CN118114771B - Function tool calling method, device, medium and equipment in trusted execution environment - Google Patents

Abstract

The embodiments of this specification disclose a method and apparatus, medium, and equipment for calling a function tool in a trusted execution environment, and relate to the field of computer technology. First, obtain the problem data input by the user, input the problem data into the target large language model deployed in the trusted execution environment, and obtain the function tool call reasoning data corresponding to the problem data; call the target function tool according to the function tool call reasoning data, so as to return the feedback result corresponding to the problem data through the target function tool. The technical solution in the embodiments of this specification can effectively improve the calling efficiency of the large language model for the function tool and improve the processing speed; secondly, it also manages the user's authority to call the function tool, and further improves the security of the data.

Description

Function tool calling method, device, medium and equipment in trusted execution environment

Technical Field

The present disclosure relates to the field of computer technology, and in particular to a method, device, medium, and equipment for calling a function tool in a trusted execution environment.

Background Art

With the rapid development of science and technology, large language models (LLM) have attracted more and more attention. Large language models are deep learning models trained based on massive text data, which can generate natural language text or understand the meaning of language text. Such models can perform a variety of natural language processing (NLP) tasks, including but not limited to text classification, question answering, and dialogue.

Since large language models will memorize a large amount of personal privacy data, in order to ensure that the personal privacy data in the large language model will not be intercepted by attackers, the large language model is generally deployed in a trusted execution environment (TEE). The trusted execution environment TEE is a secure computing environment that can protect the code and data running in it from external attacks, including attacks from the operating system, hardware, and other applications. By creating a memory-isolated execution environment inside the processor, TEE can complete privacy-protected calculations while ensuring computing efficiency.

At present, when a large language model deployed in a trusted execution environment calls a function tool based on user input, it is necessary to submit information of all supported function tools to the large language model, which determines the function tool to call. If the number of function tools is large, the input submitted to the large language model will be longer, resulting in a lower function tool calling speed of the large language model, reducing the processing efficiency of the large language model.

It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of this specification, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field.

Summary of the invention

The embodiments of this specification provide a method for calling a function tool in a trusted execution environment, a device for calling a function tool in a trusted execution environment, a storage medium, and an electronic device, with the aim of effectively improving the efficiency of calling function tools by a large language model and improving the processing speed of a large language model.

According to a first aspect of an embodiment of this specification, a function tool calling method in a trusted execution environment is provided, comprising:

Get input question data;

Inputting the problem data into a target large language model deployed in a trusted execution environment to obtain function tool call reasoning data corresponding to the problem data;

The objective function tool is called according to the inference data called by the function tool, so as to return a feedback result corresponding to the problem data through the objective function tool.

Further, in some implementations, calling the target function tool according to the function tool calling the inference data includes:

Performing authority verification on the target function tool included in the inference data called by the function tool, and determining an authority verification result;

If it is determined that the permission verification result is verified to be passed, the target function tool is called according to the inference data called by the function tool.

Furthermore, in some embodiments, the method further comprises:

If it is determined that the permission verification result is a verification failure, an error prompt indicating that the call is not authorized is returned.

Further, in some implementations, the function tool calling reasoning data includes first identification information of the target function tool;

The performing authority verification on the target function tool included in the inference data called by the function tool and determining the authority verification result includes:

Acquire a user account corresponding to the problem data, and determine a function tool calling permission list corresponding to the user account, wherein the function tool calling permission list includes second identification information of function tools that the user account has permission to call;

The first identification information is matched with the second identification information, and if the match is successful, the authority verification result is determined to be verification passed.

Further, in some implementations, the function tool call reasoning data includes a function parameter list of the target function tool;

The calling of the objective function tool according to the inference data by the function tool, so as to return the feedback result corresponding to the problem data through the objective function tool, includes:

Matching a target function tool according to the first identification information in the inference data called by the function tool;

Generate behavior request data through the target function tool and each function parameter in the function parameter list;

The feedback result corresponding to the problem data is retrieved based on the behavior request data.

Further, in some implementations, the target large language model is obtained through a model training process, and the model training process includes:

Acquire an original data set, wherein the original data set includes sample question data and sample call reasoning data corresponding to the sample question data, wherein the sample call reasoning data includes a sample function tool and a function parameter list that need to be called by the sample question data;

Dividing the original data set to determine a sample training set and an evaluation test set;

Adjusting model parameters of the trained initial large language model using the sample training set to obtain a target large language model;

Performing accuracy evaluation on the target large language model according to the evaluation test set;

If it is determined that the output accuracy of the target large language model is greater than or equal to a preset accuracy threshold, it is determined that the target large language model completes model training.

