Large language model public domain and private domain combined control system and combined control method

Abstract

The invention discloses a public domain and private domain combined control system and method for a large language model. When the large language model system receives the session, the session content is forwarded to the large language model agent module; and the big language model agent module forwards the session content to a plurality of third-party big language model systems determined according to the service agreement, and then returns the session response of each third-party big language model system to the big language model system. And the big language model system takes the session response of each third-party big language model system as training data for training, and makes a response according to the received session content after the training is completed. According to the method, responses output by other large language models are directly used as training data sources, the maturation speed of enterprise private large language models is greatly increased, the cost needed by model training is reduced on a large scale, and the training data sources of enterprise private large language models are greatly enriched under the condition that the business concentration degree of enterprises is guaranteed.

Description

Technical Field

[0001] The present invention relates to system architecture technology, in particular to the architecture technology of a joint control system of a private domain and a public domain of a large language model. Background Art

[0002] With the success of Chat-GPT, large language model systems have sprung up. Many companies, especially public Internet companies, are developing their own proprietary large language models. The large language models developed by these Internet companies usually provide professional and conventional artificial intelligence services to the society and the public. In addition to the company's proprietary large language models, open source large language models have also appeared in large numbers. Open source large language models, such as GPT-NeoX-20B, GPT-J-6b, LLaMA, Bard Nano, Mistral, MPT-7B, BLOOM, OPT-175B, XGen-7B, etc. As a result, ordinary companies can easily build their own large language model systems based on these open source large language models.

[0003] When ordinary enterprises build their large language model systems based on these open source large language models, in addition to the usual factors such as licensing restrictions, capital costs, and suitable models, they must also consider the data source required for model training. Because the accuracy of large language models in processing problems depends largely on the data source for their training. Although there are pre-trained models for existing open source large language models. However, for enterprises, the problems they need to deal with are usually relatively professional and require relatively professional data models, while pre-trained models are more versatile than professional. Enterprise professional model systems need to have enough professional training data sources. However, due to their limited size, ordinary enterprises find it difficult to have enough professional training data sources. This leads to the lack of sufficient data training for the model, making it difficult for the accuracy of its problem processing to meet its usage requirements. Summary of the invention

[0004] The problem to be solved by the present invention is: when an enterprise builds a large language model system, there is a lack of sufficient training data, especially the problem of lack of training data related to the enterprise's professional business.

[0005] To solve the above problems, the solution adopted by the present invention is as follows: According to a large language model public domain private domain joint control system of the present invention, the system includes a first-party conversation service module, a first-party large language model system and a large language model proxy module; the first-party conversation service module is connected to the first-party large language model system and the large language model proxy module; In the first-party conversation service module: when receiving a conversation from the client, forwarding the conversation content to the large language model proxy module according to pre-configuration; The large language model proxy module: when receiving the conversation content forwarded from the first-party conversation service module, forward the conversation content to a plurality of third-party large language model systems determined according to the service agreement, and send the conversation response of each third-party large language model system to the first-party large language model system; The first-party large language model system is trained using the conversation responses of each third-party large language model system as training data. After the training is completed, a response is made according to the conversation content received by the first-party conversation service module, and the conversation response is returned to the client through the first-party conversation service module.

[0006] Further, according to the joint control system of the present invention, in the first-party conversation service module, the pre-configuration is set according to a client request.

[0007] Further, according to the joint control system of the present invention, the pre-configuration includes a service type; in the first-party conversation service module, when the conversation content is forwarded to the large language model proxy module, the service type is attached; in the large language model proxy module, a third-party large language model system is determined according to the attached service type.

[0008] Further, according to the joint control system of the present invention, in the large language model proxy module, when receiving the conversation content forwarded by the first-party conversation service module, the conversation content is decomposed by keywords, and then the professional field is determined based on the keywords, and then the third-party large language model system is determined based on the determined professional field.

[0009] Further, according to the joint control system of the present invention, in the first-party conversation service module, when receiving the conversation from the client, the conversation content is decomposed by keywords, and then the service type is determined based on the keywords, and when the conversation content is forwarded to the large language model proxy module, the service type is attached; in the large language model proxy module, the third-party large language model system is determined based on the attached service type.

