and adjudicating agents for model invocation and result adjudication, the agent selection strategy is dynamically adjusted to achieve multi-model collaborative decision-making and result optimization.

Benefits of technology

It improves the accuracy and cross-domain adaptability of generated content, suppresses the knowledge illusion and coverage limitations of single models, and improves the semantic drift problem in the text generation process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114664A_ABST
    Figure CN123114664A_ABST
Patent Text Reader

Abstract

This invention discloses a model service method, system, and electronic device. The method includes: responding to a model invocation request, selecting at least one group of heterogeneous agents and one adjudicating agent within an agent application system based on agent-selected temperature parameters; sending the model invocation request to each of the heterogeneous agents via an input agent and obtaining the processing results of each heterogeneous agent; and generating a model output result through the adjudicating agent according to the model invocation request and the processing results. The technical solution of this invention, based on agent-selected temperature parameters, achieves a dynamic balance between the randomness and reliability of agent selection, enables multi-model collaborative decision-making, suppresses knowledge illusion and coverage limitations of single models, and improves cross-domain task adaptability and generation accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data processing technology, and in particular to a model service method, system and electronic device. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agent services encapsulated with large language models are being widely used in professional fields such as medical diagnosis and legal consultation due to their powerful generation capabilities.

[0003] However, existing technologies suffer from the following problems when handling specialized domain tasks: each task request typically relies on a single agent, resulting in limited coverage of model knowledge and a tendency to generate factual errors (i.e., model illusions). Some multi-agent collaborative solutions employ static configuration of agent combinations and role assignments, making dynamic adjustment difficult based on task requirements and exhibiting poor adaptability to cross-domain complex problems. Furthermore, when facing cross-domain complex problems, a single agent struggles to balance domain specificity and global consistency, compromising the accuracy of generated content. Therefore, designing an agent service system and method capable of achieving multi-agent collaboration and improving the accuracy, anti-illusion capabilities, and cross-domain adaptability of specialized domain generation tasks has become a critical technical challenge urgently needing resolution in this field. Summary of the Invention

[0004] This invention provides a model service method, system, and electronic device to address problems in existing agent services, such as agent vulnerability, poor cross-domain adaptability due to static configuration, and lack of dynamic adjustment mechanisms for model selection.

[0005] According to one aspect of the present invention, a model service method is provided, the method comprising: In response to a model invocation request, at least one set of heterogeneous agents and one adjudicating agent are selected within the agent application system based on the agent's selected temperature parameter. The model invocation request is sent to each of the heterogeneous intelligent agents through the input agent, and the processing results of each of the heterogeneous intelligent agents are obtained. The adjudicating agent generates model output results according to the model invocation request and each of the processing results.

[0006] According to another aspect of the present invention, a model service system is provided, the system comprising: The model selection module is used to select at least one set of heterogeneous agents and one adjudicating agent within the agent application system in response to a model call request, based on the agent selection temperature parameter. The model processing module is used to send the model invocation request to each of the heterogeneous intelligent agents through the input agent, and to obtain the processing results of each of the heterogeneous intelligent agents; The result adjudication module is used to generate model output results by the adjudication agent according to the model call request and each of the processing results.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model service method according to any embodiment of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute a model service method according to any embodiment of the present invention.

[0009] The technical solution of this invention involves selecting a group of heterogeneous agents and a decision agent within the agent application system based on agent selection temperature parameters upon receiving a model invocation request. The model invocation request is transmitted to different heterogeneous agents via an input agent, and the processing results of the heterogeneous agents in response to the request are obtained. The decision agent then adjudicates the processing results and the model invocation request, thereby generating the model output result. This invention dynamically selects heterogeneous agents within the agent application system using agent selection temperature parameters, decoupling the agent selection strategy from the content generation strategy. This achieves a dynamic balance between selection randomness and high reliability, improving the adaptability of the selected group of heterogeneous agents in cross-domain tasks. By using an input agent to send model invocation requests to various heterogeneous agents and obtain the processing results of each agent, multi-model collaborative decision-making can be achieved, suppressing the knowledge illusion and coverage limitations of a single model. By using an adjudicating agent to generate model output results according to the model invocation requests and processing results, the system can perform adjudication functions such as optimization and fusion of multi-model processing results based on task characteristics. This enables the optimal model combination selection and result optimization for professional tasks, achieving the beneficial effect of improving the accuracy of generated content.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a model service method provided according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of a model service method provided according to Embodiment 2 of the present invention; Figure 3 This is a flowchart of a model service method provided according to Embodiment 3 of the present invention; Figure 4 This is a diagram of a hierarchical selection decision mode provided in Embodiment 4 of the present invention; Figure 5 This is a diagram of a professional field fusion adjudication mode provided in Embodiment 4 of the present invention; Figure 6 This is a system architecture diagram of a model service method provided according to Embodiment 4 of the present invention; Figure 7 This is a schematic diagram of a semantic-based model adjudication algorithm provided in Embodiment 4 of the present invention; Figure 8 This is a schematic diagram of a professional domain fusion algorithm provided in Embodiment 4 of the present invention; Figure 9 This is a schematic diagram of a model service system provided according to Embodiment 5 of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device that implements the model service method of Embodiment Six of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. For ease of description, in this invention, "agent" refers to a service instance encapsulating an artificial intelligence model, which has independent interfaces and functions; "model" refers to the core algorithm engine within the agent. When referring to a measurement of the algorithm engine's capabilities, such as a 'professionalism rating,' the evaluation object is the "model" encapsulated by the agent; when referring to objects of interaction, selection, or scheduling in the system architecture, "agent" is used. It can be understood that the selection of an agent is essentially the selection of the model it encapsulates. Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0015] Example 1 Figure 1 This is a flowchart of a model service method provided in Embodiment 1 of the present invention. This embodiment is applicable to scenarios involving dynamic heterogeneous model scheduling and mimicry adjudication for handling task requests in professional domains. This method can be executed by a model service system, which can be implemented in hardware and / or software and can be configured in an electronic device. Figure 1 As shown, the method includes: S110. In response to the model call request, select at least one set of heterogeneous agents and one adjudicating agent within the agent application system based on the agent's selected temperature parameter.

[0016] The model invocation request can include a request issued by an Artificial Intelligence (AI) application system for a specific domain task. In some embodiments, the model invocation request may also include instructions specific to the professional domain. For example, the model invocation request may declare restrictions on professional domains such as code generation, medical diagnosis, and legal consultation. The model invocation request may include code instructions, task-related parameters, etc. The intelligent agent application system can include an AI application system built on an intelligent agent model to perform generation tasks in a specific professional domain. Heterogeneous intelligent agents refer to intelligent agent instances encapsulated with different artificial intelligence models, such as instances encapsulated with models with different architectures or training data. The adjudicating intelligent agent can include an intelligent agent model that makes consistent adjudications on the generation results of multiple heterogeneous intelligent agents. The intelligent agent selection temperature parameter is used to realize the dynamic degree of intelligent agent selection within the intelligent agent application system, which can decouple the macro-decision behavior of model selection from the micro-content generation process within a single intelligent agent.

