Program for artificial intelligence system and program for artificial intelligence system

Abstract

To provide a program for an artificial intelligence system to easily prepare data on an external information source for the artificial intelligence system, for improving the quality of a response.SOLUTION: A program for a system includes: a language character extraction step S2 of extracting language character information from an electronic information source on the basis of the electronic information source designated in an administrator terminal capable of freely communicating with a server; a language character data division step S3 of dividing data on the extracted language character information into a plurality of pieces of divided language character file data; and data conversion and storage steps S4 and S5 of converting the divided language character file data into vector coordinate system divided language data respectively using vector coordinate conversion means which is pre-trained artificial intelligence means, and storing the converted vector coordinate system divided language data in a vector index in association with the pre-conversion divided language character file data respectively.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a program for an artificial intelligence system in which a user terminal displays or outputs audio based on a server, and to the program for the artificial intelligence system. [Background technology]

[0002] Conventionally, there is known an information processing device that includes a language data receiving unit that receives language data in response to a request from a user terminal, a language data processing unit that processes the language data to generate basic summary information that forms the basis for generating a summary, a transmitting unit that transmits presentation information including the basic summary information to the user terminal, an editing information receiving unit that receives editing information related to the presentation information from the user terminal, and a summary generation unit that provides the editing information to a large-scale language model that receives input of the basic summary information and generates a summary sentence that reflects the editing information (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-73095 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the conventional information processing devices described above were limited to generating a text-format summary directly from basic summary information in some language text format using a large-scale language model, which meant there was a risk that the large-scale language model would misrecognize the content of the basic summary information, and the accuracy of understanding the content of the basic summary information was insufficient.

[0005] Therefore, the present invention solves the problems of the prior art as described above, and the object of the present invention is to provide a program for an artificial intelligence system that can easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for an artificial intelligence system to improve the quality of answers, and a program for the artificial intelligence system. [Means for solving the problem]

[0006] The invention of claim 1 is a program for an artificial intelligence system in which a user terminal displays or outputs audio based on a server, and solves the above-mentioned problem by comprising: a language character extraction step for extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources specified on an administrator terminal that can communicate with the server; a language character data division step for dividing the extracted language character information data into multiple divided language character file data; and a data conversion and storage step for converting the divided divided language character file data into vector coordinate system divided language data using a vector coordinate conversion means that is a pre-trained artificial intelligence means, and storing the converted vector coordinate system divided language data in a vector index in association with the divided language character file data before conversion.

[0007] The invention of claim 2 further solves the above-mentioned problem by, in addition to the configuration of the system program for the artificial intelligence system described in claim 1, creating language character file data of language character information for each electronic file when extracting language character information from an electronic information source in the language character extraction step.

[0008] The invention of claim 3 further solves the above-mentioned problem by including, in addition to the configuration of the system program for the artificial intelligence system described in claim 2, an update determination step for determining whether the electronic information source has been updated, a language character extraction and update step for extracting language character information from only the updated electronic file and updating the language character file data if an update has been made, a language character data re-division step for re-dividing the updated language character file data into a plurality of divided language character file data, and a data re-conversion and re-storage step for re-converting the re-divided divided language character file data into vector coordinate system divided language data using a vector coordinate conversion means which is a pre-trained artificial intelligence means, and associating the converted vector coordinate system divided language data with the divided language character file data before conversion and re-storing them in vector indexes, respectively.

[0009] The invention of claim 4 further solves the above-mentioned problem by, in addition to the configuration of the system program for the artificial intelligence system described in any one of claims 1 to 3, reading and extracting the language character information in the image using an image recognition means in the language character extraction step when the electronic information source contains image data having language character information in the image.

[0010] The invention according to claim 5 is a program for an artificial intelligence system in which a user terminal displays or outputs voice based on a server, the program comprising: a language character extraction step of extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources designated in an administrator terminal capable of communicating with the server; a language character data division step of dividing the extracted language character information data into a plurality of divided language character file data; a data conversion and storage step of converting the divided divided language character file data into vector coordinate system divided language data using vector coordinate conversion means, which is a pre-trained artificial intelligence means, and storing the converted vector coordinate system divided language data in a vector index in association with the divided language character file data before conversion; and a data conversion and storage step of converting the text data into a vector index in association with the divided language character file data before conversion, when a user inputs text on a browser of the user terminal, the user terminal converts the text data into a vector index in association with the divided language character file data before conversion. The above-mentioned problem is solved by the following steps: a question sending step in which the server sends the received question information to a server as question information; a correlation search step in which the server converts the received question information into vector coordinate system question data using a vector coordinate conversion means, searches for the vector coordinate system divided language data in the vector index that is most closely related to the vector of the vector coordinate system question data, and selects the divided language character file data corresponding to the most closely related vector coordinate system divided language data or the most closely related vector coordinate system divided language data; an answer text generation and transmission step in which the server adds the selected divided language character file data to the question information, generates text data using a text generation means which is a pre-trained artificial intelligence means, and sends the generated text data obtained by the execution to a user terminal; and a generated text display and audio output step in which the user terminal displays the received generated text data on a browser or outputs it as audio. [Effects of the Invention]

[0011] According to the system program for the artificial intelligence system of the invention of claim 1, the administrator simply specifies an electronic information source, and the language character information of the electronic information source is extracted and a division process, known as chunking, is performed, and then a vector conversion process, known as embedding, is performed, and the relationship before and after the vector conversion is associated, and a vector index storage process, known as mapping, is performed, thereby making it possible to easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system. In the artificial intelligence system, text data is sent from the user terminal to the server as question information, and the server converts the received question information into vector coordinate system question data using a vector coordinate conversion means, searches the vector index for vector coordinate system divided language data that is most closely related to the vector of the vector coordinate system question data, selects the most closely related vector coordinate system divided language data or divided language character file data corresponding to the most closely related vector coordinate system divided language data, and the server adds the selected divided language character file data to the question information and generates text data using a text generation means which is a pre-trained artificial intelligence means, and sends the generated text data obtained by the execution to the user terminal, thereby improving the quality of the answer. Furthermore, only the divided language character file data of the content that is highly relevant to the selected small portion of the question from the external information source is added to the question information and sent to the text generation means, and since the amount of data sent to the text generation means is relatively small, it is possible to avoid a long response time from sending a question to receiving an answer.

