Artificial intelligence platform interaction method based on large language model and intelligent agent

Abstract

The invention discloses an artificial intelligence platform interaction method based on a large language model and an intelligent agent. The method mainly comprises three parts of artificial intelligence platform script language development, intelligent agent construction and text conversion operation. The method comprises the following steps: firstly, developing a scripting language of an artificial intelligence platform, and defining and packaging mouse and keyboard operations such as dragging, pulling, adding, deleting and modifying nodes and connecting lines of the artificial intelligence platform into the scripting language in order to realize the interaction capability of subsequent text conversion operation; thirdly, semantic analysis is conducted on the platform requirements expressed by the natural language, corresponding operation steps and function tools needing to be called in each step are inferred, and the function tools are executed; and finally, the user reviews and corrects a text conversion operation result, and feeds back a correction result to the background database. According to the method, an intelligent interaction method of a text conversion operation user and a platform is realized; and the intelligent interaction method for completing the text conversion operation realizes effective enabling of a leading-edge large language model technology on a traditional artificial intelligence platform.

Description

Technical Field

[0001] The present invention relates to the field of intelligent interaction technology, and in particular to an artificial intelligence platform interaction method based on a large language model and an intelligent agent. Background Art

[0002] An artificial intelligence platform is a system that integrates various AI technologies and tools, designed to support data analysis, training and deployment of machine learning models, and the execution of automated tasks. The artificial intelligence platform includes a variety of feature engineering, machine learning, deep learning, and industry vertical algorithms. The use of the platform requires users to have high technical and industry backgrounds. Therefore, the platform interaction method based on natural language dialogue is of great significance to lowering the threshold for the use of artificial intelligence platforms and expanding the customer base of artificial intelligence platforms.

[0003] The intelligence of the platform is reflected in two dimensions: first, intelligent operators, which automatically extract valuable information and patterns from massive data through advanced machine learning and deep learning algorithms, and improve the accuracy and efficiency of model prediction and decision-making. Second, intelligent operations, which help users easily use the artificial intelligence platform through natural language interaction and achieve rapid development and deployment of high-precision models.

[0004] At present, artificial intelligence platforms mainly enhance the intelligence of the platforms from the perspective of improving operator accuracy and efficiency, but ignore the interactive intelligence of the platforms. Summary of the invention

[0005] In order to solve the technical problems existing in the background technology, the present invention proposes an artificial intelligence platform interaction method based on a large language model and an intelligent agent.

[0006] 8. The present invention proposes an artificial intelligence platform interaction method based on a large language model and an intelligent agent, comprising the following steps:

[0007] S1: Develop a scripting language for an AI platform to form a tool set, which includes function tools;

[0008] S2: Build a ReAct intelligent agent, which performs semantic analysis on the platform requirements expressed in natural language, infers the operation steps of the natural language requirements and the function tool formed in step 1 that needs to be called in each step to complete the text conversion operation;

[0009] S3, manually revise and improve the text conversion operation, record the revision and improve the revision operation into a script and store it in the tool set in S1;

[0010] S4. Output the revised text of S3 through the large language model.

[0011] As a further optimized solution of the present invention, in step S1, the mouse and keyboard operation definitions such as dragging, dropping, adding, deleting, and modifying the nodes and lines of the artificial intelligence platform are encapsulated into a scripting language.

[0012] As a further optimized solution of the present invention, the script functions in step S1 include: adding node function, deleting node function, node connection function, and defining data source function.

[0013] As a further optimized solution of the present invention, step S2 includes:

[0014] S20. Collect basic information of the script function tool in step 1, including: function name, function description, parameter name, parameter type, and parameter description;

[0015] S21. Use prompt engineering to guide large language models to derive tools needed to achieve natural language requirements;

[0016] S22, extracting key information and features of the natural language expressing the platform requirements, calling the corresponding function tool in step 1, and obtaining the tool calculation results;

[0017] S23. Based on the calculation result of S22, feedback in the prompting project is constructed, and the prompting project guides the large language model to derive the next function tool to be used to realize the natural language demand;

[0018] S24, looping through S22 and S23 until the tool selected by the large language model is the task end placeholder "Finish", indicating that the text conversion task is completed.

[0019] As a further optimization solution of the present invention, in step S22: the engineering guidance large language model is prompted to deduce the tools needed to implement natural language requirements according to the idea of ​​"problem>thinking>selecting tools>inputting tool parameters".

[0020] As a further optimization solution of the present invention, in step S23: the engineering guidance large language model is prompted to deduce the next function tool needed to implement the natural language requirement according to the idea of ​​"feedback>thinking>select tool>input tool parameters>tool results".

[0021] As a further optimization scheme of the present invention, the prompting project includes thinking, action, action input and observation.

[0022] Preferably, the intelligent agent in step S2 is built based on the open source large language model Qwen2.5-72B.