[0017] Specifically, upon receiving a model invocation request, a group of heterogeneous agents and a decision agent can be dynamically selected in the agent application system based on the model invocation request or pre-configured rules, using the agent's selection temperature parameters. This agent application system can manage various agents. In an exemplary implementation, taking the selection of heterogeneous agents and a decision agent based on the model invocation request as an example, after receiving the model invocation request, the instruction information carried in the model invocation request is parsed, and at least one group of heterogeneous agents and a decision agent are selected in the agent application system according to the instruction information. The instruction information may include the agent's domain of expertise, task type, and task parameters.

[0018] S120. Send the model invocation request to each heterogeneous agent through the input agent, and obtain the processing results of each heterogeneous agent.

[0019] The input agent can send model invocation requests to different heterogeneous agents. It can integrate agent components that convert general-format request information into inputs for different agents. Each agent can have various agent components. The implementation logic of the input agent is to convert the general-format request into the input format of each agent's response. The input agent can include Agent Communication Protocol (ACP), Model Context Protocol (MCP), etc. The processing results can include the response content output by each heterogeneous agent after model inference based on the input target request parameters. The processing results can include text generation results, classification labels, confidence scores, structured data, etc.

[0020] Specifically, the model service system can forward model invocation requests to different heterogeneous agents based on an input agent. In some embodiments, the input agent can invoke a model adapter to convert the model invocation request into a message format corresponding to each heterogeneous agent. After receiving the model invocation request, each heterogeneous agent can execute tasks and generate processing results.

[0021] S130. The adjudicating agent generates model output results according to the model call request and each processing result.

[0022] The model output can include the generated content that is determined and returned to the AI ​​application system after the adjudicating agent adjudicates the processing results of multiple heterogeneous agents.

[0023] Specifically, the adjudicating agent can organize model call requests and the processing results of various heterogeneous agents through adjudication, and generate model output results through the agent model. The adjudication methods of the adjudicating agent can include re-inference based on multiple results and selection of one of multiple results. The re-inference based on multiple results can include generating new results as model output results through comprehensive analysis and logical fusion of the processing results of various heterogeneous agents. The selection of one of multiple results can include selecting the optimal result as model output results through analysis and voting of each result. In this process, the adjudicating agent focuses on reasoning and selecting the differences generated by high-scoring heterogeneous agents to ensure the logical consistency, factual accuracy, and consistency with the original task intent of the final output result, thereby effectively reducing the risk of factual errors and illusions in the content generated by a single model and improving the overall reliability of the model service.

[0024] The technical solution of this invention, in response to a model invocation request, selects at least one group of heterogeneous agents and one adjudicating agent within the agent application system. Then, based on an input agent, the model invocation request is sent to each heterogeneous agent to obtain their processing results. Finally, the adjudicating agent generates the model output result according to the model invocation request and the processing results. This technical solution, by sending the model invocation request to each heterogeneous agent through an input agent and obtaining their processing results, enables multi-model collaborative decision-making, suppressing the knowledge illusion and coverage limitations of a single model. Furthermore, by having the adjudicating agent generate the model output result according to the model invocation request and the processing results, it allows for the selection of the optimal model combination for specialized tasks based on the agent's choice of temperature parameters. This achieves the beneficial effect of improving the accuracy of generated content and mitigating the semantic drift problem in the text generation process.

[0025] Furthermore, based on the above embodiments of the invention, the adjudicating agent generates model output results according to the model invocation request and each processing result, including: Determine the corresponding adjudicator's interface in the model adapter invoked by the input agent; The adaptation interface is called to convert the processing results and model call request parameters into target request parameters and processing result information that match the model parameter format of the adjudicating agent; The model output results are determined based on the target request parameters and processing result information determined by the adjudicating agent.

[0026] The processing result information may include a set of information generated by each heterogeneous agent and organized according to the model parameter format of the adjudicating agent. The processing result information can satisfy the input requirements of the adjudicating agent. This information can be obtained by each heterogeneous agent through reasoning according to the model call request.

[0027] Specifically, the adjudicating agent can find or dynamically generate the corresponding adaptation interface from the model adapter called by the input agent. The adjudicating agent calls the adaptation interface to receive the request parameters of the model call request and the processing results generated by each heterogeneous agent. It then converts the request parameters of the model call request and the processing results generated by each heterogeneous agent into target request parameters and processing result information that match the model parameter format of the adjudicating agent. It can also determine the model output result corresponding to the target request parameters and processing result information based on the adjudicating agent.

[0028] Example 2 Figure 2 This is a flowchart of a model service method provided in Embodiment 2 of the present invention. This embodiment further refines the above embodiments: like Figure 2 As shown, the method includes: S210. Determine the target professional field of the model call request, and obtain the agent selection parameters and agent selection temperature parameters, wherein the agent selection parameters include at least the number of candidate agents and the number of agents selected.

[0029] The target professional field can include the professional field category corresponding to the model invocation request, which may include code generation, medical diagnosis, legal consultation, etc. The number of candidate agents can include the number of models participating in the probability selection. The number of agents selected can include the number of heterogeneous agents ultimately selected from the agent application system. The agent selection temperature parameter T is used to adjust the entropy value of the model selection distribution. When the value of T increases, the probability of low-scoring models being selected relatively increases, thereby increasing system robustness; when the value of T decreases, the probability of high-scoring models being selected significantly increases, thereby ensuring task execution quality.

[0030] Specifically, the model service system can determine the target professional field corresponding to the task based on the professional field identifier in the model call request, and read the candidate agent selection parameters and agent selection temperature parameters to which the task belongs in the model call request. The agent selection parameters include at least the number of candidate agents and the number of agents selected.

[0031] S220. Based on the temperature parameters selected by the agent, determine the professional level score of each agent in the target professional field within the agent application system, and obtain the normalized score of each professional level score.

[0032] The professional competence score can include numerical values ​​used to quantify the level of each agent's ability in a specific professional field. The normalized score can include the professional competence score after normalization of each professional competence score. The normalized score can be the probability value obtained by transforming the professional competence scores of each agent through a function, and the sum of the normalized scores of all agents is 1.

[0033] Specifically, the model service system can obtain the professional level score of each agent in the intelligent agent application system in the target professional field based on the agent's selected temperature parameter. The professional level score can include historical evaluation values, expert scores, etc., and the system uses a normalization function to transform each professional level score to obtain a normalized score. The normalization function can include the Softmax function, piecewise linear normalization function, and ranking-based normalization function, etc. For example, the normalized score can be calculated using a normalization function formula with the agent's selected temperature parameter. This normalization function formula can include... ,in Refers to the model i In the score for the target professional field, T refers to the temperature parameter chosen by the agent, and K can represent the total number of agents. P ( i ) refers to the model i The normalized score value. By introducing the agent's selection temperature parameter T, the randomness of model selection can be dynamically adjusted, avoiding the neglect of diversity by only selecting high-scoring models in specific tasks.