[0010] Further, according to the joint control system of the present invention, the first-party large language model system stores the conversation responses of each third-party large language model system into a training data backup library.

[0011] Further, according to the joint control system of the present invention, the first-party conversation service module and the first-party large language model system are configured in the first-party intranet; the large language model proxy module is configured in the external network; the system also includes a first-party gateway connecting the first-party intranet and the external network; in the large language model proxy module, the conversation responses of each third-party large language model system are directly sent to the first-party large language model system through the first-party gateway.

[0012] Further, according to the joint control system of the present invention, in the large language model proxy module, when receiving the conversation responses of each third-party large language model system, the conversation responses of each third-party large language model system are returned to the first-party conversation service module, and the first-party conversation service module submits training to the first-party large language model system. After the training is completed, the first-party conversation service module submits the conversation content to the first-party large language model system and receives the returned conversation response, and then the first-party conversation service module sends the returned conversation response to the client.

[0013] Further, according to the joint control system of the present invention, the first-party conversation service module stores the conversation responses of each third-party large language model system into a training data backup library.

[0014] According to a large language model public domain and private domain joint control method of the present invention, the method comprises the following steps: When the first-party large language model system receives a session from the client, the first-party large language model system forwards the session content to the large language model proxy module according to a pre-configuration; When the large language model proxy module receives the conversation content forwarded from the first-party large language model system, the large language model proxy module forwards the conversation content to a number of third-party large language model systems determined according to the service agreement, and sends the conversation response of each third-party large language model system to the first-party large language model system; The first-party large language model system is trained using the conversation responses of each third-party large language model system as training data. After the training is completed, a response is made according to the conversation content received by the first-party conversation service module, and the conversation response is returned to the client through the first-party conversation service module.

[0015] The technical effects of the present invention are as follows: Through the system of the present invention, an enterprise can greatly enrich the training data source of the enterprise's private large language model while ensuring the focus of the enterprise's business.

[0016] The system of the present invention directly uses the output responses of other large language models as the source of training data, which greatly accelerates the maturity of the enterprise's private large language model and greatly reduces the cost required for model training.

[0017] The system large language model proxy module of the present invention provides a unified service proxy for other large language models, providing a platform convenience for the enterprise's private large language model to provide artificial intelligence paid services to the outside world.

[0018] The present invention can reduce data transmission costs and improve response speed while ensuring the security of the enterprise's private large language model through the cooperation of a dedicated physically isolated gateway. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the overall structure of an embodiment of the combined control system of the present invention.

[0020] Among them, MD represents the training data backup library, MM represents the large language model system, MG represents the gateway, MS represents the session service module, and MA represents the large language model proxy service module. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below with reference to the accompanying drawings.

[0022] Figure 1 An example of an association relationship architecture of a large language model public-private joint control system is provided. The joint control system is used to train a private enterprise large language model by combining a large language model that is already available on the public network and is publicly available. The large language model that is already available on the public network and is publicly available belongs to the large language model in the public domain, and the private enterprise large language model belongs to the private domain. Therefore, the joint control system of the present invention is also called a large language model public-private joint control system.

[0023] The relationship structure involves three parties: the first party, the agent, and the third party. The first party is the enterprise, and the third party is the provider of artificial intelligence services that provide large language models on the public Internet.

[0024] The facilities provided by the first party include a conversation service module MS, an enterprise private large language model system MM, a training data backup library MD and a gateway MG. Among them, the conversation service module MS, the large language model system MM, and the gateway MG are respectively referred to as the first-party conversation service module, the first-party large language model system, and the first-party gateway in the present invention. Correspondingly, the enterprise intranet is referred to as the first-party intranet in the present invention. The first-party conversation service module, the first-party large language model system, and the training data backup library MD are configured in the first-party intranet and are connected through the intranet network. The first-party gateway is used to connect the first-party intranet and the external network. The external network is also the public network. The first-party conversation service module provides a conversation interface for the first-party large language model system, that is, when the first-party enterprise intranet user needs the large language model to provide artificial intelligence services, the user connects to the first-party conversation service module through the client of the enterprise intranet, and accesses the first-party large language model system through the first-party conversation service module. The first-party conversation service module, the first-party large language model system, and the training data backup library MD are implemented by the server by executing a computer program instruction set. The first-party conversation service module, the first-party large language model system and the training data backup library MD can be deployed on different servers or server clusters, or in the same server or server cluster. The first-party gateway is implemented by the device hardware cooperating with the execution of a computer program instruction set.