Furthermore, in some embodiments, the method further comprises:

If it is determined that the output accuracy of the target large language model is less than the accuracy threshold, the following steps are repeatedly performed until the output accuracy of the new target large language model is greater than or equal to the accuracy threshold:

Updating the model parameters of the target large language model through the sample training set to obtain a new target large language model;

The accuracy of the new target large language model is evaluated according to the evaluation test set to determine the output accuracy rate corresponding to the new target large language model.

According to a second aspect of an embodiment of this specification, a function tool calling device in a trusted execution environment is also provided, including:

A question data acquisition module is used to acquire input question data;

A function tool reasoning module, used for inputting the problem data into a target large language model deployed in a trusted execution environment, and obtaining function tool call reasoning data corresponding to the problem data;

The feedback result generating module is used to call the target function tool according to the inference data called by the function tool, so as to return the feedback result corresponding to the problem data through the target function tool.

According to a third aspect of the embodiments of this specification, a computer program product is provided, wherein the computer program product stores at least one instruction, and the at least one instruction is suitable for being loaded by a processor and executing the method steps in the first aspect.

According to a fourth aspect of the embodiments of this specification, a storage medium is provided, wherein the storage medium stores a computer program, wherein the computer program is suitable for being loaded by a processor and executing the method steps in the first aspect.

According to a fifth aspect of the embodiments of this specification, there is provided an electronic device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps in the first aspect.

In the embodiment of the present specification, by obtaining the question data input by the user and inputting the question data into the target large language model deployed in the trusted execution environment, the function tool calling reasoning data corresponding to the question data is obtained; the target function tool is called according to the function tool calling reasoning data, so as to return the feedback result corresponding to the question data through the target function tool. On the one hand, by inputting the question data into the pre-trained target large language model, the function tool calling reasoning data corresponding to the question data can be obtained, and the target function tool is called by the function tool calling reasoning data to obtain the feedback result for answering the question data. Compared with submitting the information of all supported function tools to the large language model for screening, the data processing amount of the large language model can be effectively reduced, the calling efficiency of the target function tool can be improved, and thus the answer processing efficiency of the large language model for the question data can be improved; on the other hand, by selecting and calling the target function tool through the target large language model deployed in the trusted execution environment, without the help of other tools, it is possible to improve the calling efficiency of the function tool while improving the security of the personal privacy data in the target large language model.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the specification, and are used together with the specification to explain the principles of the specification. Obviously, the accompanying drawings described below are only some embodiments of the specification, and for those of ordinary skill in the art, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

FIG1 shows a schematic diagram of a system architecture of an exemplary application environment to which a method and apparatus for calling a function tool in a trusted execution environment according to an embodiment of the present disclosure can be applied.

FIG2 schematically shows a flow chart of a function tool calling method in a trusted execution environment of some embodiments of this specification;

FIG3 schematically shows a flowchart of performing authority verification on a target function tool obtained by inference in some embodiments of this specification;

FIG4 schematically shows a flow chart of a model training process of a target large language model in some embodiments of this specification;

FIG5 schematically shows a flow chart of implementing a function tool call in a trusted execution environment according to some embodiments of the present specification;

FIG6 schematically shows a schematic diagram of a function tool calling device in a trusted execution environment according to some embodiments of the present specification.

FIG. 7 schematically shows a schematic diagram of the structure of a computer system of an electronic device according to some embodiments of the present specification.

FIG. 8 schematically shows a schematic diagram of a computer-readable storage medium according to some embodiments of the present specification.

In the drawings, the same or corresponding reference numerals represent the same or corresponding parts.

DETAILED DESCRIPTION

In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

The terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit this specification. The singular forms "a", "the" and "the" used in this specification are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and/or" used in this article refers to and includes any or all possible combinations of one or more associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of this specification sheet. However, those skilled in the art will appreciate that the technical scheme of this specification sheet can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring the various aspects of this specification sheet.

Furthermore, the drawings are only schematic illustrations and are not necessarily drawn to scale. The block diagrams shown in the drawings are only functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices.

FIG1 shows a schematic diagram of a system architecture of an exemplary application environment to which a method and apparatus for calling a function tool in a trusted execution environment according to an embodiment of the present disclosure can be applied.

As shown in FIG1 , the system architecture 100 may include one or more of terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc. The terminal devices 101, 102, 103 may be various electronic devices with artificial intelligence (AI) computing capabilities, including but not limited to desktop computers, portable computers, smart phones, tablet computers, etc. It should be understood that the number of terminal devices, networks, and servers in FIG1 is merely schematic. According to implementation requirements, there may be any number of terminal devices, networks, and servers. For example, the server 105 may be a server cluster consisting of multiple servers, etc.