[0034] S230. Based on the number of candidate agents, the number of agents selected, and the normalized scores, determine the heterogeneous agents and the adjudicating agents within each agent.

[0035] Specifically, agents can be ranked according to their normalized scores (either from highest to lowest or lowest to highest). A number of candidate agents are selected based on the normalized scores. A cumulative distribution function method can be applied to select a number of agents. The agent with the lowest normalized score among the selected agents is designated as the adjudicator agent, and the remaining selected agents are designated as heterogeneous agents. Heterogeneous agents, with their higher specialized scores, ensure high-quality generation. The adjudicator agent, possessing high generalization ability, is only responsible for selecting the best result from the outputs of multiple heterogeneous agents without modification; therefore, the agent with the lowest normalized score can be selected as the adjudicator agent. This selection strategy for heterogeneous agents and the adjudicator agent can leverage the characteristics of the cumulative distribution function to achieve dynamic optimization and role allocation, effectively improving the adaptability of the model combination to the task.

[0036] S240. Determine the corresponding adaptation interface for each heterogeneous agent in the model adapter called by the input agent.

[0037] The adaptation interface in the model adapter may include an interface module for converting input data into an input format that can be received by various heterogeneous agents. The adaptation interface can be used to reduce the differences between different heterogeneous agents, so that the input data can be uniformly received and processed by a group of heterogeneous agents. The differences between different heterogeneous agents may include different input formats, different data structures, different parameter specifications, etc.

[0038] Specifically, the input agent can call the model adapter and determine the interface information corresponding to the heterogeneous agent in the model adapter. Based on the interface information corresponding to each heterogeneous agent, the input agent selects the corresponding adaptation interface from the model adapter that matches the target heterogeneous agent so that the adaptation interface can be used to perform message format conversion.

[0039] S250. Call the adaptation interface to convert the request parameters of each model call request into target request parameters that match the model parameter format of the heterogeneous intelligent agent.

[0040] The request parameters can include the set of input data parameters carried in the model invocation request. The model parameter format can include a set of formats defined by each heterogeneous agent, including input data structure, data type, and data arrangement. The target request parameters can include a set of parameters that have been converted by the adaptation interface and meet the input specifications of the corresponding heterogeneous agent.

[0041] Specifically, the model service system can parse the request parameters from each model call request, call the corresponding adaptation interface of the target heterogeneous intelligent agent, and use the adaptation interface to convert the request parameters of each model call request into target request parameters that match the model parameter format of the heterogeneous intelligent agent. The conversion methods can include conversion according to field mapping relationship, conversion according to data type, conversion according to parameter recombination method, etc.

[0042] S260. Input the request parameters for each target into the corresponding heterogeneous intelligent agent.

[0043] Specifically, target request parameters in different message formats corresponding to different heterogeneous intelligent agents can be transmitted to the corresponding heterogeneous intelligent agents. After receiving the target request parameters, the heterogeneous intelligent agents can perform inference calculations on the target request parameters according to their own model architecture and parameter configuration, and generate processing results.

[0044] S270. Call the model adapter to receive and convert the response messages returned by each heterogeneous agent to obtain the processing results of each heterogeneous agent.

[0045] The processing results can include the response content output by each heterogeneous agent after performing model inference based on the input target request parameters. The processing results can include text generation results, classification labels, confidence scores, structured data, etc.

[0046] Specifically, the model adapter can be invoked to receive the response messages returned by each heterogeneous agent, and the received response messages can be formatted according to the adaptation interface in the model adapter, so as to perform unified analysis and statistics on the processing results fed back by each heterogeneous agent, so as to generate model output results in subsequent processes.

[0047] S280. The adjudicating agent generates model output results according to the model call request and each processing result.

[0048] The model output can include the generated content that is finally determined and returned to the AI ​​application system after the adjudicating agent has adjudicated the processing results of multiple heterogeneous agents.

[0049] Specifically, the adjudicating agent can organize the model invocation request and the processing results of various heterogeneous agents to generate the model output. The adjudication methods of the adjudicating agent can include re-inference based on multiple results, selecting one of multiple results, etc.

[0050] The technical solution of this invention involves determining the target professional field of the model call request and obtaining agent selection parameters and agent selection temperature parameters. The agent selection parameters include at least the number of candidate agents and the number of agents selected. Based on the agent selection temperature parameters, a normalized score is determined for the professional level rating of each agent within the application system corresponding to the target professional field. Then, based on the number of candidate agents, the number of agents selected, and each normalized score, heterogeneous agents and adjudicating agents are determined within each agent. The agent then calls the model adapter to select the corresponding adaptation interface for each heterogeneous agent. The adaptation interface is used to convert the request parameters of each model call request into target request parameters matching the model parameter format of the heterogeneous agent. Each target request parameter is input into the corresponding heterogeneous agent to obtain processing results. Finally, the adjudicating agent generates model output results based on the model call request and each processing result. The above technical solution, by inputting the target request parameters into the corresponding heterogeneous agents to obtain the processing results, can achieve multi-model collaborative decision-making and suppress the knowledge illusion and coverage limitations of a single model. By having the adjudicating agent generate model output results according to the model call request and the processing results, it can achieve optimal model combination selection and result optimization for professional tasks, thus achieving the beneficial effect of improving the accuracy of generated content and improving the semantic drift problem in the text generation process.

[0051] Furthermore, based on the above embodiments, the invention also includes: The input agent performs information trimming on each processing result, wherein information trimming includes at least one of the following: prompt word trimming and inference process information trimming; wherein prompt word trimming includes at least: identifying and removing text segments in the processing result that are repeated with the original prompt words in the model call request; inference process information trimming includes at least: identifying and removing intermediate inference step information contained in the processing result, retaining only the final output content.

[0052] Information pruning can include the process of removing redundant content from each processing result. Cue word pruning can include the process of removing repeated prompt words from the model call request in the processing result. Cue word acquisition can be achieved by identifying and removing text segments in the processing result that repeat the original prompt words in the model call request. Inference process information pruning can include the process of removing non-final output information contained in the agent's generated results from the processing results. Non-final output information can include intermediate inference steps, internal calculation processes, etc. Inference process information pruning can be achieved by identifying and removing intermediate inference step information contained in the processing results, retaining only the final output content. Information pruning can effectively reduce redundant data and improve the accuracy and efficiency of the adjudicating agent's processing.