[0012] According to the system program for the artificial intelligence system of the invention of claim 2, in addition to the effect achieved by the invention of claim 1, language character file data is created for each electronic file, making management easy.

[0013] According to the system program for the artificial intelligence system of the invention of claim 3, in addition to the effects of the invention of claim 2, when an electronic information source is updated, language character information is extracted from only the updated electronic file, and chunking and embedding processing is performed, and this is reflected in the vector index, so-called differential update processing is performed. Therefore, information can be updated in a shorter time than in a configuration in which language character information is extracted, chunking processing, embedding processing, and vector index saving processing are performed from the beginning for the entire electronic information source.

[0014] According to the system program for the artificial intelligence system of the invention of claim 4, in addition to the effects achieved by the invention of any one of claims 1 to 3, even if the image data is one in which language characters are displayed in pixels rather than text data, the language character information can be read and extracted, making it possible to handle image data such as scanned data and photographs.

[0015] According to the program for the artificial intelligence system of the invention of claim 5, similar to the effect achieved by the invention of claim 1, the administrator simply specifies an electronic information source, and the language character information of the electronic information source is extracted and a division process, known as chunking, is performed, and then a vector conversion process, known as embedding, is performed, and the relationship before and after the vector conversion is associated, and a vector index storage process, known as mapping, is performed, thereby making it possible to easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system. In the artificial intelligence system, only the divided language character file data with content highly related to the question is added to the question information and sent to the text generation means, thereby improving the quality of the answer. Furthermore, only the divided language character file data of the content that is highly relevant to the selected small portion of the question from the external information source is added to the question information and sent to the text generation means, and since the amount of data sent to the text generation means is relatively small, it is possible to avoid a long response time from sending a question to receiving an answer. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a diagram showing the concept of an artificial intelligence system according to an embodiment of the present invention; [Figure 2] FIG. 2 is a chart showing an example of the operation of a system program of the artificial intelligence system according to an embodiment of the present invention. [Figure 3] 10A to 10D are diagrams showing an example of a reference range specification window of a browser on an administrator terminal of an artificial intelligence system according to an embodiment of the present invention. [Figure 4] FIG. 10 is an explanatory diagram showing how the artificial intelligence system according to an embodiment of the present invention divides language text data and converts the divided language character file data into vector coordinate system divided language data. [Figure 5] FIG. 10 is an explanatory diagram showing how language data after vector coordinate system division in the artificial intelligence system according to an embodiment of the present invention is stored in a vector index. [Figure 6] FIG. 1 is a chart showing an example of the operation of a user-oriented program of an artificial intelligence system according to an embodiment of the present invention. [Figure 7] A figure showing an example of an AI document screen as a talk screen in a browser on a user terminal of an artificial intelligence system according to an embodiment of the present invention. [Figure 8] FIG. 1 is an explanatory diagram showing how question information, which is a prompt of the artificial intelligence system according to an embodiment of the present invention, is converted into vector coordinate system question data. [Figure 9] 1A and 1B are explanatory diagrams illustrating correlation search between vector coordinate system question data and vector coordinate system divided language data in an artificial intelligence system according to an embodiment of the present invention. [Figure 10]FIG. 10 is an explanatory diagram showing how selected divided language character file data is added to question information, which is a prompt of the artificial intelligence system according to an embodiment of the present invention, and transmitted to the pre-trained artificial intelligence means. [Figure 11] An explanatory diagram showing how language text data is divided and the divided language character file data is converted into vector coordinate system divided language data when some documents within the specified reference range of an artificial intelligence system that is an embodiment of the present invention are updated. [Figure 12] FIG. 10 is an explanatory diagram showing how characters are extracted when pixel data is included in a specified reference range of an artificial intelligence system according to an embodiment of the present invention. [Figure 13] FIG. 1 is an explanatory diagram showing how parallel processing is performed for each block of an artificial intelligence system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] The system program (also called the setup program) of the artificial intelligence system of the present invention has a language character extraction step for extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources specified on an administrator terminal that can communicate with the server; a language character data division step for dividing the extracted language character information data into multiple divided language character file data; and a data conversion and storage step for converting the divided divided language character file data into vector coordinate system divided language data using a vector coordinate conversion means that is a pre-trained artificial intelligence means, and associating the converted vector coordinate system divided language data with the pre-conversion divided language character file data and storing them in vector indexes, so that the specific implementation can be any as long as it can easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system. The program for the artificial intelligence system of the present invention includes a language character extraction step of extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are designated electronic information sources, in an administrator terminal capable of communicating with the server; a language character data division step of dividing the extracted language character information data into a plurality of divided language character file data; a data conversion and storage step of converting the divided divided language character file data into vector coordinate system divided language data using vector coordinate conversion means, which is a pre-trained artificial intelligence means, and storing the converted vector coordinate system divided language data in a vector index in association with the divided divided language character file data before conversion; a question transmission step of transmitting the text data to the server as question information when a user inputs text on a browser of the user terminal; However, the specific implementation of the system may be any one as long as it has a correlation search step of converting the received question information into vector coordinate system question data using a vector coordinate conversion means, searching for vector coordinate system divided language data in a vector index that is most closely related to the vector of the vector coordinate system question data, and selecting divided language character file data corresponding to the most closely related vector coordinate system divided language data or the most closely related vector coordinate system divided language data; an answer text generation / transmission step in which the server adds the selected divided language character file data to the question information, generates text data using a text generation means that is a pre-trained artificial intelligence means, and transmits the generated text data obtained by the execution to the user terminal; and a generated text display / audio output step in which the user terminal displays the received generated text data on a browser or outputs it as audio.This makes it possible to easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system, and further improves the quality of answers and avoids long response times from sending a question to receiving an answer.