The function tool calling method in the trusted execution environment provided by the embodiment of the present disclosure is generally executed by the terminal devices 101, 102, and 103, and accordingly, the function tool calling device in the trusted execution environment is generally set in the terminal devices 101, 102, and 103. However, it is easy for those skilled in the art to understand that the function tool calling method in the trusted execution environment provided by the embodiment of the present disclosure can also be executed by the server 105, and accordingly, the function tool calling device in the trusted execution environment can also be set in the server 105, which is not particularly limited in this exemplary embodiment. For example, the terminal devices 101, 102, and 103 can obtain the question data input by the user, and send the question data to the server 105, and the server 105 inputs the question data into the target large language model deployed in the trusted execution environment, obtains the function tool calling reasoning data corresponding to the question data, and calls the target function tool according to the function tool calling reasoning data to return the feedback result corresponding to the question data, and the server 105 sends the feedback result to the terminal devices 101, 102, and 103, and displays it to the user.

In an embodiment of the present specification, a method for calling a function tool in a trusted execution environment is first provided. The method for calling a function tool in a trusted execution environment can be applied to a terminal device or a server. The present embodiment is not limited thereto. The following is an example of a terminal device executing the method. FIG2 schematically shows a flow chart of a method for calling a function tool in a trusted execution environment according to some embodiments of the present specification. Referring to FIG2, the method for calling a function tool in a trusted execution environment may include the following steps:

Step S210, obtaining input question data;

Step S220, inputting the problem data into a target large language model deployed in a trusted execution environment to obtain function tool call reasoning data corresponding to the problem data;

Step S230, calling the objective function tool according to the inference data called by the function tool, so as to return the feedback result corresponding to the problem data through the objective function tool.

According to the function tool calling method in the trusted execution environment in the embodiments of the present specification, on the one hand, by inputting the question data into the pre-trained target large language model, the function tool calling reasoning data corresponding to the question data can be obtained, and the target function tool is called by the function tool calling the reasoning data to obtain the feedback result for answering the question data. Compared with submitting the information of all supported function tools to the large language model for screening, the data processing amount of the large language model can be effectively reduced, and the calling efficiency of the target function tool can be improved, thereby improving the answer processing efficiency of the large language model for question data; on the other hand, the target function tool can be selected and called by the target large language model deployed in the trusted execution environment, without the help of other tools, so that the security of personal privacy data in the target large language model can be improved while improving the calling efficiency of the function tool.

Next, the function tool calling method in the trusted execution environment in this example embodiment will be further described.

In step S210, input question data is obtained.

In an example embodiment of the present specification, question data refers to instructions generated by a user through an input interface provided in advance. For example, the question data may be an instruction to query weather conditions, such as "What's the weather like today?" The question data may also be an instruction to open path navigation, such as "Navigate to the company." Of course, the question data may also be input instructions for other content, and this example embodiment does not specifically limit this.

A user interaction interface can be provided, and an input box can be provided on the user interaction interface, so that the user can manually input the question data through the input box; a voice recognition module can also be provided to obtain the question data input by the user in the form of voice by recognizing the user's voice instructions; of course, a pre-instruction control can also be provided, and the user can choose to trigger the pre-instruction control to input the question data. This example embodiment does not specifically limit the method for inputting the question data.

In step S220, the problem data is input into a target large language model deployed in a trusted execution environment to obtain function tool call reasoning data corresponding to the problem data.

In an example embodiment of this specification, a trusted execution environment refers to an execution environment obtained by isolating inside a processor, which can protect the code and data running therein from external attacks, including attacks from the operating system, hardware, and other applications. By deploying a large language model in a trusted execution environment, the leakage of personal privacy data contained in the large language model can be effectively avoided, and the security of the large language model during operation can be improved.

The large language model needs to support richer functions. In this case, the large language model will improve the richness of functions by calling function tools. However, due to the closed nature of the trusted execution environment, the large language model deployed in the trusted execution environment needs to submit relevant information of all supported function tools, such as the name, purpose, parameter list, etc. of the function tool, to the large language model at the same time as the user's question when calling the function tool, resulting in too much input data to the large language model, reducing the efficiency of the large language model calling the function tool, and thus reducing the processing speed of the large language model. In the embodiments of this specification, the problem of the efficiency of the large language model calling the function tool is solved by fine-tuning the trained target large language model.

The target big language model refers to a model that fine-tunes and optimizes a pre-trained big language model so that it can infer the function tools needed to call the user's problem data. For example, the pre-trained big language model can be a BERT (Bidirectional Encoder Representations from Transformers) model, and accordingly, the target big language model can be a BERT model; of course, the pre-trained big language model can also be a GPT (Generative Pre-trained Transformer) model, and accordingly, the target big language model can also be a GPT model; it can be understood that the target big language model can also be other types of deep learning models deployed in a trusted execution environment that perform different types of natural language processing tasks (Natural Language Processing, NLP). This example embodiment does not specifically limit the model type of the target big language model.

Function tool call reasoning data refers to information related to the function tool to be called, which is inferred by the target large language model based on the input problem data. For example, the function tool call reasoning data can be the identification information of the function tool to be called, or it can be a parameter list of the function tool to be called. Of course, the function tool call reasoning data can also be other types of information used to call the function tool, and this example embodiment does not make any special limitations on this.