[0053] Specifically, the processing results generated by various heterogeneous intelligent agents can be obtained, and each processing result can be pruned using an input agent. This pruning operation can include cue pruning and / or inference process pruning. When cue pruning is performed on the processing results based on the input agent, the original cue words in the model call request can be extracted, and text segments that repeat the original cue words in each processing result can be removed. When inference process pruning is performed on the processing results based on the input agent, the intermediate inference step information contained in the processing results can be identified, and the intermediate inference step information can be deleted, retaining only the final output content. For example, the intermediate inference step information in the processing results can use different fonts or colors, making the intermediate inference step information identifiable by the font or color. Alternatively, the intermediate inference step information can begin with a special identifier, making the intermediate inference step information in the processing results identifiable by this special identifier.

[0054] Example 3 Figure 3 This is a flowchart of a model service method provided in Embodiment 3 of the present invention. This embodiment further refines the above embodiments: like Figure 3 As shown, the method includes: S310. Determine the target professional field of the model call request, and obtain the agent selection parameters and agent selection temperature parameters, wherein the agent selection parameters include at least the number of candidate agents and the number of agents selected.

[0055] S320. Extract the historical score values ​​of each intelligent agent in the target professional field within the intelligent agent application system.

[0056] The historical score can include the evaluation value obtained by the agent in performing tasks in the target professional field within a historical period. The historical score can be obtained based on expert evaluation or user historical usage data.

[0057] Specifically, it can extract the target professional domain of the model call request and obtain the historical score of each agent in the intelligent agent application system corresponding to the target professional domain.

[0058] S330. Determine the first ratio between each historical score and the temperature parameter selected by the agent, and use each first ratio as an index to perform exponential calculation to obtain the score for each level of expertise.

[0059] The first ratio may include the quotient obtained by dividing each historical score by the temperature parameter selected by the agent.

[0060] Specifically, for each agent, the model service system can calculate the first ratio of the historical score to the temperature parameter selected by the agent. Each first ratio can be converted into an exponent with the same preset base, such as base e or base 10. The converted exponent can be used as the professional level score of the current agent. By iterating through all agents, the professional level score corresponding to each agent can be obtained. The preset base can be any constant.

[0061] S340. Determine the second ratio between each professional level score and the sum of all professional level scores, and use each second ratio as the normalized score for the corresponding agent.

[0062] The second ratio may include the quotient obtained by dividing the expertise score of each agent by the sum of the expertise scores of all agents.

[0063] Specifically, the sum of the professional competence scores of all agents can be calculated, and the professional competence score of each agent can be divided by the sum of the professional competence scores of all agents to obtain the normalized score corresponding to the current agent. This process can be repeated for all agents to obtain the normalized score corresponding to each agent.

[0064] For example, the normalized score of each agent can be obtained by the following formula: ,in Refers to the model i The historical score value, where T refers to the agent's chosen temperature parameter (T>0), e refers to the natural constant, and K can represent the total number of agents. P ( i ) refers to the model i Normalized score S350. Based on the number of candidate agents, the number of agents selected, and the normalized scores, determine the heterogeneous agents and the adjudicating agents within each agent.

[0065] S360, Determine the adaptation interface for each heterogeneous intelligent agent within the input agent.

[0066] S370. Call the adaptation interface to convert the request parameters of each model call request into target request parameters that match the model parameter format of the heterogeneous intelligent agent.

[0067] S380. Input the target request parameters into the corresponding heterogeneous intelligent agents to obtain the processing results.

[0068] S390. The adjudicating agent generates model output results according to the model call request and each processing result.

[0069] The technical solution of this invention involves determining the target professional domain of the model call request and obtaining agent selection parameters and agent selection temperature parameters. The agent selection parameters include at least the number of candidate agents and the number of agents selected. Historical scores of each agent in the target professional domain are obtained. A first ratio of each historical score to the agent selection temperature parameter is then determined, and each first ratio is used as an exponent for exponential calculation to obtain a professional level score. A second ratio of each professional level score to the sum of all professional level scores is then determined, and each second ratio is used as a normalized score for the corresponding agent. Heterogeneous agents and adjudicating agents are then determined within each agent according to the number of agents selected and the normalized scores. Adaptation interfaces for each heterogeneous agent within the input agent are then determined. These adaptation interfaces are then called to convert the request parameters of each model call request into target request parameters matching the model parameter format of the heterogeneous agents. Each target request parameter is then input into the corresponding heterogeneous agent to obtain processing results. Finally, the adjudicating agent generates model output results according to the model call request and the processing results. The aforementioned technical solution, through an adjudicating agent generating model outputs based on model call requests and processing results, enables multi-model collaborative decision-making. This suppresses the knowledge illusion and coverage limitations of single models, achieving optimal model combination selection and result optimization for specialized tasks. This results in improved accuracy of generated content and mitigates semantic drift issues in the text generation process. Furthermore, by introducing a temperature parameter for agent selection, the system can balance the exploratory nature of low-scoring models by adjusting the temperature value while ensuring high-scoring models are selected to maintain generation quality, avoiding local optima and significantly enhancing the adaptability of model services in complex cross-domain scenarios.

[0070] Furthermore, based on the above embodiments of the invention, heterogeneous agents and adjudicating agents are determined within each agent according to the number of candidate agents, the number of agents selected, and each normalized score, including: In the agent application system, select a number of agent agents as candidate agents according to the normalization scores from high to low. The cumulative distribution function is applied to the normalized scores of each candidate agent for screening, and the agents selected based on the number of times they are selected are considered as the selected agents. The selected agent with the lowest normalized score within each normalized score range is identified as the adjudicating agent, and the remaining selected agents are identified as heterogeneous agents.

[0071] Candidate agents can be a set of agents selected in descending order of their normalized scores, and then ordered by the number of candidates. The cumulative distribution function (F(x)) can be a mathematical function that applies a probability distribution for sampling. For any random variable X, F(x) means: F(x) = P(X≤x), which is the total probability that the random variable takes a value less than or equal to x. Cumulative distribution functions can include those based on the Softmax probability distribution, those based on the uniform distribution, etc. The selected agents can include the set of agents selected from the candidate agents after screening them by applying the cumulative distribution function according to the normalized scores. For example, the SOFTMAX function is called to calculate the corresponding SOFTMAX probability values ​​p1, p2, p3, p4 and p5 for each normalized score 1, 2, 3, 4 and 5. Then, the CDF function is used to process the SOFTMAX probability values ​​corresponding to the normalized scores of each agent to obtain the cumulative probability values ​​corresponding to different SOFTMAX probability values: F(1)=p1, F(2)=p1+p2, F(3)=p1+p2+p3, F(4)=p1+p2+p3+p4, F(5)=p1+p2+p3+p4+p5. The agents in the combination of agents that are within a certain cumulative probability range can be selected as the selected agents.