[0018] For example, administrator terminals and user terminals can be devices that send and receive information, such as desktop personal computer terminals, notebook personal computer terminals, smartphone terminals, and tablet terminals, and can be any device that can be freely connected to the server via a communication network, including a wide area network such as the Internet, a local network, or a telephone line. The servers may be a single server or multiple servers on a cloud. The pre-trained artificial intelligence means may be composed of an Embedding model, which is an example of a vector coordinate conversion means that converts input data such as language text data into data in a vector coordinate system, or an interactive text generation means that generates text data and is also called a large-scale language model, such as ChatGPT (Generative Pre-trained Transformer) (hereinafter referred to as ChatGPT), and may be configured to be installed on one server or multiple servers on the cloud. Furthermore, the electronic information source may be any information that can be specified as electronic data, such as electronic file information such as a document file, URL information of a website, or database information. Here, a document file refers to an electronic data file whose contents contain text information, and the data file format may be a simple text data format, a document data format, a cell-display type spreadsheet file format, or an image file format such as a pixel-display type photograph or video, which can be read using an image recognition means or a voice recognition means, and can be any format that directly recognizes characters or indirectly recognizes characters by converting voice to characters. The characters also include the source code of the program. Furthermore, the URL information (Uniform Resource Locator information), which is the location information for the specified information that specifies the electronic information source, can be anything that can identify an HTML format website page or document file as the specified information obtained by accessing a web server. [Example]

[0019] An artificial intelligence system 100 according to an embodiment of the present invention will be described below with reference to FIGS. 1 is a diagram showing the concept of an artificial intelligence system 100 according to an embodiment of the present invention; FIG. 2 is a chart showing an example of the operation of a system program for the artificial intelligence system 100 according to an embodiment of the present invention; FIGS. 3(A) to 3(D) are diagrams showing an example of a reference range specification window 151 of a browser on an administrator terminal 150 of the artificial intelligence system 100 according to an embodiment of the present invention; FIG. 4 is an explanatory diagram showing how language character file data TX of the artificial intelligence system 100 according to an embodiment of the present invention is divided and the divided language character file data DTX is converted into vector coordinate system divided language data VTX; FIG. 5 is an explanatory diagram showing how the vector coordinate system divided language data VTX of the artificial intelligence system 100 according to an embodiment of the present invention is stored in a vector index 121; FIG. 6 is a chart showing an example of the operation of a user-oriented program for the artificial intelligence system 100 according to an embodiment of the present invention; 9(A) is a diagram showing the vector coordinate system question data VPT of the artificial intelligence system 100 according to an embodiment of the present invention, and FIG. 9(B) is a diagram showing a correlation search between the vector coordinate system question data VPT of the artificial intelligence system 100 according to an embodiment of the present invention and the vector coordinate system divided language data VTX. FIG. 10 is a diagram showing the process of adding selected divided language character file data DTX to the question information PT, which is the prompt of the artificial intelligence system 100 according to an embodiment of the present invention, and transmitting the result to the pre-trained artificial intelligence means. FIG. 11 is a diagram showing the process of dividing the language character file data TX when some documents in the specified reference range of the artificial intelligence system 100 according to an embodiment of the present invention are updated, and converting the divided language character file data DTX into vector coordinate system divided language data VTX. FIG. 12 is a diagram showing the process of extracting characters when pixel data is included in the specified reference range of the artificial intelligence system 100 according to an embodiment of the present invention.FIG. 1 is an explanatory diagram showing how parallel processing is performed for each block BK in an artificial intelligence system 100 according to an embodiment of the present invention.

[0020] As shown in Figure 1, an artificial intelligence system 100, an embodiment of the present invention, includes a user terminal 110, a first server 120, a second server 130, and a third server 140 as examples of servers, and an administrator terminal 150. Of these, the first server 120 has a database, a vector index 121 which also serves as a spatial map, and a bot 122 which is also a display program, and the database contains information on the web page of the AI ​​document screen 112, which is the talk screen of the artificial intelligence system 100 displayed by the browser 111 of the user terminal 110. Here, the vector index 121 refers to a list, database, or conceptual spatial map in a storage unit for quickly finding a specific object from among a large amount of vector coordinate system data.

[0021] The second server 130 also has a vector coordinate conversion means 131 as pre-trained artificial intelligence means. The vector coordinate conversion means 131 is also called an embedding model and is provided to convert input data into data in a vector coordinate system. For example, it determines vector components based on each element of the data to be converted and its intensity and frequency of appearance. Furthermore, the third server 140 has a chat GPT, which is an example of a text generation means 141 also called a large-scale language model (LLM) as a pre-trained artificial intelligence means. As a technical concept, the first server 120, the second server 130, and the third server 140 may have a common physical configuration, or may have physically different configurations.

[0022] In addition, in the interaction between the user terminal 110 and the chat GPT, which is an example of the text generation means 141, the administrator terminal 150 specifies the external information source to be referenced by the chat GPT, which is an example of the text generation means 141, based on the administrator's operation. Here, the external information source refers to an external information source that has not been pre-trained, as opposed to the information that chat GPT, which is an example of the text generation means 141, has pre-trained and internally stored. When the user inputs text in the browser 111 of the user terminal 110, the user terminal 110 transmits the text data to the first server 120 as question information PT.

[0023] Then, the first server 120 executes text data generation using chat GPT, which is an example of the text generation means 141, which is the pre-trained artificial intelligence means of the third server 140. Then, the first server 120 transmits the generated text data GT obtained through the execution to the user terminal 110. Then, the user terminal 110 is configured to display the received generated text data GT on the browser 111 or output it as voice.