[0025] The facility provided by the agent is a large language model agent service platform. The large language model agent service platform is a comprehensive service platform, which is usually configured on the public network, connecting the first-party conversation service module and the first-party large language model system through the public network and the first gateway, and connecting various third-party large language model systems through the public network.

[0026] There are two types of third-party big language model systems: the first is the enterprise professional big language model, and the second is the public network general big language model. The enterprise professional big language model comes from the big language model of each enterprise's own architecture. Its big language model has a strong correlation with the business of each enterprise, and is highly professional and precise. The public network general big language model comes from the general big language model for the general public on the Internet, such as the existing big language models Chart-GPT, Baidu Wenxin, Ali Tongyi, etc. This type of big language model needs to be oriented to the general public of the Internet, with strong versatility but lack of professionalism.

[0027] The large language model proxy service platform provides proxy services and paid support for these third-party large language model systems. In order to manage the proxy services of these large language models, the large language model proxy service platform needs to classify these large language models according to their characteristics: for the first type of enterprise professional large language models, they are classified according to the enterprise business type and field; for the public network general large language models, they are divided into free and paid categories. The free public network general large language models refer to those large language models that provide free services to Internet users.

[0028] In this embodiment, the large language model proxy module referred to in the present invention is a module in the large language model proxy service platform, which is implemented by executing a computer program instruction set through a server.

[0029] The large language model public-domain and private-domain joint control system of this embodiment includes at least a first-party conversation service module, a first-party large language model system, and a large language model proxy module. The cooperation between the components is as follows: In the first-party conversation service module: when receiving the client's conversation, the conversation content is forwarded to the large language model proxy module according to the pre-configuration.

[0030] The client session comes from the request sent by the client. How the client sends the request, how the client establishes a connection with the first-party session service module, or how the first-party service module performs user identity authentication on the client is not within the scope of the present invention and need not be elaborated in this specification.

[0031] The pre-configuration here can be a configuration file, or can be pre-specified by the connected client. In this embodiment, the pre-configuration includes at least two items: third-party model auxiliary options and service types.

[0032] The third-party model assistance option is a Boolean value. If the third-party model assistance option is false, the first-party session service module directly submits the received client session content to the first-party large language model system, which then responds to the client; if the model assistance option is true, the first-party session service module forwards the session content to the large language model proxy module MA.

[0033] When forwarding the session content to the large language model proxy module MA, the service type may be included or not. The service type here is determined by the service agreement between the enterprise as a user of the large language model proxy service platform and the large language model proxy service platform. Different enterprises and large language model proxy service platforms have different service agreements, different service types, and corresponding service contents.

[0034] In the large language model proxy module MA: when receiving the conversation content forwarded from the first-party conversation service module, the conversation content is forwarded to several third-party large language model systems determined according to the service agreement, and the conversation responses of each third-party large language model system are sent to the first-party large language model system.

[0035] The above process requires the large language model proxy service platform to perform enterprise user identity authentication on the connection of the first-party conversation service module. The first-party conversation service module establishes a connection with the large language model proxy service platform, and the large language model proxy service platform performs enterprise user identity authentication on the connection of the first-party conversation service module is not within the scope of the present invention and need not be elaborated.

[0036] The service agreement here corresponds to the aforementioned service type. In this embodiment: By default, the first-party conversation service module does not attach a service type when forwarding the conversation content to the large language model proxy module MA. At this time, the large language model proxy module MA uses the aforementioned free public network universal large language model as the determined third-party large language model systems.