[0072] Specifically, the model service system can sort agents according to their normalized scores from highest to lowest, and select a number of agents as candidate agents based on the same normalized score ranking. Then, it calculates the probability distribution P(i) of each candidate agent across all candidate agent scores based on its normalized score, constructs a cumulative distribution function, and selects a number of agents from all candidate agents based on this cumulative distribution function as the selected agents. The agent with the lowest normalized score among the selected agents is selected as the adjudicator agent, and the other selected agents besides the adjudicator agent are defined as heterogeneous agents. Heterogeneous agents, with their higher professional scores, ensure high generation quality; the adjudicator agent has high generalization ability and is only responsible for selecting the best result from the heterogeneous agents' outputs without modification. Compared to traditional hard threshold screening, this selection strategy utilizes the characteristics of the cumulative distribution function to achieve dynamic optimization and role allocation, effectively improving the adaptability of the model combination to the task.

[0073] Furthermore, based on the above embodiments of the invention, determining heterogeneous agents and adjudicating agents within each agent according to the number of candidate agents, the number of agents selected, and each normalized score further includes: For each professional domain, a number of candidate agents are selected within the agent application system according to the normalized scores from high to low. The agent application system includes agents from multiple professional domains and at least one agent from a general domain. The cumulative distribution function is applied to the normalized scores of each candidate agent for screening, and agents with the number of selected agents in each professional field are selected as heterogeneous agents. The scores of each agent in the general domain agent pool are normalized based on the temperature parameter selected by the agent, and one general domain agent is selected as the adjudicator agent based on the normalization result.

[0074] Among them, general-domain intelligent agents can include intelligent agents that are capable of judging and adjudicating cross-domain comprehensive problems.

[0075] Specifically, the model service system groups agents according to their professional fields. Within each professional field, agents are ranked from highest to lowest based on their normalized scores, and a number of top-ranked agents are selected as candidate agents in descending order. Then, for the candidate agent set within each professional field, the probability distribution of each candidate agent within that professional field is calculated based on its normalized score, and a cumulative distribution function is constructed. Using the cumulative distribution function, a number of agents within the professional field are selected as heterogeneous agents. Based on the agent selection temperature parameter, the scores of each agent in the general domain agent pool are normalized, and one general domain agent is selected as the adjudicator agent based on the normalization result. In this embodiment of the invention, the cumulative distribution function randomly selects agents according to the number of candidate agents, ensuring that the agent combination is selected when specific conditions are met. This ensures the comprehensiveness of the agent combination and guarantees that various business scenarios can be handled by the agent combination, thereby mitigating the result deviation caused by the inconsistency between the professional level scoring scenario and the real scenario.

[0076] Furthermore, based on the above embodiments of the invention, the adjudicating agent generates model output results according to the model invocation request and each processing result, including: Determine the corresponding adapter interface of the adjudicator agent in the model adapter called by the input agent; The adaptation interface is called to convert the processing results and model call request parameters into target request parameters and processing result conversion information that match the model parameter format of the adjudicating agent; The model output results are determined based on the target request parameters and processing result transformation information determined by the adjudicating agent.

[0077] The processing result conversion information can include the processing results generated by each heterogeneous agent and the converted information. This processing result conversion information can match the model parameter format of the adjudicating agent and meet the input requirements of the adjudicating agent. The target request parameter can be the request parameter in the model call request after format conversion. This target request parameter can conform to the model parameter format of the adjudicating agent.

[0078] Specifically, the adjudicating agent can find or dynamically generate the corresponding adaptation interface from the model adapter. The adaptation interface of the adjudicating agent receives the request parameters of the model call request and the processing results generated by each heterogeneous agent, and converts the request parameters of the model call request and the processing results generated by each heterogeneous agent into target request parameters and processing result information that match the model parameter format of the adjudicating agent. It can also determine the model output result corresponding to the target request parameters and processing result information based on the adjudicating agent.

[0079] Furthermore, based on the above embodiments of the invention, the model output result corresponding to the target request parameters and processing result information determined by the adjudicating agent includes at least one of the following: The control and adjudication agent selects the model output result from each processing result information according to the target request parameters; The control and adjudication agent integrates the processing results into the model output based on the target request parameters.

[0080] Specifically, the adjudicating agent can receive target request parameters and processing result information, and determine the model output corresponding to the processing result information. Determining the model output corresponding to the target request parameters and processing result information can include two methods: controlling the adjudicating agent to select the model output from each processing result according to the target request parameters, and controlling the adjudicating agent to fuse each processing result into a single model output according to the target request parameters.

[0081] Example 4 The model service method described in this embodiment employs a dynamic heterogeneous model selection mechanism, specifically: First, construct a professional domain capability matrix, which includes professional domains, models, and scoring indicators to quantify the professional capabilities of the intelligent agent model.

[0082] Secondly, a dynamic scheduling algorithm is designed to retain Top-K candidate models based on ability scores and eliminate interference from low-scoring models. A Softmax function is applied to the Top-K models to generate probability distributions. The randomness of selection is adjusted by using a temperature parameter T (T>0) for agent selection (T<1 strengthens high-scoring models, T>1 enhances exploratory behavior). Unlike traditional techniques where the temperature parameter only affects the content generated by a single model, this invention decouples the agent's selection temperature parameter from the generation temperature parameter used during inference within each heterogeneous agent. This allows for independent control and optimization of the macro-level model combination strategy and the micro-level content generation behavior, thereby maintaining the robustness and diversity of model combinations while ensuring high-quality and high-determinism of the generated results. This scheme significantly improves the adaptability of model services in complex cross-domain scenarios by dynamically adjusting parameters rather than statically configuring them.

[0083] Figure 4 This is a diagram of a hierarchical selection decision mode provided in Embodiment 4 of the present invention.

[0084] For details, see Figure 4 This embodiment describes a hierarchical selection and adjudication mode applicable to models M1, M2, and M3 all belonging to the same professional domain. This hierarchical selection and adjudication mode differs from traditional data-based voting mechanisms by employing generative adjudication. The output of the generative model includes: constructing a prompt containing the original request and the results generated by each heterogeneous agent, instructing the adjudicating agent to perform re-reasoning based on logical consistency. Specifically, high-scoring models M1 and M2 generate results R1 and R2. The adjudicating agent M3, based on the original input and the results R1 and R2, constructs the prompt and performs reasoning to regenerate the model, selecting one result from R1 and R2 to address potential factual errors and illusions in content generated by a single model.

[0085] Figure 5 This is a diagram of a professional field fusion adjudication mode provided according to Embodiment 4 of the present invention.

[0086] For details, see Figure 5 This embodiment describes a domain-integrated adjudication model applicable to scenarios where M1 and M2 represent different domains, and M3 is a general-purpose model. Unlike traditional data-comparison-based voting mechanisms, this domain-integrated adjudication model employs generative adjudication (where high-scoring models M1 and M2 generate results R1 and R2, and M3, based on the original input and R1 / R2, performs reasoning and regenerates, fusing R1 / R2), thus resolving the semantic merging problem of text generation from two different domain models.