[0024] In this embodiment, when an administrator specifies an external information source using the administrator terminal 150, the system program of the artificial intelligence system 100 sets up the external information source, thereby improving the quality of responses from chat GPT, which is an example of text generation means 141, when a user interacts with chat GPT, which is an example of text generation means 141, a pre-trained artificial intelligence means, using the user terminal 110. The system program for the artificial intelligence system 100 according to the embodiment of the present invention will be described in detail below.

[0025] As shown in FIG. 2, in step S1, the first server 120, which is an example of a server, determines whether or not an electronic information source has been designated as an external information source on the administrator terminal 150 that can communicate with the first server 120. More specifically, as shown in FIG. 3(A), when the administrator terminal 150 logs in as an administrator, communicates with the first server 120, and performs a predetermined operation on the administrator screen, the browser on the administrator terminal 150 opens a reference range specification window 151. The reference range specification window 151 has a "File Upload" tab, a "URL" tab, a "Folder" tab, and a "Database" tab.

[0026] When the "File Upload" tab is selected, the reference range specification window 151 displays a drag area 151a as a predetermined area. As an example, suppose an administrator uses a mouse to drag the icon of a document file DF, for example, "The Appeal of Electric Cars".pptx, from the drive in the storage section of the administrator terminal 150 or the file server to the drag area 151a and then operates the send button. In this way, when the icon of the document file DF is dragged into the drag area 151a as the specified area and the send button is operated, it is determined that an electronic information source has been specified and the process proceeds to step S2; on the other hand, if it is determined that no such source has been specified yet, step S1 is repeated.

[0027] Furthermore, as shown in FIG. 3(B), when the "URL" tab is selected in the reference range specification window 151, the reference range specification window 151 displays a URL input field 151b and a document name input field. You can enter URL information in this URL input field 151b and operate the send button to specify the URL information that is the location information for the specified information, or you can enter a file name in the document name input field and operate the send button.

[0028] Furthermore, as shown in FIG. 3(C), when the "Folder" tab is selected in the reference range specification window 151, the reference range specification window 151 displays a folder specification field 151c that can freely display a list of folder tree structures. Then, the user may select a desired folder in the folder specification field 151c and operate the send button to specify the electronic information source. Furthermore, as shown in FIG. 3(D), when the "Database" tab is selected in the reference range specification window 151, the reference range specification window 151 displays a database URL input field 151d, an ID input field, and a password input field. The user may input URL information in the database URL input field 151d, ID information in the ID input field, and password information in the password input field, and then operate the send button. When the drag area 151a in FIG. 3(A) is clicked, the state changes to that shown in FIG. 3(C), and a list of folders is displayed in the folder specification field 151c of the reference range specification window 151. The folder containing the desired document file DF may be selected and designated from the list of folders in this folder designation field 151c.

[0029] In this way, the method of specifying an electronic information source can be such that the administrator drags the icons of one or more document files DF in the reference range specification window 151 displayed on the browser of the administrator terminal 150 to specify one or more document files DF as the reference range of the electronic information source, or the administrator can specify one or more URL information, or the administrator can specify the reference range of the electronic information source by specifying a folder that stores multiple document files DF in the database of a file server or first server 120, such as a folder in the database of a core system.

[0030] In step S2, as a language character extraction step, the first server 120, as an example of a server, extracts language character information from the electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources specified on the administrator terminal 150. More specifically, as shown in FIG. 4, as an example, language character information is extracted from a document file DF of "The Appeal of Electric Cars".pptx. As a higher-level concept of the technical idea, it is sufficient to extract language character information, and when extracted, the extracted language character information may or may not be in a file format such as the language character file data TX shown in Figure 4. Furthermore, although the first server 120 as an example of a server extracts language character information from an electronic information source, the second server 130 may extract it.

[0031] In step S3, as a language character data dividing step, the first server 120 as an example of a server performs a so-called chunking process to divide the extracted language character information data into a plurality of divided language character file data DTX. More specifically, as shown in Figure 4, as an example, language character file data TX, which is language character information extracted from a document file DF called "The Appeal of Electric Vehicles".pptx, is divided into multiple divided language character file data DTX, for example, each containing 500 characters. In consideration of variations in the length of each sentence, the sentence may be divided into a character length of 400 to 600 characters as an example of the predetermined range, or each sentence may be divided into a plurality of parts. Furthermore, the first server 120, which is an example of a server, performs chunking processing, but the second server 130 may perform chunking processing.

[0032] In step S4, as a data conversion and storage step, the first server 120, which is an example of a server, performs a so-called embedding process in which the divided post-division language character file data DTX is converted into vector coordinate system post-division language data VTX using a vector coordinate conversion means 131, which is a pre-trained artificial intelligence means. More specifically, as shown in FIG. 4, the divided language character file data DTX is converted into vector coordinate system divided language data VTX each having a plurality of parameters. Here, the "vector coordinate system divided language data VTX" is, for example, vector data consisting of 1536 dimensions (parameters). In other words, it is vector data specified by 1536 components. The number of components of a vector may be any number. Furthermore, the first server 120 as an example of a server performs the embedding process, but the second server 130 may perform the embedding process.

[0033] In step S5, as a data conversion and storage step, the first server 120, as an example of a server, performs a so-called mapping process in which the converted vector coordinate system divided language data VTX is associated with the pre-conversion divided language character file data DTX and stored in the vector index 121. More specifically, the vector coordinate system divided language data VTX shown in FIG. 4 is stored in the vector index 121 of the first server 120, as shown in FIG. At this time, the vector coordinate system divided language data VTX is stored in association with the pre-conversion divided language character file data DTX. Although the location of the vector index 121 is the first server 120 as an example, it does not have to be the first server 120. Furthermore, the first server 120, which is an example of a server, performs the mapping process, but the second server 130 may perform the mapping process.