[0037] If the first-party conversation service module forwards the conversation content to the large language model proxy module MA with the service type: If the service type is a general paid type. The large language model proxy module MA uses several or one paid public network general large language models agreed upon by the enterprise user and the large language model proxy service platform as the determined several third-party large language model systems. That is to say, in this case, the large language model proxy module MA needs to determine the paid public network general large language model agreed upon by the service type. The paid public network general large language models agreed upon by different enterprises and the large language model proxy service platform are different.

[0038] If the service type is professional-related. The large language model proxy module MA uses the enterprise professional large language models corresponding to the several professional fields agreed upon by the enterprise user and the large language model proxy service platform as the determined several third-party large language model systems. That is to say, in this case, the large language model proxy module MA needs to find out the professional field of the service agreement according to the service type, and then determine the enterprise professional large language model according to the professional field. The professional fields agreed upon by different enterprises and the large language model proxy service platform are different.

[0039] In another optional implementation, if the service type is professional-related and the enterprise user and the large language model proxy service platform have not agreed on a professional field, the large language model proxy module first decomposes the conversation content by keywords, then determines the professional field based on the keywords, and finally determines the third-party large language model system based on the professional field determined by the keywords.

[0040] In another optional implementation, if the enterprise user is a high-level user, the first-party conversation service module does not need to include the service type when forwarding the conversation content to the large language model proxy module MA. The large language model proxy module also performs keyword decomposition on the conversation content, and then determines the professional field based on the keywords, and finally determines the third-party large language model system based on the professional field determined by the keywords.

[0041] When determining a third-party large language model system based on the professional field determined by the keywords, if the determined professional field can find a corresponding enterprise professional large language model, the corresponding enterprise professional large language model will be used as the determined third-party large language model systems; if the professional field does not correspond to an enterprise professional large language model, the paid public network general large language model agreed upon in the service will be used as the determined third-party large language model systems.

[0042] Taking into account that the above-mentioned service types are quite complex, in another optional implementation, before the first-party conversation service module forwards the conversation content to the large language model proxy module MA, the first-party conversation service module decomposes the conversation content by keywords, and then determines the service type based on the keywords, and then determines the service type based on the keywords when forwarding the conversation content to the large language model proxy module MA.

[0043] In the first-party large language model system: the first-party large language model system is trained with the conversation responses of each third-party large language model system as training data. After the training is completed, a response is made according to the conversation content received by the first-party conversation service module, and the conversation response is returned to the client through the first-party conversation service module.

[0044] In this embodiment, preferably, after completing the training, the first-party large language model system also stores the conversation responses of each third-party large language model system into the training data backup library.

[0045] The large language model proxy module MA sends the conversation responses of each third-party large language model system to the first-party large language model system usually via the first-party conversation service module. That is, the large language model proxy module MA returns the conversation responses of each third-party large language model system to the first-party conversation service module, and then the first-party conversation service module submits the training to the first-party large language model system. After the training is completed, the first-party conversation service module submits the conversation content to the first-party large language model system and receives the returned conversation response, and then the first-party conversation service module sends the returned conversation response to the client.

[0046] In this embodiment, the first-party gateway is a specially customized gateway that cooperates with the large language model proxy module. When the large language model proxy module MA sends the conversation responses of each third-party large language model system to the first-party large language model system, the conversation responses of each third-party large language model system are directly sent to the first-party large language model system through the first-party gateway.

[0047] In addition, it should be pointed out that the training data backup library and the storage of the conversation responses of each third-party large language model system in the training data backup library are optional. The existence of the training data backup library is only to reserve training data resources for the later reconstruction of the first-party large language model system.

[0048] In addition, the first-party large language model system may also store the conversation responses of each third-party large language model system into the training data backup library through the first-party conversation service module, especially under the implementation mode in which the large language model proxy module MA sends the conversation responses of each third-party large language model system to the first-party large language model system usually via the first-party conversation service module. When the large language model proxy module MA returns the conversation responses of each third-party large language model system to the first-party conversation service module, the first-party conversation service module may store the conversation responses of each third-party large language model system into the training data backup library.

