Construction and intelligent scheduling method of multi-model service based on large model

Abstract

The invention discloses a multi-model service construction and intelligent scheduling method based on a large model. The multi-model service construction and intelligent scheduling method based on the large model comprises the following steps: S1, constructing a multi-model integration platform; s2, registering and evaluating the model; s3, the semantic vector of the user service request is disassembled and understood; s4, model selection and priority screening; s5, sorting service response priorities and intelligently assigning resources; s6, optimizing and feeding back a feedback result; s7, performing hit labeling and result training; according to the multi-model service construction and intelligent scheduling method based on the large model, an artificial intelligence integration platform is constructed, multiple artificial intelligence models including the large model are subjected to unified services and interfaces, and intelligent understanding of user service requests, intelligent model selection, intelligent service response scheduling and feedback result optimization are achieved; and therefore, the multi-scene artificial intelligence technology application requirements of the user can be met.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method for constructing and intelligently scheduling multi-model services based on a large model. Background Art

[0002] With the rapid development of artificial intelligence technology, users' demand for artificial intelligence services is increasing, and the demand scenarios are becoming increasingly diverse. However, traditional artificial intelligence applications are usually limited to a single technology and cannot meet users' needs for multi-scenario artificial intelligence technology applications under a single application. Therefore, how to build a system and method that can integrate multiple artificial intelligence technologies and meet users' diverse needs through intelligent scheduling has become an urgent problem to be solved in the current field of artificial intelligence technology.

[0003] The current technical solution has two drawbacks:

[0004] First, when using large model technology, general large models cannot achieve domain depth in a specific industry. Improving generation capabilities requires training of large parameter models, which has high resource costs. Industry models focus on deep training in a certain field, but have poor general capabilities. At the same time, due to the limitations of large model technology, when completing specific business scenarios, it is necessary to use a combination of large and small models, which requires the application system's technology stack to be more complicated than a single large model application.

[0005] Second, in application scenarios with diverse requirements, a single model cannot meet the scenario requirements. It is necessary to decompose the target requirements into tasks and complete different sub-task requirements through a combination of multiple models. If this problem is not fully designed in the early stage of application planning, it will have a great impact on the subsequent changes in the technical architecture. In the case of relatively single traditional artificial intelligence applications, this defect is not obvious, but with the widespread application of artificial intelligence technology and the trend of complex and changing application scenarios, the defect has become more and more obvious. Summary of the invention

[0006] The purpose of the invention is to provide a method for constructing and intelligently scheduling multi-model services based on a large model, which solves the problems in the background technology.

[0007] The present invention is implemented as follows: a method for constructing and intelligently scheduling a multi-model service based on a large model, the method for constructing and intelligently scheduling a multi-model service based on a large model comprises the following steps:

[0008] S1. Build a multi-model integration platform;

[0009] S2, model registration and evaluation;

[0010] S3, semantic vector decomposition and understanding of user service requests;

[0011] S4, model selection and priority screening;

[0012] S5, service response priority sorting and resource intelligent allocation;

[0013] S6. Feedback result optimization and feedback;

[0014] S7. Hit labeling and result training.

[0015] A further technical solution of the present invention is: the construction of a multi-model integration platform is: the platform integrates multiple artificial intelligence models including large models, and provides unified services and interfaces; the platform adopts a standardized and modular design to facilitate the access and expansion of different models.

[0016] A further technical solution of the present invention is: the registration and evaluation of the model are as follows: the capabilities, task objectives, resource overhead, and calculation types of the model built on the platform are registered, the capabilities of the model are evaluated, evaluation indicators, capability ranges, access rules are formed, and the scheduling model is fine-tuned and trained; an indicator matrix is ​​formed by capability labeling and weight calculation, and the model evaluation indicators are defined according to the matrix, and finally an indicator score for the model is formed. When multiple models have the same business capability, they can be selected based on the evaluation indicators.

[0017] A further technical solution of the present invention is: the semantic vector decomposition and understanding of the user service request is as follows: when the user makes a service request in the form of a natural language dialogue, the system first decomposes the user request into a semantic vector, converts the user's natural language input into a semantic vector that can be understood by the computer, and obtains the specific parameters of the user request. Then, the system performs semantic understanding on the request to clarify the user's specific needs.

[0018] A further technical solution of the present invention is that the model selection and priority screening are as follows: the system searches and matches the rule base in the platform according to the specific parameters of the user request, and determines the model with the highest parameter index corresponding to the service demand. If there are multiple models that meet the index, they are screened according to the preset priority rules, and the most appropriate model is selected for service response.

[0019] A further technical solution of the present invention is that the service response priority sorting and resource intelligent allocation are as follows: the system prioritizes the service responses according to factors such as the urgency and importance of the user's request, and establishes a response queue. Then, the system intelligently allocates the response resources according to the calculation form and resource requirements of the model, ensuring that high-priority service requests are processed first, while avoiding resource waste and bottleneck phenomena.

[0020] A further technical solution of the present invention is: the feedback result optimization and feedback are: obtaining the results of model feedback, and assembling and optimizing the results of multiple model feedback; finally, the system feeds back the optimized results to the end user to meet the user's application needs; the optimization process can be carried out according to preset rules or algorithms to improve the accuracy and availability of the results.