[0034] This allows the administrator to simply specify an electronic information source, and the language and character information of the electronic information source is extracted and a division process known as chunking is performed, and then a vector conversion process known as embedding is performed, and the relationship before and after the vector conversion is associated, and a vector index saving process known as mapping is performed. As a result, data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system 100, can be easily generated.

[0035] Next, the user-facing program of the artificial intelligence system 100 will be described. As shown in Figure 6, in step S11, as a question operation determination step (question sending step), the user terminal 110 determines whether the user has performed a text input operation on the AI ​​document screen 112 of the browser 111 of the user terminal 110 and whether a sending operation has been performed.

[0036] More specifically, as shown in FIG. 7, it is determined whether or not text is entered into an input field 113 on an AI document screen 112 and the send button is operated. Note that data input by voice by operating the microphone icon button 114 may be converted into text data by the browser 111 and input into the input field 113 . As an example, suppose someone inputs the question "What are the advantages of electric vehicles compared to gasoline-powered vehicles?" and sends it. In this way, if it is determined that text has been input and a send operation has been performed, the process proceeds to step S12, and if it is determined that no such operation has been performed yet, step S11 is repeated.

[0037] In step S12, as a question transmission step, user terminal 110 transmits text data to first server 120 as question information PT, which is an example of instruction and command information. More specifically, as an example, question information PT of "What are the advantages of electric cars compared to gasoline cars?" is transmitted to the first server 120.

[0038] In step S13, as a question information vector conversion step (correlation search step), the first server 120 (an example of a server) uses the vector coordinate conversion means 131 to convert the received question information PT into vector coordinate system question data VPT. More specifically, as shown in FIG. 8, for example, question information PT of "What are the advantages of electric vehicles compared to gasoline vehicles?" is converted into vector coordinate system question data VPT of "2, 6, 1, 1, 7, 7, 6, 3, 1 . . . ." FIG. 9(A) shows question information PT in text format data converted into vector coordinate system question data VPT. This display is, for example, two-dimensional, but the actual vector coordinates are, for example, 1536-dimensional. Although the first server 120, which is an example of a server, performs the embedding process on the question information PT, the second server 130 may perform the embedding process.

[0039] In step S14, as shown in Figure 9(B), as a correlation search step, the first server 120, as an example of a server, searches for the vector coordinate system divided language data VTX that has the closest relationship to the vector of the vector coordinate system question data VPT in the vector index 121. Then, the divided language character file data DTX corresponding to the most closely related vector coordinate system divided language data VTX or the most closely related vector coordinate system divided language data VTX is selected. Although the first server 120, which is an example of a server, performs the correlation search, the second server 130 may perform the correlation search.

[0040] In step S15, as shown in Figure 10, as a reference information attachment and transmission step (answer text generation and transmission step), the first server 120, as an example of a server, adds the selected divided language character file data DTX to the question information PT and transmits it to the chat GPT, which is an example of a text generation means 141, which is a pre-trained artificial intelligence means. More specifically, reference information of "Electric vehicles are electric...", which is an example of the selected divided language character file data DTX, is added to "What are the advantages of electric vehicles compared to gasoline vehicles?", which is an example of question information PT, and is sent to chat GPT, which is an example of text generation means 141, which is a pre-trained artificial intelligence means. At this time, the bot 122 of the first server 120 adds the selected divided language character file data DTX to the question information PT and sends it to the chat GPT, which is an example of a text generation means 141 that is a pre-trained artificial intelligence means, as a so-called prompt, which is standard instruction / command information for answering the question information PT by referring to the divided language character file data DTX. Although the first server 120, which is an example of a server, transmits the message to the chat GPT, the second server 130 may transmit the message to the chat GPT.

[0041] In step S16, as a response text generating and transmitting step, the first server 120 generates text data using chat GPT, which is an example of the text generating means 141, which is a pre-trained artificial intelligence means. In other words, the chat GPT, which is an example of a text generation means 141 that is a pre-trained artificial intelligence means, generates generated text data GT as an answer based on the received prompt (an instruction / command to answer the question information PT by referring to the divided language character file data DTX). Then, as an example of generated text data GT, the following is generated: "Since everything is electronically controlled, it is compatible with automated driving. Just input your destination by voice and then depart..."

[0042] Then, the first server 120 transmits the generated text data GT obtained through the execution to the user terminal 110. More specifically, an example of generated text data GT is transmitted to the user terminal 110 as follows: "Since everything is electronically controlled, it is compatible with automated driving. Just input your destination by voice and then depart..."

[0043] In step S17, as a generated text display / voice output step, the user terminal 110 displays the received generated text data GT on the browser 111 or outputs it as voice. More specifically, the user terminal 110 displays, as an example of generated text data GT, the following on the AI ​​document screen 112 of the browser 111, or outputs it aloud using a speaker: "Since everything is electronically controlled, it is compatible with autonomous driving. Just enter your destination by voice and then depart..."

[0044] As a result, as mentioned above, the administrator simply specifies the electronic information source, and the language character information of the electronic information source is extracted and a division process, known as chunking, is performed, and then a vector conversion process, known as embedding, is performed, and the relationship before and after the vector conversion is associated, and a vector index storage process, known as mapping, is performed. As a result, data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system 100, can be easily generated.

[0045] Then, in the artificial intelligence system 100, only the divided language character file data DTX having content highly related to the question is added to the question information PT and sent to the chat GPT, which is an example of the text generating means 141. As a result, the quality of the answers can be improved. Furthermore, only the divided language character file data DTX containing content that is highly relevant to a selected small portion of questions from external information sources is added to the question information PT and sent to the chat GPT, which is an example of text generation means 141, and the amount of data sent to the chat GPT, which is an example of text generation means 141, is relatively small. As a result, it is possible to avoid a long response time from sending a question to receiving a reply.