[0087] Furthermore, the model service method described in this embodiment adopts a zero-intrusive proxy architecture. Its bidirectional adaptation layer design is compatible with existing intelligent agent model API interface specifications for AI application systems, supports multiple models on both local and cloud platforms, and encapsulates heterogeneous model calls. A dynamic expansion mechanism provides a hot-swappable interface, supporting model expansion by registering new models through a capability matrix and enabling real-time scheduling. It also supports the expansion of adjudication logic, allowing users to extend and customize consensus algorithms. Multiple professional domains are managed in separate zones, with capability matrices and adjudication strategies maintained independently for each domain, ensuring domain exclusivity.

[0088] Figure 6 This is a system architecture diagram of a model service method provided according to Embodiment 4 of the present invention.

[0089] For details, see Figure 6 The system architecture diagram of the model service method described in this embodiment includes a system management module, an input agent module, a model scoring matrix module, an agent selector module, a arbitrator module, and a model adapter module, as detailed below: The system management module manages business domain registration and model parameter configuration. It dynamically loads adjudication algorithm plugins and triggers the input agent to open new ports when a new domain is added. The system management module can send port configuration commands to the input agent and receive model adapter registration information. The input agent module opens independent ports for each domain and supports multiple API format conversions. The input agent module receives AI application requests and passes domain identifiers to the agent selector module. The model scoring matrix module quantifies the model capability database; scores are derived from authoritative evaluation datasets and benchmark data from professional domain literature. The model scoring matrix module can manage the writing of new professional level scores. The agent selector module dynamically selects the scheduling engine, filtering models based on professional level scores, the number of candidate agents (Top-K), and the agent selection temperature parameter T, applying the Softmax normalization function algorithm and the cumulative distribution function method during the model selection process. The agent selector module can be triggered by the input agent to schedule and output a set of models to the adjudicator. The arbitrator module is the core of collaborative decision-making. This module incorporates three types of arbitrator components: dynamic arbitrator components based on regression models, classification arbitrator components based on expertise scores, and dynamic arbitrator components based on agent models. These components can be categorized as generative, regression, and classification components. The arbitrator also supports custom plugins and can orchestrate multi-stage model call flows. The arbitrator module can receive agent selector output, invoke model adapters, and return arbitration results to the output agent. The model adapter module includes a built-in general adapter and provides open message format conversion interfaces for user extension. The input agent can send requests to cloud or local models; this model adapter module can include adaptation interfaces for various agents.

[0090] AI application system agent developers select and set up ports for their respective professional domains based on the expertise required for their business needs. A complete business request includes: the agent sending a request prompt to the mimicry model management framework; the input agent notifying the agent selector of the prompt and the professional domain identifier as parameters; the agent selector querying the number of models required by the mimicry adjudicator based on the professional domain identifier, and selecting the model set based on parameters such as Top-K, model selection temperature, and the professional level rating of agents within the model scoring matrix module.

[0091] The adjudicator's processing includes: querying the dynamic adjudication component based on the domain identifier; calling the adaptation interface provided by the model adapter to convert user parameters into the target model parameter format; the dynamic adjudication component concurrently calling multiple agent models and adjudicating the results returned by the agent models; the dynamic adjudication component can call agent models for inference in stages according to domain requirements, for example, the first stage calls the agent model of the first domain for inference, and the second stage calls the agent model of the second domain for inference. The domain knowledge used by the agent models called in the first and second stages can be different. The output agent organizes the final processing result into the message format required by the agents of the AI ​​application system and returns it to the caller.

[0092] Furthermore, in the model service method described in this embodiment, under the hierarchical optimization decision mode, the specific content of the model selection algorithm is as follows: Each professional domain corresponds to one model adjudication component. This component contains the required number of models N, the Top-K candidate agents configured by the user through the system management component, and the agent selection temperature parameter T. The following explanation uses professional domain 1, with N=3 selected models, Top-K candidate models=5, and T=0.9 as an example.

[0093] Based on Top-K=5, models 1, 2, 3, 5, and 6 are selected. An example of the professional domain rating matrix in the model rating matrix module is shown in Table 1. Table 1. Example of a rating matrix for a professional field The agent selects temperature parameters and applies the Softmax function to each score in the professional domain D, with the following formula: ,in Representation Model iIn the professional field, D represents the score, which is the quantitative value of professional ability; T represents the model selection temperature coefficient, which is used to adjust the entropy value of the model selection distribution. When the T value increases, the probability of low-scoring models being selected increases relatively, so as to increase the robustness of the system; when the T value decreases, the probability of high-scoring models being selected increases significantly, so as to ensure the quality of task execution. P(i) express i The selection probability is calculated. Finally, using the cumulative distribution function method, three models (e.g., Model 1, Model 3, and Model 6) are selected as the inference model set.

[0094] Figure 7 This is a schematic diagram of a semantic-based model hierarchical optimization decision algorithm provided in Embodiment 4 of the present invention.

[0095] For details, see Figure 7 The semantic-based model adjudication algorithm described in this embodiment is as follows: First, the selected inference models are ranked from highest to lowest based on their professional level scores: Model 6, Model 3, and Model 1. The top two models (Model 6 and Model 3) are used for inference to generate results; the third model (Model 1) is used for inference selection. The adjudication component uses model adapters to convert the user request into message formats for Model 6 and Model 3, respectively. The adjudication component then sends the converted messages to Model 6 and Model 3, respectively. The adjudicator uses the Model 1 adapter to convert the hierarchical selection adjudication system prompt (i.e., the system prompt in hierarchical selection adjudication mode), the original user request message, and the generated results from Model 6 and Model 3 into a request message for Model 1. The hierarchical selection adjudication system prompt guides the agent model to use, for example, the following methods based on the original user request message: Figure 4 The hierarchical optimization decision-making mode shown will analyze the generation results of Model 6 and Model 3 and select the better model generation result. Finally, the arbitrator sends the generated request message to Model 1; based on the result returned by Model 1 (Model 6 or Model 3), the arbitrator selects the generation result of the corresponding model, converts it into a message format compatible with the intelligent agent, and returns it to the AI ​​application system.

[0096] The specific details of the model selection algorithm under the professional field integration adjudication model are as follows: Two professional domains that need to be merged correspond to one model adjudication component, requiring a total of N=3 models. The user configures the Top-K number of candidate agents and the agent selection temperature parameter T through the system management component. The following explanation uses the fusion of professional domains 1 and 2, with N=3, Top-K=2, and T=0.9 as an example.

[0097] Based on Top-K=2, models 3 and 6 are selected for Specialty Domain 1, models 2 and 4 are selected for Specialty Domain 2, and models 1 and 5 are selected for the General Domain. An example of the domain scoring matrix is ​​shown in Table 2. Table 2 Example of Domain Scoring Matrix The agent selects temperature parameters and applies the Softmax function to each score in the professional domain D, with the following formula: ,in Representation Model i In the professional field, D represents the score, which is the quantitative value of professional ability; T represents the agent selection temperature parameter, which is used to adjust the entropy value of the model selection distribution. When the value of T increases, the probability of low-scoring models being selected increases relatively, so as to increase the robustness of the system; when the value of T decreases, the probability of high-scoring models being selected increases significantly, so as to ensure the quality of task execution. P(i) express i The selection probability is calculated. Finally, for each professional and general domain, the cumulative distribution function method is used to select one model (e.g., Model 6, Model 2, and Model 1) as the inference model set.