[0023] The present invention proposes a method for intelligent interaction of an artificial intelligence platform, which mainly includes three parts: development of an artificial intelligence platform script language, construction of an intelligent agent, and text conversion operation (Text2Operation). First, the script language of the artificial intelligence platform is developed. In order to realize the interactive ability of the subsequent Text2Operation, the mouse and keyboard operation definitions such as dragging, pulling, adding, deleting, and modifying nodes and lines of the artificial intelligence platform are encapsulated into a script language; then, based on the open source large language model Qwen2.5-72B, an intelligent agent (Agent) of collaborative reasoning + action (ReAct) is constructed, and the platform requirements expressed in natural language are semantically analyzed, and the corresponding operation steps and the function tools (Tools) that need to be called in each step are inferred, and the function tools are executed; finally, the user reviews and corrects the Text2Operation results, and feeds back the correction results to the background database, and uses historical data to improve the accuracy of Text2Operation. This method implements the intelligent interaction method between Text2Operation users and the platform: develops the scripting language of the artificial intelligence platform, builds the underlying foundation of the Text2Operation intelligent interaction method, builds the ReAct intelligent agent, infers the operation logic expressed in natural language, calls built-in tools, and completes the Text2Operation intelligent interaction method, realizing the effective empowerment of traditional artificial intelligence platforms with cutting-edge large language model technology.

[0024] In the present invention, the proposed artificial intelligence platform interaction method based on a large language model and an intelligent agent realizes the effective combination of cutting-edge large language model technology and a traditional artificial intelligence platform. Users only need to input natural language requirements and the artificial intelligence platform can output the model construction required by the user.

[0025] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Implementation flow chart of the present invention;

[0027] Figure 2 A schematic diagram of the Transformer architecture on which the large language model of the present invention is based;

[0028] Figure 3 The main datasets constructed for modeling intelligent platforms in some inventions;

[0029] Figure 4 This is the homepage of the smart platform constructed using the present method in some embodiments. DETAILED DESCRIPTION

[0030] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0031] like Figure 1-Figure 2 The artificial intelligence platform interaction method based on a large language model and an intelligent agent shown includes the following steps:

[0032] S1: Develop a scripting language for the artificial intelligence platform to form a tool set; encapsulate the mouse and keyboard operation definitions such as dragging, dropping, adding, deleting, and modifying nodes and connections of the artificial intelligence platform into a scripting language;

[0033] The script functions that need to be developed include: adding nodes, deleting nodes, connecting nodes, defining data sources, etc. Since the hyperparameters of each node are different and there are many of them, we do not develop the node hyperparameter definition function, but preset appropriate values ​​for each hyperparameter. In theory, you can run it directly after adding nodes, connecting nodes, and defining data sources. Figure 3 Shown are the main data sets constructed by the modeling intelligence platform in some embodiments.

[0034] S2: Build a ReAct (Reasoning and Acting) intelligent agent, which performs semantic analysis on the platform requirements expressed in natural language, and infers the operation steps and function tools that need to be called at each step.

[0035] Specifically, the intelligent agent is built based on the open source large language model Qwen2.5-72B;

[0036] like Figure 1 As shown, the implementation steps of ReAct are as follows:

[0037] S20. Collect basic information of the script function tool in step 1, including: function name, function description, parameter name, parameter type, parameter description, etc.;

[0038] S21. Use the prompt engineering to guide the large language model to deduce the tools needed to achieve natural language requirements according to the idea of ​​"question > thinking > selecting tools > inputting tool parameters";

[0039] S22, extracting key information and features of the natural language expressing the platform requirements, calling the corresponding function tool in step 1, and obtaining the tool calculation results;

[0040] S23. Based on the calculation results of S22, feedback in the prompt project is constructed, and the prompt project guides the large language model to deduce the next function tool needed to implement the natural language requirements according to the idea of ​​"feedback > thinking > selecting tools > inputting tool parameters > tool results";

[0041] S24, looping through S22 and S23 until the tool selected by the large language model is the task end placeholder "Finish", indicating that the Text2Operation task is completed;

[0042] S3. Manually revise and improve the Text2Operation results, and record the script functions corresponding to the mouse operations; record the improved script functions into the database of the tool set;

[0043] The recorded user-perfected Text2Operation results will be added to the prompt project in the form of examples when subsequent users perform similar requirements, improving the accuracy of the ReAct intelligent agent;

[0044] The method further includes S4: outputting the modified calculation result through the large language model.

[0045] Like the existing large language model, this large language model is designed based on the Transformer architecture.

[0046] like Figure 2 As shown, the Transformer architecture includes an encoder and a decoder. The encoder is Figure 2 The left side is used to process the input sequence, and the decoder is Figure 2 The right side of is used to generate the output sequence;

[0047] The encoder and decoder are composed of 6 (N x =N y =6) are stacked with the same layers. Each layer of the encoder contains a multi-head attention layer and a feedforward network layer (FNN), which are connected using a residual mechanism and a normalization operation. In addition to the two layers of the encoder, the decoder inserts a multi-head attention layer to fuse the output of the encoder with the input of the decoder. The model adds an upper triangular mask matrix to the attention part of the decoder to prevent information leakage during model training and ensure that the model is not affected by the subsequent information when calculating the information at the current position.