[0046] In addition, a speaker icon button 115 is provided at each exchange location on the AI ​​document screen 112. When the speaker icon button 115 is operated, the contents of the exchange, such as the question information PT and the generated text data GT, are converted into voice data and output as voice using the speaker of the user terminal 110. This allows for interaction by looking, speaking, and listening without typing. As a result, the user can smoothly interact with the chat GPT, which is an example of the text generation means 141, which is a pre-trained artificial intelligence means, while viewing the contents of the AI ​​document screen 112.

[0047] In addition, in this embodiment, in the language character extraction step of the system program of the artificial intelligence system 100, when the first server 120 extracts language character information from an electronic information source, the first server 120 creates language character file data TX of the language character information for each electronic file. As a result, language character file data TX is created for each electronic file of the electronic information source. As a result, management can be easily performed.

[0048] Furthermore, in this embodiment, the system program of the artificial intelligence system 100 further includes an update presence / absence determination step, a language character extraction / update step, and a data reconversion / re-storage step. In the update determination step, the first server 120 or the server where the electronic information source is stored determines whether the electronic information source has been updated.

[0049] More specifically, as shown in FIG. 11, there are a large number of document files DF, which are electronic information sources, from file 1 to file 515, and it is assumed that file 13 of these has been updated. In the language character extraction and update step, when there is an update, the first server 120 extracts language character information only from the updated electronic file and updates the language character file data TX. More specifically, as an example, language character information is extracted from only the 13th file, and the language character file data TX is updated.

[0050] In the language character data re-division step, the first server 120 re-divides the updated language character file data TX into a plurality of divided language character file data DTX. More specifically, as shown in FIG. 11, for example, the language character file data TX, which is the content of the 13th file, is again divided into a plurality of divided language character file data DTX.

[0051] In the data reconversion and restoring step, the first server 120 reconverts the re-divided divided language character file data DTX into vector coordinate system divided language data VTX using a vector coordinate conversion means 131, which is a pre-trained artificial intelligence means. At the same time, the first server 120 associates the converted vector coordinate system divided language data VTX with the pre-conversion divided language character file data DTX and stores them again in the vector index 121.

[0052] As a result, when an electronic information source is updated, language character information is extracted from only the updated electronic file, chunking processing and embedding processing are performed, and the information is reflected in the vector index 121, thereby performing a so-called differential update processing. As a result, information can be updated in a shorter time than in a configuration in which language character information is extracted, chunked, embedded, and vector index saved from the beginning for the entire electronic information source.

[0053] In this embodiment, as shown in FIG. 12, when the electronic information source contains image data PH having language character information in the image, in the language character extraction step of the system program of the artificial intelligence system 100, the first server 120 reads and extracts data PH1 of the language character information in the image using an image recognition means not shown. More specifically, suppose that an electronic information source contains image data PH in pixel format, and the image of this image data PH contains, as an example, the words "Come to Kyoto! Tourism Promotion Project" displayed in pixel format. In this case, since it is not font data, character information cannot be extracted directly from the image data PH.

[0054] Therefore, the image data PH is virtually displayed on a computer using an image recognition means, and the displayed image is virtually scanned. If characters are included, character recognition OCR (Optical Character Recognition / Reader) is performed and the characters "Come to Kyoto! Tourism Promotion Project" are extracted as language character information data PH1. It should be noted that the image recognition means may be trained in advance to recognize things other than text in the image, such as scenery and figures. For example, if a photo shows Kiyomizu-dera Temple in Kyoto and autumn leaves, the content may be extracted as text such as <Photo: Kiyomizu-dera Temple, autumn leaves>.

[0055] Furthermore, the extracted language character information data PH1 is divided into a plurality of divided language character file data DTX as described above. Then, the divided language character file data DTX is converted into vector coordinate system divided language data VTX using vector coordinate conversion means 131, which is pre-trained artificial intelligence means. Furthermore, the converted vector coordinate system divided language data VTX is stored in the vector index 121 in association with the pre-conversion divided language character file data DTX. As a result, even if the image data PH is one in which language characters are displayed in pixels rather than as text data in a font format, the data PH1 of the language character information is read and extracted. As a result, it can also handle image data PH such as scanned data and photographs.

[0056] Regarding the embedding process and correlation search process, the configuration has been described in which the divided post-division language character file data DTX is converted into vector coordinate system post-division language data VTX using vector coordinate conversion means 131, and when performing correlation search, all vector coordinate system post-division language data VTX is searched at once, but this is not limited to this. For example, as a variant example, as shown in Figure 13, there is a huge amount of information in an electronic information source, and content information identified by multiple document files DF or multiple URL information is defined as one block BK, and there are multiple blocks BK of content information identified by multiple document files DF or multiple URL information. The content information within the reference range of the specified electronic information source is several blocks BK in length.

[0057] Furthermore, the first server 120 receives the question information PT from the user terminal 110 and converts it into vector coordinate system question data VPT using the vector coordinate conversion means 131 of the second server 130. Thereafter, the first server 120 searches for the vector coordinate system divided language data VTX that is most closely related to the vector of the vector coordinate system question data VPT in the vector index 121 for each block BK by parallel processing. Then, the first server 120 picks up the closest related vector coordinate system divided language data VTX in each block BK, or a plurality of top-ranked closest related vector coordinate system divided language data VTX, by parallel processing.

[0058] Next, the first server 120 further picks up the closest related vector coordinate system divided language data VTX picked up in all blocks BK, or the closest related vector coordinate system divided language data VTX from among the multiple closest related vector coordinate system divided language data VTX, or the most closely related vector coordinate system divided language data VTX from among the multiple closest related vector coordinate system divided language data VTX. Then, the first server 120 further selects the most closely related vector coordinate system divided language data VTX picked up, or the divided language character file data DTX corresponding to the most closely related vector coordinate system divided language data VTX.