After obtaining the problem data input by the user, the problem data can be input into the target large language model deployed in the trusted execution environment to directly obtain the function tool call inference data corresponding to the problem data, so as to call the required function tool according to the function tool call inference data. It is not necessary to use the relevant information of all supported function tools as input, thereby effectively reducing the amount of data processed by the target large language model.

In step S230, the objective function tool is called according to the inference data called by the function tool, so as to return a feedback result corresponding to the problem data through the objective function tool.

In an example embodiment of the present specification, the objective function tool refers to a function tool obtained by inference of a target large language model and required to be called when replying to question data. For example, the objective function tool may be an application programming interface (Application Programming Interface, API) or a software development kit (Software Development Kit, SDK). This example embodiment does not specifically limit the type of the objective function tool.

For example, suppose the question data input by the user is "How is the weather today?" and the question data is input into the target large language model deployed in the trusted execution environment. The target large language model can understand that the intention of the question data is to ask about today's weather. Then the function tool call inference data output by the target large language model includes the identification information of the weather query API and the parameter list of the weather query API. The weather query API is then matched according to the identification information, and a weather query request is generated in combination with the parameter list of the weather query API. The weather query API can return the feedback result of the weather query with a date of today based on the weather query request, and display the feedback result to the user, completing the function tool call of this question data.

By inputting the question data into the pre-trained target large language model, the function tool calling inference data corresponding to the question data can be obtained, and the function tool calls the inference data to call the target function tool to obtain the feedback result used to answer the question data. Compared with submitting the information of all supported function tools to the large language model for screening, it can effectively reduce the data processing volume of the large language model, improve the calling efficiency of the target function tool, and thus improve the answer processing efficiency of the large language model for question data; by deploying the target large language model in a trusted execution environment to select and call the target function tool, without the help of other tools, it is possible to improve the calling efficiency of the function tool while improving the security of personal privacy data in the target large language model.

Next, steps S210 to S230 are described in detail.

In an example embodiment of the present specification, before calling the target function tool according to the function tool call reasoning data, the target function tool included in the function tool call reasoning data obtained by the target large language model can be authenticated. Referring to FIG3 , it can specifically include:

Step S310, performing authority verification on the target function tool included in the inference data called by the function tool, and determining the authority verification result;

Step S320: If it is determined that the permission verification result is verified to be passed, the target function tool is called according to the inference data of the function tool.

Step S330: If it is determined that the permission verification result is a verification failure, an error prompt indicating that the call is not authorized is returned.

Among them, the permission verification result refers to the verification result used to characterize whether the user has the authority to call the target function tool. For example, if the permission verification result is "True", it may indicate that the verification is successful, that is, the user has the authority to call the function tool to call the target function tool contained in the reasoning data; if the permission verification result is "False", it may indicate that the verification failed, that is, the user does not have the authority to call the function tool to call the target function tool contained in the reasoning data; of course, the permission verification result can also be "0" or "1", where "0" can indicate verification failure and "1" can indicate verification success. This example embodiment does not specifically limit the form of expression of the permission verification result.

When it is determined that the permission verification result is passed, the target function tool can be called according to the inference data of the function tool, and the feedback result corresponding to the problem data is returned through the called target function tool; when it is determined that the permission verification result is failed, an error prompt of unauthorized call can be returned to notify the user in time.

By verifying the permissions of the target function tool contained in the inference data output by the target large language model, the call control of the function tool is realized. Only when the user has the permission to call the relevant function tool, the call of the relevant function tool is allowed to return the feedback result, which can further ensure the security of the data and avoid the possibility of personal privacy data leakage.

Optionally, the function tool call reasoning data may include the first identification information of the target function tool and the function parameter list corresponding to the target function tool. The first identification information refers to the data used to identify the target function tool. For example, the first identification information may be the name of the target function tool, or the unique code of the target function tool, or of course, the calling path of the target function tool. This example embodiment does not specifically limit the type of the first identification information of the target function tool.

A function parameter list refers to a set of input data used to define the specific behavior and content of a function tool request. For example, a function parameter can be a path parameter or a query parameter, and of course it can also be a request body parameter, a request header parameter, an operating system parameter, a software version parameter, etc. The function parameter list can contain a combination of one or more of the above function parameters. This example embodiment does not impose any special restrictions on the specific type and number of function parameters in the function parameter list.

The following steps can be used to verify the permissions of the target function tool contained in the function tool call reasoning data and determine the permission verification result:

The user account corresponding to the problem data can be obtained, and the function tool calling permission list corresponding to the user account can be determined. The function tool calling permission list can include the second identification information of the function tools that the user account has the authority to call; the first identification information can be matched with the second identification information, and if the match is successful, the permission verification result is determined to be verification passed.