[0098] Figure 8 This is a schematic diagram of a professional domain fusion algorithm provided in Embodiment 4 of the present invention.

[0099] For details, see Figure 8 The professional domain fusion algorithm described in this embodiment is as follows: First, Model 6 is selected for inference to generate results for domain 1; Model 3 is selected for inference to generate results for domain 2; the adjudication component uses model adapters to convert user requests into message formats for Model 6 and Model 3 respectively; the adjudication component sends the converted messages to Model 6 and Model 3 respectively; the adjudicator uses the adapter of Model 1 to convert the domain fusion adjudication system prompt (i.e., the system prompt in the domain fusion adjudication mode), the original user request message, and the generated results of Model 6 and Model 3 into a request message for Model 1; among them, the domain fusion adjudication system prompt guides the agent model to adopt, for example, based on the original user request message. Figure 5 The professional domain fusion adjudication model shown will analyze the generated results of Model 6 and Model 3 to generate the final adjudication result. Finally, the adjudicator will send the generated request message to Model 1; based on the result returned by Model 1, the adjudicator will convert it into a message format compatible with the intelligent agent and return it to the AI ​​application system.

[0100] Furthermore, the model service method described in this embodiment includes a dynamic adjudication component for regression models, whose adjudication algorithm includes output mean and median optimization. For example, output mean refers to the adjudicator selecting two models, performing concurrent inference, and calculating the average as the final result; median optimization refers to the adjudicator selecting three models, performing concurrent inference, and selecting the middle value of the inference results as the final result. The classification adjudication component based on expertise scoring includes majority voting and equal division arbitration. For example, majority voting refers to the adjudicator selecting three models, performing concurrent inference, and using the majority adjudication (≥2 of the same category) as the final classification result; equal division arbitration refers to breaking ties by scoring the classification results of each model according to expertise; if the three classification results are all different, the result with the higher original score is adopted.

[0101] The technical solution of this invention, through the collaborative work of the model scoring matrix and the agent selector, can achieve quantitative evaluation of the agent model's capabilities in a professional domain and real-time selection of the optimal model combination, effectively solving the problems of limited knowledge coverage of a single model and insufficient generalization ability across multiple domains. Through the collaboration of the generative mimicry adjudication component and the model adapter in the adjudication layer, cross-validation, optimization, or fusion of the generated model content can be achieved, significantly improving the accuracy and security of the generated results. This embodiment constructs two adjudication modes: hierarchical optimization and professional domain fusion. Combined with a zero-intrusive agent architecture, it realizes dynamic collaboration of multi-domain model capabilities, significantly enhancing the system's scalability and task adaptability.

[0102] Example 5 Figure 9 This is a schematic diagram of a model service system provided according to Embodiment 5 of the present invention. Figure 9 As shown, the system includes: a model selection module 410, a model processing module 420, and a result adjudication module 430. Among them, The model selection module 410 is used to select at least one set of heterogeneous agents and one adjudicating agent within the agent application system in response to a model call request, based on the agent selection temperature parameter.

[0103] The model processing module 420 is used to send the model invocation request to each of the heterogeneous intelligent agents through the input agent, and to obtain the processing results of each of the heterogeneous intelligent agents.

[0104] The result adjudication module 430 is used to generate model output results by the adjudication agent according to the model call request and each of the processing results.

[0105] The technical solution of this invention involves a model selection module responding to a model invocation request by selecting at least one group of heterogeneous agents and one adjudicating agent within the agent application system. A model processing module sends the model invocation request to each heterogeneous agent based on an input agent, obtaining the processing results from each agent. Finally, a result adjudication module generates the model output result based on the model invocation request and the processing results from the adjudicating agent. This technical solution, by sending the model invocation request to each heterogeneous agent through an input agent and obtaining their processing results, enables multi-model collaborative decision-making, suppressing the knowledge illusion and coverage limitations of a single model. Furthermore, by having the adjudicating agent generate the model output result based on the model invocation request and the processing results, it achieves optimal model combination selection and result optimization for specialized tasks, thus improving the accuracy of the generated content.

[0106] Optionally, the model selection module 410 includes: The information extraction unit is used to determine the target professional field of the model call request and to obtain the agent selection parameters and the agent selection temperature parameters, wherein the agent selection parameters include at least the number of candidate agents and the number of agents selected.

[0107] The normalized scoring unit is used to determine the professional level score of each agent in the application system corresponding to the target professional field based on the temperature parameter selected by the agent, and to obtain the normalized score of each professional level score.

[0108] An agent selection unit is configured to determine each heterogeneous agent and the adjudicating agent within each of the agents according to the number of candidate agents, the number of agents selected, and the normalized scores of each agent.

[0109] Based on the above-described embodiments, the normalized scoring unit is specifically used for: extracting historical scoring values ​​of each agent in the intelligent agent application system corresponding to the target professional field; determining a first ratio between each historical scoring value and the temperature parameter selected by the agent, and using each first ratio as an exponent to perform exponential calculation to obtain each professional level score; determining a second ratio between each professional level score and the sum of each professional level score, and using each second ratio as the normalized score corresponding to each agent.

[0110] Based on the above-described embodiments, the agent selection unit is specifically configured to: select the number of agents as candidate agents within the agent application system according to the normalized scores from high to low; apply a cumulative distribution function to the normalized scores of each candidate agent to filter and select the number of agents selected as admitted agents; determine the admitted agent with the lowest normalized score among the normalized scores as the adjudicating agent, and determine the remaining admitted agents as the heterogeneous agents.

[0111] Based on the above-described embodiments, the agent selection unit is further configured to: for each professional domain, select the number of agent candidates as candidate agent within the agent application system according to the normalized scores from high to low; wherein the agent application system includes agent candidates from multiple professional domains and at least one agent from a general domain; apply a cumulative distribution function to the normalized scores of each candidate agent to filter them, so as to select the number of agent candidates selected in each professional domain as the heterogeneous agent; normalize the scores of each agent in the general domain agent pool based on the agent selection temperature parameter, and select one agent from the general domain as the adjudicating agent based on the normalization result.

[0112] Optionally, the model processing module 420 is specifically used for: determining the adaptation interface corresponding to each heterogeneous agent in the model adapter invoked by the input agent; calling the adaptation interface to convert the request parameters of the model invocation request into target request parameters that match the model parameter format of the heterogeneous agent; inputting each of the target request parameters into the corresponding heterogeneous agent; and calling the model adapter to receive and convert the response messages returned by each heterogeneous agent to obtain the processing results of each heterogeneous agent.