[0059] Next, the first server 120 adds the selected divided language character file data DTX to the question information PT. Furthermore, the first server 120 performs text data generation using chat GPT, which is an example of a text generation means 141, which is a pre-trained artificial intelligence means of the third server 140, based on the data in which the divided language character file data DTX is added to the question information PT. The first server 120 may then be configured to transmit the generated text data GT obtained through the execution to the user terminal 110.

[0060] As a result, when the number of content information identified by document files DF or URL information as pre-specified electronic information sources is large, for example, 400 files or more, a correlation search is performed in parallel for each block BK, and the results of each block BK are viewed across all blocks BK, and the most closely related vector coordinate system divided language data VTX, or the divided language character file data DTX corresponding to the top multiple most closely related vector coordinate system divided language data VTX, is selected and added to the question information PT, and text data generation is performed using chat GPT, which is an example of a text generation means 141, which is a pre-trained artificial intelligence means. As a result, the response time from input of question information PT such as a question by the user to display of the answer can be significantly shorter than in a configuration that performs serial processing, allowing for smoother communication with the user.

[0061] In addition, instead of a configuration in which a correlation search is performed in parallel within all blocks BK, a configuration may be adopted in which a correlation search is performed on an outline of each block BK, and a correlation search is performed only within the block BK with the highest correlation, or within the top several blocks BK in the correlation result ranking. More specifically, the first server 120 creates one piece of vector coordinate system data for each block BK based on the vector coordinate system divided language data VTX in the block BK. For example, one vector coordinate system data is created using the average value of all the vector coordinate system divided language data VTX in the block BK. Then, the first server 120 receives the query information PT from the user terminal 110 and converts it into vector coordinate system query data VPT using the vector coordinate conversion means 131 of the second server 130, and then searches for the vector coordinate system data in each block BK that has the closest relationship to the vector of the vector coordinate system query data VPT. Next, the block BK corresponding to the vector coordinate system data that is the closest to the block BK or the vector coordinate system data that is the closest to the block BK is selected.

[0062] Furthermore, for each selected block BK, the vector index 121 is searched for the vector coordinate system divided language data VTX that is most closely related to the vector of the vector coordinate system question data VPT. Next, the closest related vector coordinate system divided language data VTX in the selected block BK or a plurality of top-ranked closest related vector coordinate system divided language data VTX are picked up in parallel processing. Furthermore, the divided language character file data DTX corresponding to the picked up most closely related vector coordinate system divided language data VTX or the most closely related vector coordinate system divided language data VTX is selected.

[0063] Next, the first server 120 adds the selected divided language character file data DTX to the question information PT. Then, the first server 120 executes text data generation using chat GPT, which is an example of the text generation means 141, which is the pre-trained artificial intelligence means of the third server 140. Furthermore, the first server 120 may be configured to transmit the generated text data GT obtained by the execution to the user terminal 110.

[0064] As a result, when the number of content information specified by the preset document file DF or URL information is large, for example, 800 files or more, a correlation search is not performed in parallel within each block BK, but rather a correlation search is performed on the outline of each block BK, and a correlation search is performed only within the block BK with the highest correlation, or within the top multiple blocks BK in the correlation result ranking, and the most closely related vector coordinate system divided language data VTX or the divided language character file data DTX corresponding to the top multiple most closely related vector coordinate system divided language data VTX is selected and added to the question information PT, and text data generation is performed using chat GPT, which is an example of a text generation means 141, which is a pre-trained artificial intelligence means. As a result, the response time from input of question information PT such as a question by the user to display of the answer can be significantly shorter than in a configuration that performs serial processing, allowing for smoother communication with the user. In other words, compared to a configuration in which correlation searches are performed in parallel within all blocks BK, the processing time can be significantly reduced, and the response time from input of question information PT such as a question by the user to display of the answer can be shortened.

[0065] The system program for the artificial intelligence system 100, an embodiment of the present invention, obtained in this manner includes a language character extraction step S2 for extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources specified on an administrator terminal 150 that can communicate with a first server 120, which is an example of a server; a language character data division step S3 for dividing language character file data TX, which is data of the extracted language character information, into multiple divided language character file data DTX; and data conversion and storage steps S4 and S5 for converting the divided divided language character file data DTX into vector coordinate system divided language data VTX using a vector coordinate conversion means 131, which is a pre-trained artificial intelligence means of the second server 130, and storing the converted vector coordinate system divided language data VTX in association with the pre-conversion divided language character file data DTX in the vector index 121, respectively.This makes it possible to easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system 100.

[0066] Furthermore, in the language character extraction step S2, when language character information is extracted from the electronic information source, language character file data TX of the language character information is created for each electronic file, which simplifies management.

[0067] In addition, the system program of the artificial intelligence system 100 further includes an update determination step for determining whether the electronic information source has been updated; a language character extraction and update step for extracting language character information from only the updated electronic file and updating the language character file data TX if an update has been made; a language character data re-division step for re-dividing the updated language character file data TX into a plurality of divided language character file data DTX; and a data re-conversion and re-save step for converting the re-divided divided language character file data DTX into vector coordinate system divided language data VTX again using vector coordinate conversion means 131, which is the pre-trained artificial intelligence means of the second server 130, and associating the converted vector coordinate system divided language data VTX with the pre-conversion divided language character file data DTX and re-save each in the vector index 121, respectively.As a result, a so-called differential update process is performed, and information can be updated in a shorter time than a configuration in which language character information extraction, chunking process, embedding process, and vector index save process are performed for the entire electronic information source from scratch.

[0068] Furthermore, if the electronic information source contains image data PH having language character information in the image, in the language character extraction step S2, data PH1 of the language character information in the image is read and extracted using an image recognition means not shown, so that even if the image data PH has language characters displayed in pixels rather than text data, it can also be used for image data PH such as scanned data or photographs.