[0049] In addition, it should be pointed out that the coordination and interaction process between the components of the joint control system of the present invention can also be used as a method process, and the functions implemented by each module component can be used as steps of the method, and they correspond one to one. This method is the public-private joint control method of the large language model referred to in the present invention.

[0050] In addition, the first-party conversation service module, as a conversation access module of the first-party large language model system, can be regarded as a part of the first-party large language model system in a broader sense. Therefore, the cooperation performed by the above-mentioned first-party conversation service module can be regarded as a function included in the first-party large language model system in a broader sense. The present invention separates the first-party conversation service module from the first-party large language model system only for the convenience of description, and the essential interaction process remains unchanged.

Claims

1. A large language model public domain private domain joint control system, characterized in that: The system includes a first-party conversation service module, a first-party large language model system and a large language model proxy module; the first-party conversation service module is connected to the first-party large language model system and the large language model proxy module; In the first-party conversation service module: when receiving a conversation from the client, forwarding the conversation content to the large language model proxy module according to pre-configuration; The large language model proxy module: when receiving the conversation content forwarded from the first-party conversation service module, forward the conversation content to a plurality of third-party large language model systems determined according to the service agreement, and send the conversation response of each third-party large language model system to the first-party large language model system; The first-party large language model system is trained using the conversation responses of each third-party large language model system as training data. After the training is completed, a response is made according to the conversation content received by the first-party conversation service module, and the conversation response is returned to the client through the first-party conversation service module.

2. The combined control system according to claim 1, characterized in that: In the first-party conversation service module, the pre-configuration is set according to a client request.

3. The combined control system according to claim 2, characterized in that: The pre-configuration includes a service type; in the first-party conversation service module, when forwarding the conversation content to the large language model proxy module, the service type is attached; in the large language model proxy module, a third-party large language model system is determined according to the attached service type.

4. The combined control system according to claim 1, characterized in that: In the large language model proxy module, when receiving the conversation content forwarded by the first-party conversation service module, the conversation content is decomposed by keywords, and then the professional field is determined according to the keywords, and then the third-party large language model system is determined according to the determined professional field.

5. The combined control system according to claim 1, characterized in that: In the first-party conversation service module, when a conversation from a client is received, the conversation content is decomposed by keywords, and then the service type is determined based on the keywords. When the conversation content is forwarded to the large language model proxy module, the service type is attached. In the large language model proxy module, a third-party large language model system is determined based on the attached service type.

6. The combined control system according to claim 1, characterized in that: The first-party large language model system stores the conversation responses of each third-party large language model system in a training data backup library.

7. The combined control system according to any one of claims 1 to 6, characterized in that: The first-party conversation service module and the first-party large language model system are configured in the first-party intranet; the large language model proxy module is configured in the external network; the system also includes a first-party gateway connecting the first-party intranet and the external network; in the large language model proxy module, the conversation responses of each third-party large language model system are directly sent to the first-party large language model system through the first-party gateway.

8. The combined control system according to any one of claims 1 to 5, characterized in that: In the large language model proxy module, when the conversation responses of each third-party large language model system are received, the conversation responses of each third-party large language model system are returned to the first-party conversation service module, and the first-party conversation service module submits training to the first-party large language model system. After the training is completed, the first-party conversation service module submits the conversation content to the first-party large language model system and receives the returned conversation response, and then the first-party conversation service module sends the returned conversation response to the client.

9. The combined control system according to claim 8, characterized in that: The first-party conversation service module stores the conversation responses of each third-party large language model system into a training data backup library.

10. A method for joint control of public and private domains of a large language model, characterized in that: The method comprises the following steps: When the first-party large language model system receives a session from the client, the first-party large language model system forwards the session content to the large language model proxy module according to a pre-configuration; When the large language model proxy module receives the conversation content forwarded from the first-party large language model system, the large language model proxy module forwards the conversation content to a number of third-party large language model systems determined according to the service agreement, and sends the conversation response of each third-party large language model system to the first-party large language model system; The first-party large language model system is trained using the conversation responses of each third-party large language model system as training data. After the training is completed, a response is made according to the conversation content received by the first-party conversation service module, and the conversation response is returned to the client through the first-party conversation service module.

Images