The function tool calling permission list refers to a pre-configured set of identifiers of function tools that the user has the permission to call. The second identification information refers to identification information used to identify the function tools that the user account has the permission to call. For example, the second identification information can be the name, unique code, etc. of the function tool that the user account has the permission to call, but this embodiment is not limited to this.

After obtaining the first identification information of the target function tool to be called, the user account of the input problem data is obtained, and the function tool calling permission list corresponding to the user account is retrieved, and the first identification information and the second identification information are matched. If the match is successful, it can be determined that the permission verification result is verified passed.

It can be understood that when there are multiple target function tools to be called, the first identification information corresponding to each target function tool can be matched in turn. If all matches are successful, the authority verification result can be determined as passed; if at least one first identification information is not matched successfully, the authority verification result can be determined as failed.

The permission verification is achieved by organizing the function tools that the user account has the permission to call into a function tool call permission list containing the second identification information, and matching the first identification information in the function tool call inference data output by the target large language model with the second identification information in the function tool call permission list. The calculation amount is small, the permission verification accuracy is high, and the permission verification efficiency is high.

Of course, all supported function tools can also be prioritized and different priorities can be assigned to different user accounts. The user account's calling authority for the function tools can be determined based on the priority. This example embodiment does not specifically limit the method of authority verification.

In an example embodiment of the present specification, the following steps may be performed to implement calling the target function tool according to the inference data of the function tool, so as to return the feedback result corresponding to the problem data through the target function tool, which may specifically include:

The target function tool can be matched according to the first identification information in the inference data called by the function tool, and the behavior request data can be generated through the target function tool and each function parameter in the function parameter list, and then the feedback result corresponding to the problem data can be retrieved based on the behavior request data.

Among them, behavior request data refers to input data used to define the specific behavior and content of the function tool request. For example, the question data may be "How is the weather today?", and the target function tool to be called may be the weather query API. The behavior request data may be a request instruction consisting of a path parameter including the user account ID and a query parameter including query data for the current area and time period. Of course, this is only an illustrative example and should not impose any special limitations on this example embodiment.

After the problem data is input into the target large language model deployed in the trusted execution environment, the target large language model can output the first identification information of the target function tool to be called and the function parameter list input to the target function tool, and then the target function tool can be matched according to the first identification information in the function tool call reasoning data, and behavior request data can be generated through each function parameter in the function parameter list, and then the behavior request data can be input into the target function tool to retrieve the feedback result corresponding to the problem data; in this way, there is no need to use the relevant information of all supported function tools as the input of the target large language model, which can effectively reduce the amount of data that the large language model needs to process, improve the efficiency of the large language model in calling the function tool, and thus improve the processing efficiency of the large language model.

In an exemplary embodiment of the present specification, the target large language model is obtained through a model training process, and the model training process of the target large language model can be implemented through the steps in FIG. 4 , which may specifically include:

Step S410, obtaining an original data set, wherein the original data set includes sample question data and sample call reasoning data corresponding to the sample question data, wherein the sample call reasoning data includes a sample function tool and a function parameter list that need to be called by the sample question data;

Step S420, dividing the original data set to determine a sample training set and an evaluation test set;

Step S430, adjusting model parameters of the trained initial large language model using the sample training set to obtain a target large language model;

Step S440, performing accuracy evaluation on the target large language model according to the evaluation test set;

Step S450: If it is determined that the output accuracy of the target large language model is greater than or equal to a preset accuracy threshold, it is determined that the target large language model has completed model training.

The original data set can be directly obtained from an open source database, or can be obtained by collecting source data and filtering and constructing based on the source data. This example embodiment does not specifically limit the way in which the original data set is constructed. The original data set can include sample problem data and sample call reasoning data corresponding to the sample problem data. The sample call reasoning data can include a sample function tool and a function parameter list that the sample problem data needs to call.

The initial large language model refers to a pre-trained deep learning model that can perform different types of natural language understanding tasks. For example, the initial large language model can be a BERT model or a GPT model. This example embodiment does not specifically limit the type of the initial large language model.

When dividing the original data set, a preset division ratio can be used for division, for example, the division ratio can be 8:2 or 6:3; of course, the data in the original data set can also be divided by random division to obtain a sample training set and an evaluation test set, and this example embodiment does not make any special limitations on this.

The model parameters of the initial large language model can be adjusted through the sample training set constructed by the original data set to obtain the target function tool and function parameter list required to infer and answer the question data; then the target large language model that has completed the preliminary training can be verified and optimized through the evaluation test set constructed by the original data set, so that the target large language model reaches the accuracy index and the training of the target large language model is completed.

The accuracy threshold refers to a parameter set for verifying the optimization target large language model. For example, the accuracy threshold can be 99.5% or 95%. The accuracy threshold can be customized according to actual usage. This example embodiment does not impose any special restrictions on the value of the accuracy threshold.