[0048] The multi-head attention layer consists of multiple self-attention mechanisms with scaled dot product attention. The attention mechanism maps the query and a set of key-value pairs to a high-dimensional space, that is, it performs a weighted sum of the values, where the weights are calculated by the query and the key. The scaled dot product attention performs vector multiplication of the query vector Q and the key vector K, and then performs a dot product of size The zoom, Represents the vector dimension of Q and K. After normalization, multiply it with the value vector V to get the final output, which is expressed as:

[0049]

[0050] Multi-head attention is to concatenate the output results of multiple scaled dot product attentions and obtain the result through a fully connected layer, which can be expressed as:

[0051] MultiHead(Q,K,V)=Concat(head 1 ,...,head h )W o

[0052] In addition, the attention mechanism, like CNN, cannot represent temporal order. Compared with RNN, it performs poorly in tasks such as natural language processing, so position encoding (PE) is introduced. Using the absolute position encoding strategy, the position information of each moment is composed by sine and cosine functions of different frequencies:

[0053]

[0054] Among them, pos is the position, i is the position coordinate, d model The maximum length accepted by the model, that is, the maximum value of the coordinate.

[0055] Transformer's FNN consists of two fully connected layers plus a ReLU activation function, and the formula is:

[0056] FFN(x)=ReLU(xW 1 +b 1 )W 2 +b 2 (1)

[0057] Among them, W 1 and b 1 are the weight and bias of the first FFN, W 2 and b 2 are the weights and biases of the second FFN.

[0058] The intelligent agent of the present invention is built on the open source large language model Qwen2.5-72B.

[0059] The X-AI platform using this intelligent agent is defined as the D2-based X-AI platform, and X-AI and D2-based X-AI are analyzed. Specifically, the performance of the two platforms is evaluated by task completion rate (TCR), task execution time (TET) and number of steps (NOS). Specifically:

[0060]

[0061] Where: Ti is the time spent on completing each task; Si is the number of steps for each task; n is the number of completed tasks.

[0062] The analysis results are:

[0063]

[0064] In summary, this method realizes an intelligent interaction method of artificial intelligence platform based on large language model and intelligent agent, and realizes the effective combination of cutting-edge large language model technology and traditional artificial intelligence platform.

[0065] like Figure 4 As shown, the artificial intelligence platform constructed using this method includes a vector database, and the vector database Milvus stores the user manual and technical documents of the artificial intelligence platform. The intelligent platform is a visual modeling platform, which can automatically form a model through the natural language of the user's intention.

[0066] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An artificial intelligence platform interaction method based on a large language model and an intelligent agent, characterized in that: The steps include: S1: Develop a scripting language for an AI platform to form a tool set, which includes function tools; S2: Build a ReAct intelligent agent, which performs semantic analysis on the platform requirements expressed in natural language, infers the operation steps of the natural language requirements and the function tool formed in step 1 that needs to be called in each step to complete the text conversion operation; S3, manually revise and improve the text conversion operation, record the revision and improve the revision operation into a script and store it in the tool set in S1; S4. Output the revised text of S3 through the large language model.

2. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 1, characterized in that: In step S1, the mouse and keyboard operation definitions of dragging, dropping, adding, deleting and modifying nodes and lines of the artificial intelligence platform are encapsulated into a scripting language.

3. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 2, characterized in that: The script functions in step S1 include: adding node function, deleting node function, connecting node function, and defining data source function.

4. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 1, characterized in that: Step S2 includes: S20. Collect basic information of the script function tool in step 1, including: function name, function description, parameter name, parameter type, and parameter description; S21. Use prompt engineering to guide large language models to derive tools needed to achieve natural language requirements; S22, extracting key information and features of the natural language expressing the platform requirements, calling the corresponding function tool in step 1, and obtaining the tool calculation results; S23. Based on the calculation result of S22, feedback in the prompting project is constructed, and the prompting project guides the large language model to derive the next function tool to be used to realize the natural language demand; S24, looping through S22 and S23 until the tool selected by the large language model is the task end placeholder "Finish", indicating that the text conversion task is completed.

5. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 4, characterized in that: In step S22: the project guides the large language model to deduce the tools needed to implement natural language requirements according to the idea of ​​"problem>thinking>selecting tools>entering tool parameters".

6. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 4, characterized in that: In step S23: the project guides the large language model to deduce the next function tool needed to implement the natural language requirement according to the idea of ​​"feedback>thinking>select tool>input tool parameters>tool results".

7. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 4, characterized in that: Prompt engineering includes thinking, action, action input, and observation.

8. The artificial intelligence platform interaction method based on a large language model and an intelligent agent according to claim 1, characterized in that: The intelligent agent in step S2 is built based on the open source large language model Qwen2.5-72B.

Images