[0021] A further technical solution of the present invention is: the hit marking and result training are: recording the hit situation of each feedback result, evaluating the customer's satisfaction with the hit result, accurately hitting the platform, and marking the hit result, modifying the indicator score of each model in the evaluation index, and completing self-learning and iterative training of model selection.

[0022] Beneficial effects of the invention: The method for constructing and intelligently scheduling multi-model services based on big models of the invention unifies the services and interfaces of various artificial intelligence models including big models by building an artificial intelligence integration platform, realizes intelligent understanding of user service requests, intelligent selection of models, intelligent scheduling of service responses, and optimization of feedback results, so as to meet users' multi-scenario artificial intelligence technology application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of a method for constructing and intelligently scheduling a multi-model service based on a large model provided by the present invention;

[0024] Figure 2 It is a module structure diagram of a method for constructing and intelligently scheduling a multi-model service based on a large model provided by the present invention. DETAILED DESCRIPTION

[0025] Embodiment 1: Figure 1 A method for constructing and intelligently scheduling a multi-model service based on a large model is shown, and the method for constructing and intelligently scheduling a multi-model service based on a large model includes the following steps:

[0026] S1. Build a multi-model integration platform;

[0027] S2, model registration and evaluation;

[0028] S3, semantic vector decomposition and understanding of user service requests;

[0029] S4, model selection and priority screening;

[0030] S5, service response priority sorting and resource intelligent allocation;

[0031] S6. Feedback result optimization and feedback;

[0032] S7. Hit labeling and result training.

[0033] The construction of the multi-model integration platform is as follows: the platform integrates multiple artificial intelligence models including large models, and provides unified services and interfaces; the platform adopts a standardized and modular design to facilitate the access and expansion of different models.

[0034] The registration and evaluation of the model are as follows: register the capabilities, task objectives, resource overhead, and computing types of the model built on the platform, evaluate the capabilities of the model, form evaluation indicators, capability ranges, and access rules, and fine-tune the scheduling model rules; model registration forms an information matrix for the model, and the matrix example is as follows:

[0035]

[0036] The indicator matrix is ​​formed by capability annotation and weight calculation. The model evaluation indicators are defined according to the matrix, and finally the indicator score of the model is formed. When multiple models have the same business capability, they can be selected by evaluation indicators. The example is as follows:

[0037]

[0038] The semantic vector decomposition and understanding of the user service request is as follows: when the user makes a service request in the form of natural language dialogue, the system first decomposes the user request into a semantic vector, converts the user's natural language input into a semantic vector that can be understood by the computer, and obtains the specific parameters of the user's request. Then, the system semantically understands the request and clarifies the user's specific needs.

[0039] The model selection and priority screening is as follows: the system searches and matches the rule base in the platform according to the specific parameters of the user's request, and determines the model with the highest parameter index corresponding to the service requirement. If there are multiple models that meet the index, they are screened according to the preset priority rules and the most appropriate model is selected for service response.

[0040] The service response priority sorting and resource intelligent allocation are as follows: the system prioritizes service responses and establishes a response queue based on factors such as the urgency and importance of user requests. Then, the system intelligently allocates response resources based on the calculation form and resource requirements of the model to ensure that high-priority service requests are processed first while avoiding resource waste and bottlenecks.

[0041] The feedback result optimization and feedback are as follows: obtaining the results of model feedback, and assembling and optimizing the results of multiple model feedbacks; finally, the system feeds back the optimized results to the end user to meet the user's application needs; the optimization process can be performed according to preset rules or algorithms to improve the accuracy and availability of the results.

[0042] The hit marking and result training are as follows: recording the hit situation of each feedback result, evaluating the customer's satisfaction with the hit result, accurately hitting the platform, marking the hit result, modifying the indicator score of each model in the evaluation index, and completing self-learning and iterative training of model selection.

[0043] The method for constructing and intelligently scheduling multi-model services based on big models of the present invention unifies the services and interfaces of various artificial intelligence models including big models by building an artificial intelligence integration platform, thereby realizing intelligent understanding of user service requests, intelligent selection of models, intelligent scheduling of service responses, and optimization of feedback results, so as to meet users' multi-scenario artificial intelligence technology application needs.

[0044] Embodiment 2: Figure 2 A system for constructing a multi-model service based on a large model and an intelligent scheduling method is shown, the system comprising:

[0045] User interface module: This module combines natural language processing (NLP) with large-model semantic understanding technology to build an efficient user interaction interface that can accurately capture and analyze natural language service requests input by users, perform in-depth analysis and understanding of semantic vectors, and ensure accurate recognition capabilities for user requests.

[0046] Model management module: As the core control unit of the system, the model management module is responsible for intelligently managing and scheduling a variety of AI model resources, including large models. Based on the characteristics of the request (such as complexity and real-time requirements), the module can automatically select the optimal model combination and perform dynamic loading and unloading operations to optimize system performance and response speed.