The sample problem data in the evaluation test set can be input into the target large language model that has been preliminarily trained, and the target function tool and function parameter list can be inferred to output. Then, the target function tool and function parameter list obtained by inference are compared with the function tool and sample function parameter list in the sample call inference data to determine the comparison result. After multiple rounds of evaluation tests, the output accuracy of the current target large language model can be obtained. When the output accuracy is greater than or equal to the accuracy threshold, it can be considered that the trained target large language model meets the use conditions and the training of the target large language model can be completed.

Optionally, when it is determined that the output accuracy of the target large language model is less than the accuracy threshold, the following steps may be repeatedly performed until the output accuracy of the new target large language model updated most recently is greater than or equal to the accuracy threshold, and the training of the target large language model is completed. The repeating steps may include:

The model parameters of the target large language model can be updated through the sample training set to obtain a new target large language model. Then, the accuracy of the new target large language model can be evaluated according to the evaluation test set to determine the output accuracy rate corresponding to the new target large language model.

By dividing the original data set into a sample training set and an evaluation test set, and adjusting the model parameters of the target large language model through the sample training set, and verifying and evaluating the target large language model through the evaluation test set, combined with the accuracy threshold, the target large language model can be continuously adjusted and optimized to obtain a target large language model with better reasoning performance, which can effectively improve the accuracy of the target large language model's call to the objective function tool, thereby improving the accuracy of the answer to the question data.

FIG5 schematically shows a flowchart of implementing a function tool call in a trusted execution environment according to some embodiments of the present specification.

As shown in reference to FIG5 , step S502 is to obtain an original data set, and divide the original data set into a sample training set and an evaluation test set; step S504 is to adjust the model parameters of the initial large language model through the sample training set to obtain a target large language model that has been preliminarily trained; step S506 is to verify and evaluate the target large language model that has been preliminarily trained through the evaluation test set to determine the output accuracy of the target large language model; step S508 is to determine whether the output accuracy is greater than or equal to a preset accuracy threshold, if it is determined that the output accuracy of the target large language model is greater than or equal to the preset accuracy threshold, then execute step S510, otherwise execute step S504; step S510 is to complete the model training of the target large language model; step S512 is to deploy the trained target large language model to a trusted execution environment of a terminal device or a server; step S514 is to obtain the question data input by the user; step S516 is to deploy the question data Input into the target large language model to obtain function tool call reasoning data, the function tool call reasoning data includes the first identification information and function parameter list of the target function tool; step S518, determine the function call permission list according to the user account corresponding to the problem data, the function call permission list includes the second identification information of the function tool with permission to call; step S520, compare the first identification information and the second identification information, and determine the permission verification result according to the comparison result; step S522, determine whether the permission verification result is successful, if it is determined that the permission verification result is successful, execute step S524, otherwise execute step S526; step S524, generate behavior request data according to the target function tool and the function parameter list, and input the behavior request data to the target function tool, so that the target function tool returns the feedback result corresponding to the problem data, and ends the current process; step S526, return an error prompt for unauthorized call, and end the current process.

In summary, by obtaining the question data input by the user and inputting the question data into the target large language model deployed in the trusted execution environment, the function tool call reasoning data corresponding to the question data is obtained; the target function tool is called according to the function tool call reasoning data, so as to return the feedback result corresponding to the question data through the target function tool. On the one hand, by inputting the question data into the pre-trained target large language model, the function tool call reasoning data corresponding to the question data can be obtained, and the target function tool is called by the function tool call reasoning data to obtain the feedback result for answering the question data. Compared with submitting the information of all supported function tools to the large language model for screening, the data processing amount of the large language model can be effectively reduced, the calling efficiency of the target function tool can be improved, and thus the answer processing efficiency of the large language model for question data can be improved; on the other hand, the target function tool can be selected and called through the target large language model deployed in the trusted execution environment, without the help of other tools, so that the security of personal privacy data in the target large language model can be improved while improving the calling efficiency of the function tool.

It should be noted that, although the steps of the method in this specification are described in a specific order in the drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and/or one step may be decomposed into multiple steps, etc.

In addition, in the embodiment of the present specification, a function tool calling device in a trusted execution environment is also provided. Referring to FIG6 , the function tool calling device 600 in the trusted execution environment includes: a problem data acquisition module 610, a function tool reasoning module 620, and a feedback result generation module 630. Among them:

The question data acquisition module 610 is used to acquire input question data;

A function tool reasoning module 620 is used to input the problem data into a target large language model deployed in a trusted execution environment to obtain function tool call reasoning data corresponding to the problem data;

The feedback result generating module 630 is used to call the target function tool according to the inference data called by the function tool, so as to return the feedback result corresponding to the problem data through the target function tool.

In an exemplary embodiment of the present specification, based on the above solution, the feedback result generating module 630 includes:

The function tool permission verification unit can be used to verify the permission of the target function tool included in the function tool calling reasoning data and determine the permission verification result;

The target function tool calling unit may be configured to call the target function tool according to the function tool calling inference data when it is determined that the permission verification result is a passed verification.