[0113] Optionally, the apparatus provided in this embodiment of the invention further includes: an information trimming module, used to trim information of each of the processing results through the input agent, wherein the information trimming includes at least one of the following: prompt word trimming and inference process information trimming; wherein the prompt word trimming includes at least: identifying and removing text segments in the processing results that are repeated with the original prompt words in the model call request; the inference process information trimming includes at least: identifying and removing intermediate inference step information contained in the processing results, retaining only the final output content.

[0114] Optionally, the result adjudication module 430 is specifically used to: determine the adaptation interface of the adjudication agent in the model adapter called by the input agent; call the adaptation interface to convert the processing result and the request parameters of the model call request into target request parameters and processing result conversion information that match the model parameter format of the adjudication agent; and determine the model output result corresponding to the target request parameters and the processing result conversion information according to the adjudication agent.

[0115] Optionally, the model output result determined by the result adjudication module 430 based on the adjudication agent, corresponding to the target request parameters and the processing result transformation information, includes at least one of the following: The adjudicating agent is controlled to select the model output result from each of the processing result transformation information according to the target request parameters; The decision-making agent is controlled to fuse the processing results into the model output result according to the target request parameters.

[0116] The model service apparatus provided in this embodiment of the invention can execute the model service method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0117] Example 6 Figure 10 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0118] like Figure 10As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0119] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0120] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as model servicing methods.

[0121] In some embodiments, the model servicing method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the model servicing method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the model servicing method by any other suitable means (e.g., by means of firmware).

[0122] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0123] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0124] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0125] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0126] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0127] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0128] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0129] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A model service method, characterized in that, The method includes: In response to a model invocation request, at least one set of heterogeneous agents and one adjudicating agent are selected within the agent application system based on the agent's selected temperature parameter. The model invocation request is sent to each of the heterogeneous intelligent agents through the input agent, and the processing results of each of the heterogeneous intelligent agents are obtained. The adjudicating agent generates model output results according to the model invocation request and each of the processing results.

2. The method according to claim 1, characterized in that, The step of responding to a model invocation request by selecting at least one set of heterogeneous agents and one adjudicating agent within the agent application system based on the agent's selected temperature parameter includes: The target professional domain of the model invocation request is determined, and agent selection parameters and agent selection temperature parameters are obtained, wherein the agent selection parameters include at least the number of candidate agents and the number of agents selected; Based on the temperature parameters selected by the agent, the professional level score of each agent in the agent application system corresponding to the target professional field is determined, and the normalized score of each professional level score is obtained. Based on the number of candidate agents, the number of agents selected, and the normalized scores, each heterogeneous agent and the adjudicating agent are determined within each of the agents.

3. The method according to claim 2, characterized in that, The process of determining the professional competence score of each agent in the intelligent agent application system corresponding to the target professional field based on the temperature parameters selected by the agent, and obtaining a normalized score for each professional competence score, includes: Extract the historical score values ​​of each intelligent agent in the intelligent agent application system corresponding to the target professional field; Determine a first ratio between each of the historical score values ​​and the temperature parameter selected by the agent, and use each of the first ratios as an index to perform exponential calculation to obtain each of the professional level scores; A second ratio is determined between each of the aforementioned professional competence scores and the sum of the aforementioned professional competence scores, and each of the second ratios is used as the normalized score corresponding to each of the aforementioned agents.

4. The method according to claim 2, characterized in that, The step of determining the heterogeneous agents and the adjudicating agent within each of the agents according to the number of candidate agents, the number of agents selected, and the normalized scores of each agent includes: According to the normalized scores, the number of candidate agents are selected as candidate agents within the agent application system in descending order. The cumulative distribution function is applied to the normalized scores of each candidate agent to filter them, and the number of agents selected by each agent is selected as the included agents. The selected agent with the lowest normalized score among the normalized scores is determined as the adjudicating agent, and the remaining selected agents are determined as the heterogeneous agents.

5. The method according to claim 2, characterized in that, The step of determining the heterogeneous agents and the adjudicating agent within each of the agents according to the number of candidate agents, the number of agents selected, and the normalized scores of each agent includes: For each professional domain, the number of candidate agents are selected as candidate agents within the agent application system according to the normalized scores from high to low; wherein, the agent application system includes agents from multiple professional domains and at least one agent from a general domain. The normalized scores of each candidate agent are filtered using a cumulative distribution function to select the number of agents selected in each professional field as the heterogeneous agents. Based on the temperature parameter selected by the agent, the scores of each agent in the general domain agent pool are normalized, and one general domain agent is selected as the adjudicating agent according to the normalization result.

6. The method according to claim 1, characterized in that, The step of sending the model invocation request to each of the heterogeneous agents through the input agent and obtaining the processing results of each of the heterogeneous agents includes: Determine the adaptation interface corresponding to each heterogeneous agent in the model adapter invoked by the input agent; The adapter interface is invoked to convert the request parameters of the model invocation request into target request parameters that match the model parameter format of the heterogeneous agent; Each of the target request parameters is input into the corresponding heterogeneous intelligent agent; The model adapter is invoked to receive and transform the response messages returned by each heterogeneous agent to obtain the processing results of each heterogeneous agent.

7. The method according to claim 1 or 6, characterized in that, Before the process of generating the model output result by the adjudicating agent according to the model invocation request and each of the processing results, it further includes: The input agent performs information trimming on each of the processing results, wherein the information trimming includes at least one of the following: prompt word trimming, reasoning process information trimming; The prompt word trimming includes at least: identifying and removing text segments in the processing result that are repeated with the original prompt words in the model call request; The information trimming in the reasoning process includes at least: identifying and removing intermediate reasoning step information contained in the processing result, and retaining only the final output content.

8. The method according to claim 1, characterized in that, The step of generating model output results through the adjudicating agent according to the model invocation request and each of the processing results includes: Determine the corresponding adaptation interface of the adjudicating agent in the model adapter invoked by the input agent; The adaptation interface is invoked to convert the processing result and the request parameters of the model call request into target request parameters and processing result conversion information that match the model parameter format of the adjudicating agent; The model output result corresponding to the target request parameters and the processing result conversion information is determined by the adjudicating agent.

9. The method according to claim 8, characterized in that, The model output result corresponding to the target request parameter and the processing result transformation information determined by the adjudicating agent includes at least one of the following: The adjudicating agent is controlled to select the model output result from each of the processing result transformation information according to the target request parameters; The decision-making agent is controlled to fuse the processing results into the model output result according to the target request parameters.

10. A model service system, characterized in that, The system includes: The model selection module is used to select at least one set of heterogeneous agents and one adjudicating agent within the agent application system in response to a model call request, based on the agent selection temperature parameter. The model processing module is used to send the model invocation request to each of the heterogeneous intelligent agents through the input agent, and to obtain the processing results of each of the heterogeneous intelligent agents; The result adjudication module is used to generate model output results by the adjudication agent according to the model call request and each of the processing results.

11. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model service method according to any one of claims 1-9.