[0047] Model service standardization module: This module aims to achieve the standardization and unification of multi-model service interfaces, and ensure seamless integration and efficient collaboration between different models by defining unified service interface specifications, service forms and technical parameter standards. This standardization process simplifies the complexity of service development and integration, and improves the scalability and maintainability of the system.

[0048] Model rule evaluation and optimization module: This module builds a comprehensive performance evaluation system to comprehensively evaluate the functions, performance, resource consumption and weight parameters of multiple models. Based on the evaluation results, the module can automatically adjust the service rules, optimize the model combination and scheduling strategy, and ensure that the system always responds to user requests in the most optimized state.

[0049] Intelligent scheduling engine module: As the core of resource management and task scheduling, the intelligent scheduling engine module adopts advanced scheduling algorithms to achieve intelligent task allocation and efficient resource utilization based on the priority of service requests, resource requirements and system load. The module supports real-time monitoring and dynamic adjustment to ensure the stability and efficiency of system operation.

[0050] Result optimization and integration module: This module focuses on improving the accuracy and usability of model feedback results. By integrating multiple optimization algorithms and strategies, the results of multiple model outputs are intelligently integrated and optimized, redundant information is removed, missing details are supplemented, and the display method is adjusted to provide users with clearer, more accurate and easy-to-understand results.

[0051] Resource dynamic scheduling module: This module is responsible for fine-grained resource allocation and scheduling based on the hardware resource requirements (such as CPU and GPU computing power) during model execution. By dynamically adjusting resource quotas, optimizing resource allocation strategies, and realizing automatic resource recovery and release, this module ensures efficient utilization and maximum conservation of system resources.

[0052] Feedback and evaluation module: As a key link in the interaction between the system and users, the feedback and evaluation module is responsible for presenting the optimized service results to users and collecting user feedback and satisfaction evaluation. By recording and tracking the service process, this module provides important data support for the continuous optimization and performance analysis of the system.

[0053] This system of multi-model service construction and intelligent scheduling methods based on large models achieves accurate and efficient processing of user natural language service requests through highly integrated artificial intelligence technology and refined resource management strategies.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing and intelligently scheduling multi-model services based on a large model, characterized in that: The method for constructing and intelligently scheduling a multi-model service based on a large model includes the following steps: S1. Build a multi-model integration platform; S2, model registration and evaluation; S3, semantic vector decomposition and understanding of user service requests; S4, model selection and priority screening; S5, service response priority sorting and resource intelligent allocation; S6. Feedback result optimization and feedback; S7. Hit labeling and result training.

2. According to claim 1, a method for constructing and intelligently scheduling multi-model services based on a large model is characterized in that: The construction of the multi-model integration platform is: the platform integrates multiple artificial intelligence models including large models, and provides unified services and interfaces.

3. According to claim 1, a method for constructing and intelligently scheduling multi-model services based on a large model is characterized in that: The registration and evaluation of the model are as follows: registering the capabilities, task objectives, resource overhead, and computing types of the model built on the platform, evaluating the capabilities of the model, forming evaluation indicators, capability ranges, and access rules, and fine-tuning the scheduling model rules; An indicator matrix is ​​formed by using capability labeling and weight calculation, and the model evaluation indicators are defined according to the matrix, ultimately forming the indicator score for the model. When multiple models have the same business capability, they can be selected based on the evaluation indicators.

4. According to claim 1, a method for constructing and intelligently scheduling multi-model services based on a large model is characterized in that: The semantic vector decomposition and understanding of the user service request is as follows: when the user makes a service request in the form of natural language dialogue, the system first decomposes the user request into a semantic vector, converts the user's natural language input into a semantic vector that can be understood by the computer, and obtains the specific parameters of the user's request. Then, the system semantically understands the request and clarifies the user's specific needs.

5. According to claim 1, a method for constructing and intelligently scheduling multi-model services based on a large model is characterized in that: The model selection and priority screening is as follows: the system searches and matches the rule base in the platform according to the specific parameters of the user's request, and determines the model with the highest parameter index corresponding to the service requirement. If there are multiple models that meet the index, they are screened according to the preset priority rules and the most appropriate model is selected for service response.

6. The method for constructing and intelligently scheduling multi-model services based on a large model according to claim 1, characterized in that: The service response priority sorting and resource intelligent allocation are as follows: the system prioritizes service responses and establishes a response queue based on factors such as the urgency and importance of user requests. Then, the system intelligently allocates response resources based on the calculation form and resource requirements of the model to ensure that high-priority service requests are processed first while avoiding resource waste and bottlenecks.

7. The method for constructing and intelligently scheduling multi-model services based on a large model according to claim 1, characterized in that: The feedback result optimization and feedback are as follows: obtaining the results of model feedback, and assembling and optimizing the results of multiple model feedbacks; finally, the system feeds back the optimized results to the end user to meet the user's application needs.

8. The method for constructing and intelligently scheduling multi-model services based on a large model according to claim 1, characterized in that: The hit marking and result training are as follows: recording the hit situation of each feedback result, evaluating the customer's satisfaction with the hit result, accurately hitting the platform, marking the hit result, modifying the indicator score of each model in the evaluation index, and completing self-learning and iterative training of model selection.

Images