[0069] The program of the artificial intelligence system 100 according to the embodiment of the present invention includes a language character extraction step S2 of extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are electronic information sources designated on an administrator terminal 150 that can communicate with a first server 120 that is an example of a server; a language character data division step S3 of dividing language character file data TX, which is data of the extracted language character information, into a plurality of divided language character file data DTX; data conversion and storage steps S4 and S5 of converting the divided divided language character file data DTX into vector coordinate system divided language data VTX using vector coordinate conversion means 131, which is pre-trained artificial intelligence means of the second server 130, and storing the converted vector coordinate system divided language data VTX in association with the pre-conversion divided language character file data DTX in the vector index 121; and question sending steps S11 and S12 in which the first server 120 sends the text data as question information PT to the first server 120; correlation search steps S12 to S14 in which the first server 120 converts the received question information PT into vector coordinate system question data VPT using the vector coordinate conversion means 131 of the second server 130, searches the vector index 121 for vector coordinate system divided language data VTX that is most closely related to the vector of the vector coordinate system question data VPT, and selects divided language character file data DTX corresponding to the most closely related vector coordinate system divided language data VTX or the most closely related vector coordinate system divided language data VTX; answer text generating and sending steps S15 and S16 in which the first server 120 adds the selected divided language character file data DTX to the question information PT, and generates text data using chat GPT, which is an example of text generating means 141 that is a pre-trained artificial intelligence means of the third server 140, and sends the generated text data GT obtained by the execution to the user terminal 110;The system includes a generated text display / audio output step S17 for displaying the received generated text data GT on the browser 111 or outputting it as audio, which makes it possible to easily create data for so-called RAG (Retrieval Augmented Generation), which is an external information source for the artificial intelligence system 100, and further has significant advantages, such as improving the quality of answers and avoiding long response times from sending a question to receiving an answer. [Explanation of symbols]

[0070] 100 ··· Artificial Intelligence Systems 110 User terminal 111 ··· (User device) browser 112 ··· AI document screen (talk screen) 113 ··· Input field 114 ··· Microphone icon button 115 ··· Speaker icon button 120 First Server (Server) 121 ··· Vector index (spatial map) 122 ··· Bot 130 Second Server (Server) 131 ··· Vector coordinate transformation means (pre-trained artificial intelligence means, embedding model) 140 ··· Third Server (Server) 141 ··· Text generation methods (pre-trained AI methods, large-scale language models (LLM), chat GPT) 150 ··· Administrator terminal 151 Reference Range Specification Window 151a··· Drag region 151b··· URL input field 151c··· Folder specification field 151d··· Database URL input field DF: Document file (electronic information source, electronic file) PH: Image data (in pixel format) (electronic source, electronic file) PH1: Language character information data (extracted from image data) TX: (Reference target) language character file data (language character information) DTX: Post-division language character file data VTX: Vector coordinate system divided language data PT ··· Question information VPT... Vector coordinate system query data GT ··· Generated text data BK Block

Claims

1. A program for an artificial intelligence system in which a user terminal displays or outputs audio based on a server, a language character extraction step of extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are designated in an administrator terminal capable of communicating with the server; a language character data dividing step of dividing the extracted language character information data into a plurality of divided language character file data; A program for an artificial intelligence system, characterized in that it includes a data conversion and storage step of converting the divided language character file data into vector coordinate system divided language data using a vector coordinate conversion means which is a pre-trained artificial intelligence means, and associating the converted vector coordinate system divided language data with the divided language character file data before conversion and storing them in vector indexes.

2. A program for an artificial intelligence system as described in claim 1, characterized in that in the language character extraction step, when language character information is extracted from an electronic information source, language character file data of the language character information is created for each electronic file.

3. an update determination step of determining whether the electronic information source has been updated; a language character extraction and update step for extracting language character information from only the updated electronic file and updating the language character file data when an update is detected; a language character data re-division step of re-dividing the updated language character file data into a plurality of divided language character file data; A program for an artificial intelligence system as described in claim 2, further comprising a data reconversion / re-saving step in which the re-divided divided language character file data is converted again into vector coordinate system divided language data using a vector coordinate conversion means which is a pre-trained artificial intelligence means, and the converted vector coordinate system divided language data is associated with the divided language character file data before conversion and saved again in a vector index.

4. A program for an artificial intelligence system according to any one of claims 1 to 3, characterized in that if the electronic information source contains image data having language character information in the image, in the language character extraction step, the language character information in the image is read and extracted using an image recognition means.

5. A program of an artificial intelligence system that a user terminal displays or outputs audio based on a server, a language character extraction step of extracting language character information from an electronic information source based on at least one of electronic file information, website URL information, and database information, which are designated in an administrator terminal capable of communicating with the server; a language character data dividing step of dividing the extracted language character information data into a plurality of divided language character file data; a data conversion and storage step of converting the divided language character file data into vector coordinate system divided language data using a vector coordinate conversion means which is a pre-trained artificial intelligence means, and storing the converted vector coordinate system divided language data in a vector index in association with the divided language character file data before conversion; a question sending step in which, when a user inputs text on a browser of the user terminal, the user terminal sends text data as question information to a server; a correlation search step in which the server converts the received question information into vector coordinate system question data using a vector coordinate conversion means, searches for vector coordinate system divided language data in the vector index that is most closely related to the vector of the vector coordinate system question data, and selects divided language character file data corresponding to the most closely related vector coordinate system divided language data or the most closely related vector coordinate system divided language data; an answer text generating and transmitting step in which the server adds the selected divided language character file data to the question information, generates text data using a text generating means which is a pre-trained artificial intelligence means, and transmits the generated text data to the user terminal; A program for an artificial intelligence system, characterized in that the program comprises a generated text display / voice output step in which the user terminal displays the received generated text data on a browser or outputs it as voice.

Images