In an exemplary embodiment of the present specification, based on the above-mentioned solution, the function tool calling device 600 in the trusted execution environment further includes a verification failure feedback unit, which is configured as follows:

If it is determined that the permission verification result is a verification failure, an error prompt indicating that the call is not authorized is returned.

In an exemplary embodiment of the present specification, based on the aforementioned solution, the function tool calling reasoning data includes the first identification information of the target function tool;

The function tool permission verification unit is configured as:

Acquire a user account corresponding to the problem data, and determine a function tool calling permission list corresponding to the user account, wherein the function tool calling permission list includes second identification information of function tools that the user account has permission to call;

The first identification information is matched with the second identification information, and if the match is successful, the authority verification result is determined to be verification passed.

In an exemplary embodiment of the present specification, based on the above solution, the function tool calling reasoning data includes a function parameter list of the target function tool; the feedback result generation module 630 is configured as follows:

Matching a target function tool according to the first identification information in the inference data called by the function tool;

Generate behavior request data through the target function tool and each function parameter in the function parameter list;

The feedback result corresponding to the problem data is retrieved based on the behavior request data.

In an exemplary embodiment of the present specification, based on the above-mentioned solution, the target large language model is obtained through a model training process, and the function tool calling device 600 in the trusted execution environment also includes a model training module, which includes:

The data acquisition unit can be used to obtain an original data set, wherein the original data set includes sample problem data and sample call reasoning data corresponding to the sample problem data, wherein the sample call reasoning data includes a sample function tool and a function parameter list that need to be called by the sample problem data;

A data set division unit, which can be used to divide the original data set to determine a sample training set and an evaluation test set;

A model parameter adjustment unit, which can be used to adjust the model parameters of the trained initial large language model through the sample training set to obtain a target large language model;

The accuracy evaluation unit can be used to perform accuracy evaluation on the target large language model according to the evaluation test set; if it is determined that the output accuracy of the target large language model is greater than or equal to a preset accuracy threshold, it is determined that the target large language model has completed model training.

In an exemplary embodiment of the present specification, based on the above solution, the model training module further includes a model parameter updating unit, which is configured to:

If it is determined that the output accuracy of the target large language model is less than the accuracy threshold, the following steps are repeatedly performed until the output accuracy of the new target large language model is greater than or equal to the accuracy threshold:

Updating the model parameters of the target large language model through the sample training set to obtain a new target large language model;

The accuracy of the new target large language model is evaluated according to the evaluation test set to determine the output accuracy rate corresponding to the new target large language model.

The specific details of each module of the function tool calling device in the above-mentioned trusted execution environment have been described in detail in the corresponding function tool calling method in the trusted execution environment, so they will not be repeated here.

It should be noted that, although several modules or units of the function tool calling device in the trusted execution environment are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of this specification, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

In addition, in an exemplary embodiment of the present specification, an electronic device capable of implementing the function tool calling method in the above-mentioned trusted execution environment is also provided.

Those skilled in the art will appreciate that various aspects of this specification may be implemented as a system, method, or program product. Therefore, various aspects of this specification may be specifically implemented in the following forms, namely: a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which may be collectively referred to herein as a "circuit," "module," or "system."

The electronic device 700 according to this embodiment of the present specification is described below with reference to Fig. 7. The electronic device 700 shown in Fig. 7 is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present specification.

As shown in FIG7 , the electronic device 700 is in the form of a general computing device. The components of the electronic device 700 may include, but are not limited to: the at least one processing unit 710 , the at least one storage unit 720 , a bus 730 connecting different system components (including the storage unit 720 and the processing unit 710 ), and a display unit 740 .

The storage unit stores a program code, and the program code can be executed by the processing unit 710, so that the processing unit 710 executes the steps described in the "Exemplary Method" section of the present specification according to various exemplary embodiments of the present specification. For example, the processing unit 710 can execute step S210 as shown in FIG2 to obtain input problem data; step S220 to input the problem data into the target large language model deployed in the trusted execution environment to obtain the function tool call reasoning data corresponding to the problem data; step S230 to call the target function tool according to the function tool call reasoning data, so as to return the feedback result corresponding to the problem data through the target function tool.

The storage unit 720 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 721 and/or a cache memory unit 722 , and may further include a read-only memory unit (ROM) 723 .

The storage unit 720 may also include a program/utility 724 having a set (at least one) of program modules 725, such program modules 725 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

Bus 730 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

The electronic device 700 may also communicate with one or more external devices 770 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 700, and/or communicate with any device that enables the electronic device 700 to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication may be performed through an input/output (I/O) interface 750. In addition, the electronic device 700 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and/or public networks, such as the Internet) through a network adapter 760. As shown, the network adapter 760 communicates with other modules of the electronic device 700 through a bus 730. It should be understood that, although not shown in the figure, other hardware and/or software modules may be used in conjunction with the electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of this specification can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of this specification.

Referring to FIG8 , a program product 800 for implementing the function tool calling method in the above-mentioned trusted execution environment according to an embodiment of the present specification is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present specification is not limited thereto, and in this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

The program product may adopt any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

Program code for performing the operations of the present specification may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

In addition, the above-mentioned figures are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present specification, and are not intended to be limiting. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiments of this specification can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of this specification.

Those skilled in the art will readily come up with other embodiments of the specification after considering the specification and practicing the invention disclosed herein. The embodiments of the specification are intended to cover any variations, uses, or adaptations of the technical solutions provided, which follow the general principles of the specification and include common knowledge or customary technical means in the art that are not disclosed in the specification. The specification and embodiments are to be regarded as exemplary only.

It should be understood that the present description is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims (9)

1. A method for calling a function tool in a trusted execution environment, the method comprising: Get input question data; Inputting the problem data into a target large language model deployed in a trusted execution environment to obtain function tool call reasoning data corresponding to the problem data; Calling the objective function tool according to the inference data of the function tool, so as to return the feedback result corresponding to the problem data through the objective function tool; The target large language model is obtained through a model training process, and the model training process includes: Acquire an original data set, wherein the original data set includes sample question data and sample call reasoning data corresponding to the sample question data, wherein the sample call reasoning data includes a sample function tool and a function parameter list that need to be called by the sample question data; Dividing the original data set to determine a sample training set and an evaluation test set; Adjusting model parameters of the trained initial large language model using the sample training set to obtain a target large language model; Performing accuracy evaluation on the target large language model according to the evaluation test set; If it is determined that the output accuracy of the target large language model is greater than or equal to a preset accuracy threshold, it is determined that the target large language model completes model training. 2. The function tool calling method according to claim 1, wherein calling the target function tool according to the function tool calling inference data comprises: Performing authority verification on the target function tool included in the inference data called by the function tool, and determining an authority verification result; If it is determined that the permission verification result is verified to be passed, the target function tool is called according to the inference data called by the function tool. 3. The function tool calling method according to claim 2, further comprising: If it is determined that the permission verification result is a verification failure, an error prompt indicating that the call is not authorized is returned. 4. The function tool calling method according to claim 2, wherein the function tool calling reasoning data comprises first identification information of the target function tool; The performing authority verification on the target function tool included in the inference data called by the function tool and determining the authority verification result includes: Acquire a user account corresponding to the problem data, and determine a function tool calling permission list corresponding to the user account, wherein the function tool calling permission list includes second identification information of function tools that the user account has permission to call; The first identification information is matched with the second identification information, and if the match is successful, the authority verification result is determined to be verification passed. 5. The function tool calling method according to claim 1, wherein the function tool calling reasoning data comprises a function parameter list of the target function tool; The calling of the objective function tool according to the inference data by the function tool, so as to return the feedback result corresponding to the problem data through the objective function tool, includes: Matching a target function tool according to the first identification information in the inference data called by the function tool; Generate behavior request data through the target function tool and each function parameter in the function parameter list; The feedback result corresponding to the problem data is retrieved based on the behavior request data. 6. The function tool calling method according to claim 1, further comprising: If it is determined that the output accuracy of the target large language model is less than the accuracy threshold, the following steps are repeatedly performed until the output accuracy of the new target large language model is greater than or equal to the accuracy threshold: Updating the model parameters of the target large language model through the sample training set to obtain a new target large language model; The accuracy of the new target large language model is evaluated according to the evaluation test set to determine the output accuracy rate corresponding to the new target large language model. 7. A function tool calling device in a trusted execution environment, comprising: A question data acquisition module is used to acquire input question data; A function tool reasoning module, used for inputting the problem data into a target large language model deployed in a trusted execution environment, and obtaining function tool call reasoning data corresponding to the problem data; A feedback result generating module, used for calling the target function tool according to the inference data called by the function tool, so as to return the feedback result corresponding to the problem data through the target function tool; The target large language model is obtained through a model training process, and the model training process includes: Acquire an original data set, wherein the original data set includes sample question data and sample call reasoning data corresponding to the sample question data, wherein the sample call reasoning data includes a sample function tool and a function parameter list that need to be called by the sample question data; Dividing the original data set to determine a sample training set and an evaluation test set; Adjusting model parameters of the trained initial large language model using the sample training set to obtain a target large language model; Performing accuracy evaluation on the target large language model according to the evaluation test set; If it is determined that the output accuracy of the target large language model is greater than or equal to a preset accuracy threshold, it is determined that the target large language model completes model training. 8. A storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method according to any one of claims 1 to 6 when executed by a processor. 9. An electronic device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the method as claimed in any one of claims 1 to 6.